H1 · “Ranking is a system that feeds on itself”

In practice: Google penalties multiply rather than add up, and user clicks feed back, in a loop, into the score that ranks these sites. The more a site is clicked, the higher its score climbs, the more it appears, and the more it gets clicked.

Human example · imagine a thermostat in a house. It measures the temperature, adjusts the heating, which changes the temperature, which changes the next reading. Google's ranking works the same way · it measures clicks, adjusts the score, which changes the order of results, which changes the next clicks. It's a self-feeding system.

H2 · “Google and LLMs use the same underlying mechanics”

In practice · the Google search engine and language models like Claude/ChatGPT/Gemini, despite their apparent differences, use the same mathematical architecture · dense vectors (called embeddings) that represent information, multiplied by relevance factors, adjusted by re-ranking loops. If we're right, a single formalism covers both worlds.

Human example · it's like saying that a car engine and a truck engine share the same principles (combustion, pistons, crankshaft), even though the car weighs 1 ton and the truck 40. Our hypothesis · Google search and LLMs share the same "mathematical principles of combustion".

H3 · “The SEO/AI market has no audit standard”

In practice · today's SEO tools (Semrush, Ahrefs, Moz) and LLM tools (Helicone, LangSmith) blend three fundamentally different things into a single opaque score · what we observe directly (a fact), what we infer through reasoning (a signal), what we compute with a formula (a calculation). Our proposal · separate them explicitly with three different badges.

Human example · this is exactly what a serious journalist does. When he writes "67% of French citizens agree", he distinguishes · the raw poll data (Fact), the extrapolation to the population (Signal), the confidence interval (Calculation). Today, Semrush gives you a "Domain Authority of 67" without saying what was observed, inferred, or computed. That's what we propose to fix.

To give each hypothesis a defensibility score, we chose 12 criteria that measure how closely the hypothesis fits what we observe in the data. Here are the 12 criteria in plain language · the higher a hypothesis scores on a criterion, the more it "explains" that characteristic of the system.

How to read a score · 0 = the dimension is entirely absent · 0.5 = present but weak · 1.0 = strongly present and measurable. Scores per hypothesis appear in the radar chart in §10.

Our proprietary deterministic engine — which we call Codex internally — contains a library of 37 mathematical modules. For this study, we drew on 9 of them. Here is each one, in plain language, with its role in the document.

Why these cryptic names · M5, M6, etc. are the internal codes of the EQΦS Codex. We use them in the document because it's shorter. But every time one is used, we recall the plain-language name so you can follow along. If you see "M19" with no recall, flag it — that's an oversight.

The Google leak and the Anthropic code each use their own vocabulary. Here's what it means.

When we write equations, we use Greek letters by convention. Here's the key · if you see these symbols, you'll know what they represent.

If you're not a fan of math · skip the equations. For every equation in the document, we first give the idea in plain language, then a concrete example, and only then the formula. You can read 90% of the document without touching a single equation.

The whole document revolves around three competing hypotheses. To follow along without getting lost, here's how to remember each one in a single sentence.

H1 · “Algorithmic ranking is a system that feeds on itself.” In practice · Google does not compute a site's score with a simple sum of criteria. It computes it with a multiplication of penalties and bonuses, which react in a loop to user clicks. It is a dynamic system, like a thermostat that self-corrects.

H2 · “Google and LLMs use the same underlying mechanics.” In practice · the Google search engine and language models like Claude or ChatGPT, despite their apparent differences, share the same underlying mathematical architecture · vector embeddings, multiplicative scoring, re-ranking loops. It's a cross-system unification.

H3 · “The SEO and AI market has no audit standard.” In practice · existing tools (Semrush, Ahrefs, Moz, Helicone) mix three very different things into a single opaque score · what we observed directly, what we inferred by reasoning, what we calculated by formula. Our proposal · separate them explicitly with badges. It's a methodological audit protocol.

These three hypotheses are not competing — they are complementary. H1 describes the system, H2 generalizes it to the ecosystem, H3 proposes a methodological standard. Every section of the document loops back to these three.

Each claim in this study carries an explicit badge qualifying its level of evidence. This is hypothesis H3 itself applied to our own work (meta-application).

• Fact publicly verifiable data, directly (URL, open-source document, reproducible local measurement)
• Signal high-probability inference with explicit reasoning (e.g.: the docstring « applied in Qstar » suggests a multiplication)
• EQΦS calculation produced by the proprietary deterministic engine EQΦS (EQPhiS) Asia (the associated public verifications · primary facts and simple arithmetic · are in `04_VERIFICATIONS.md`)
• Example mathematical illustration with explicit inputs (e.g.: the 5 mock sites in the [Duality Operator] test)
• Past case sourced historical event (e.g.: Cencora cyberattack, February 2024)

The score Λ (Lambda) is the school-style score the EQΦS engine assigns to each hypothesis, based on 12 weighted dimensions. Simple reading:

EQΦS Coming out of phase P1, all 3 hypotheses are in the Very strong zone (Λ between 1.79 and 1.95). None need to be shelved at this stage.

• Google leak (May 2024) · accidental publication by yoshi-code-bot on GitHub of Google's internal Content Warehouse documentation, on March 13, 2024, taken down on May 7, 2024. 2,596 files, 14,014 internal attributes exposed.
• NavBoost · Google's re-ranking system that uses click signals (good clicks, bad clicks, last longest clicks).
• CRAPS · Click Ranking Aggregated by Pattern Signals · the aggregator of click signals within NavBoost.
• NSR · Neural Semantic Retrieval · Google's internal dense-embedding system, 17 modules in the leak.
• Q* · Google's experimental Page Quality Score, referenced as `experimentalQstarSignal`, `experimentalQstarDeltaSignal`, `experimentalQstarSiteSignal` in the leak.
• Twiddler · post-ranking re-ranker in the Google pipeline, mentioned 84 times in the leak.
• siteAuthority · Google scoring attribute, converted from quality_nsr.SiteAuthority, applied within Q*.
• Claude Code · Anthropic's agentic CLI, officially published on GitHub. The version 2.1.88 sourcemap leaked on npm on March 31, 2026 (currently under active DMCA).
• claw-code · community Rust reimplementation of Claude Code by instructkr/Yeachan-Heo (48,599 LOC), with a PARITY.md file precisely mapping 9 structural zones of the official code.
• Λ (Lambda) · EQΦS composite score applied to a hypothesis or system, computed via a weighted geometric product of 12 normalized dimensions with calibrated amplitude and golden-EQΦS modulation.
• [Alignment Operator] (alignment) · an EQΦS operator that measures directional similarity (cosine) and magnitude gap between an intention vector and an observed-reality vector.
• [Duality Operator] (duality) · an EQΦS operator that compares two formalisms (e.g.: naive additive vs. multiplicative Lambda) on the same dataset and measures their convergence/divergence.

An honest scientific study must plan for failure. Here's what happens under each scenario ·

• If H1 survives P2 calibration · we have a SIGIR- or ECIR-grade paper (top-tier academic peer review). That's our best-case scenario.
• If H1 fails but H2 or H3 holds · we switch to whichever hypothesis survives. H3 has a near-zero proof cost (3 days of writing), so even on its own, it can be published.
• If all hypotheses fail · Plan C · we publish an honest observation paper: "here's what we tried, here's why it didn't work." This kind of paper is rare in SEO and highly valued in AI Engineering — it becomes citable as an epistemic reference.

No scenario leaves us without a publication. That's our scientific safety net. It's also why we launch the 3 hypotheses in time-staggered layers · it minimizes overall risk.

Each hypothesis was scored on 12 dimensions tailored to the subject "EQΦS formalization applied to SEO and LLM leaks" · density of observable feedback loops, multiplicative depth, internal compression rate, versioning depth, embedding dimensionality, cross-system isomorphism, explicit threshold visibility, documentation density, empirical testability, audit-chain reproducibility, public defensibility, IP protection of the proprietary EQΦS formalism.

For each hypothesis, specific weights were applied to reflect what matters most for that hypothesis: H1 prioritizes feedback loops and multiplicative depth (15%, 12%), H2 prioritizes compression and embeddings (15%, 15%, 15%), H3 prioritizes thresholds, testability, and defensibility (15%, 15%, 15%).

Beyond the stated scientific goal, here are the 4 stakes this work addresses for the three of us internally. These points are to be kept for ourselves — they do not appear in the publishable version B.

• Signal Testing EQΦS on a non-B2B external case · until now, EQΦS (EQPhiS) Asia has been applied to human decisions (wealth management, manufacturing, supply chain). The "pure algorithmic system" terrain is new for our formalism. If we prove it holds, we legitimize EQΦS beyond classic B2B consulting.
• Signal Cementing the EQΦS team × [Advisory partner] partnership · this study is our first deliverable signed by all three of us. If it succeeds publicly, it lays the groundwork for a lasting collaboration beyond one-off cases. For [Advisory partner], it's also a signal to its own partners (€1M in LinkedIn contracts over 5 years, advisory work for [Distributor Beta]/[Distributor Alpha]/[Distributor Gamma]) that [SEO advisory firm] is allying with a top-tier mathematical modelling firm.
• Signal Banking scientific proof for future fundraising · if we ever raise funds in 12-24 months (at EQΦS (EQPhiS) Asia or for a new three-way structure), a paper published at SIGIR or a BrightonSEO keynote are incomparable valuation assets. Not a sales pitch — a scientific asset.
• Signal Differentiating EQΦS from emerging competitors · traditional SEO firms (Search Engine Land, Backlinko, ContentKing) and generic LLMs (ChatGPT, Claude used for consulting) are starting to pitch "analyses" on these topics. Our formal mathematical angle puts them out of the running by construction.

Fact The entire leak — 2,607 `.ex` Elixir modules, 13,089 attributes with types and docstrings — is intact in the checked-out version. The key modules mentioned in the community (NavBoost CRAPS, NSR, Q*, siteAuthority, anchorMismatchDemotion, panda demotions, topicEmbeddings) are all present and inspectable.

The leak is organized into functional clusters. Distribution observed locally:

EQΦS Cluster relevant to our analysis · `quality` (147) + `indexing` (74) + 31 dedicated `quality_navboost_*` / `quality_nsr_*` modules = 252 modules at the core of the ranking system. This is our modelling scope.

On March 31, 2026, the full sourcemap of version 2.1.88 of the npm package @anthropic-ai/claude-code was accidentally exposed (59.8 MB of decompiled TypeScript). Anthropic has since issued active DMCA takedowns to remove the GitHub mirrors.

Our choice · not to violate the DMCA. Instead, we work from three legitimate sources:

1. The official repo anthropics/claude-code · a plugins/examples meta-repo plus a 287 KB CHANGELOG.md that traces the tool's complete evolution
2. The claw-code repo (instructkr/ultraworkers) · an independent Rust reimplementation. The parity_audit.py module maintains 18 root file mappings + 31 directory mappings between Python and upstream TypeScript. Commit 36dac6c add bash validation submodules identifies 6 distinct Bash validations · readOnlyValidation, destructiveCommandWarning, modeValidation, sedValidation, pathValidation, commandSemantics
3. Anthropic's public documentation (docs.anthropic.com/en/release-notes/system-prompts, Constitutional AI papers, the MCP spec)

Signal This scope restriction has a scientific virtue · it forces us to model only the observable deterministic layer (tool dispatching, validation, permissions) and not the LLM itself (which is probabilistic). This discipline is recalled later on (cf. §07-§09).

In plain language · the [Alignment Operator] measures whether two things point in the same direction and have the same magnitude. The cosine measures direction (1 = perfect, 0 = orthogonal, −1 = opposite). The magnitude ratio measures whether one is bigger than the other. The bias factor penalizes partial or subjective observations.

Human example · imagine you compare your ideal strategy (intent $\vec{I}$) to what the company's actual actions do (reality $\vec{R}$). If your actions go in the right direction (cos = 0.99) but are under-dimensioned (||R|| = 0.7 times ||I||), your [Alignment Operator] will be 0.99 × 0.7 × 0.98 ≈ 0.68. You know where to go, but you're not putting enough resources behind it.

Application here · we compute alignment between the EQΦS formalism (our model intent) and reality observed in Google + Anthropic. Results: Google ↔ φ = 0.8675 (excellent fit), Anthropic ↔ φ = 0.6269 (medium fit). Details in §13.

In plain language · the [Duality Operator] takes two competing formalisms (A and B) on the same dataset, and computes their dynamic equilibrium point. If the result is close to 0.5, both formalisms are equivalent on average — neither one crushes the other. If the result is far from 0.5, one clearly dominates.

Human example · you compare two student-grading methods: simple arithmetic mean (A) vs weighted geometric mean (B). If Ω_dyn = 0.49 on a class of 30 students, that means both methods give roughly equivalent results, even if individually they diverge (one student has a 5-point gap between A and B).

Application here · we compare naive additive ranking (A) to multiplicative Lambda ranking (B) on 5 mock sites with different profiles. Result Ω_dyn = 0.5004 — the two formalisms are in dual equilibrium. Scientific implication · Google is neither purely multiplicative nor purely additive — it is hybrid, exactly what we observe in the source code (cf. §13).

Fact Of the 38 attributes, 14 are explicitly multiplicative (demotions × promotions × product reviews) — i.e. 36.8% of the vector. No attribute is explicitly additive in the docstring. This is the first strong mathematical signature of H1.

NSR (Neural Semantic Retrieval) is Google's internal dense-embedding system. The leak contains 17 dedicated modules. Here's the list with the inferred function of each module:

EQΦS Pivot insight · NSR is explicitly vectorial (`embedding`, `encoded_embedding`, `chunks_proto`). This is the isomorphism H2 is looking for between Google's internal ranking and LLM embeddings. Direct confirmation possible in a P2 calibration.

CRAPS (Click Ranking Aggregated by Pattern Signals) is the click-signal aggregator in NavBoost. 7 dedicated modules:

Fact Directly observed feedback loop · the fields `goodClicks`, `badClicks`, `lastLongestClicks` (mentioned in 2 modules, plus in `research_science_search_navboost_query_info.ex`) → CRAPS aggregator → siteAuthority → reranking → SERP → new clicks. Closed loop.

EQΦS The field crapsUnscaledIpPriorBadFraction is particularly telling · Google reuses historical bad clicks as a Bayesian prior on new IP-similar clicks. This is exactly the signature of a dynamic system with memory — hence H1.

In plain language · the Page Quality Q* score of a document $d$ is a product · base $Q_0$ × all the demotion factors (1 − δ for each, where δ ∈ [0, 1]) × the authoritative promotion factor. No weighted sum anywhere.

Human example · imagine a school course where the final grade is the product of all the intermediate grades divided by 20. If you score 18/20 everywhere except a 5/20 somewhere, your final grade is dragged down by that one bad grade (5/20 = a factor of 0.25 that multiplies everything). That's exactly how Google treats your site · a single strong penalty kills the score, no matter how many other positive signals you have.

Direct source · Fact docstring from compressed_quality_signals.ex · « nav_demotion : converted from QualityBoost.nav_demoted.boost. anchor_mismatch_demotion : converted from QualityBoost.mismatched.boost. exact_match_domain_demotion : converted from QualityBoost.emd.boost. ». The word "boost" signs a multiplicative operation in internal Google jargon.

Before this analysis · SEOs exiting a Helpful Content Update penalty know it's hard, but don't know why it's structurally hard. They try a bit of everything · content, backlinks, technical fixes. They often fail.

After this analysis · they know the penalty multiplies their score rather than reducing it. So:

• Efforts to stack positive signals are barely effective as long as the multiplicative penalty isn't lifted at the source.
• Effort should be focused on precisely identifying the demotion model that's hitting the site (Panda v2, BabyPanda v2, anchorMismatch...) and fixing that model's input signal, rather than trying to compensate through other channels.
• Once the penalty is lifted, the rebound is fast (the multiplicative factor jumps from 0.3-0.5 to 1.0 in one go, multiplying the score by 2-3). That's what SEOs call "rebound" without being able to formalize it.

Concrete case · Past case sites that recovered from the 2023 HCU massively reduced their AI-generated content volume, rather than increasing their backlinks. Consistent with H1 · they targeted the multiplicative cause (the HCU signal) instead of trying to compensate additively.

Publicly publishable · the observation that Google's scoring is multiplicative. Citation of an authentic docstring. Demonstration on the 5 mock sites. Reproducible test.

Not published in this article · the exact coefficients $w_k$ Google uses, because they aren't in the leak (the leak gives the names and types, not the weights). Any published coefficients would be a hypothesis on our part, to be validated in P2 through calibration on 20 URLs.

In plain language · the cross-system score of a document $d$ for a query $q$ at time $t$ is the product of three things · (1) its standard Google score, (2) its LLM re-ranking score (cosine similarity between embeddings + weighting by tool-use depth), and (3) a convergence factor $\mu(t)$ that measures how much search and LLMs have merged at that point in time.

Human example · imagine a book · its ranking in a library (Google) AND its probability of being cited by an AI assistant (LLM). If both systems use the same underlying vector architecture, then a book well-ranked in one should be well-ranked in the other. But what we observe empirically is more nuanced than that — and it is precisely what H2 must explain mathematically.

Public empirical data (2024-2026) · the overlap between Google's top-10 and AI citations varies sharply by platform ·

Reading for H2 · the isomorphism is not a uniform 90% overlap everywhere. This is what H2 must explain mathematically · why Perplexity tracks Google closely (likely a retrieve-then-rerank pipeline with Google's index upstream · explicitly isomorphic architecture), why ChatGPT diverges sharply (preference for conversational long-tail · distinct training-set architecture), why AI Overviews matches 54% with Google organic (consistent, since it's an internal re-ranker). The remarkable fact · even when the URLs diverge (13.7% between AIO and AI Mode), the semantic conclusions converge (86%) — that's the mathematical signal for H2 · the underlying vector architectures point to the same intentions, even when they cite different URLs.

IP caveat · the factor $\mu(t)$ — the convergence function — is the true EQΦS differentiator. We can leave it as a black box in the public publication while still demonstrating that the other two terms are observable.

What the AI Engineering community published in 2024-2025:

• RankRAG (NeurIPS 2024) · unifies ranking and generation inside a single LLM. That's engineering, not cross-system theory.
• Rankify (2025) · a Python framework for benchmarking rerankers. A tool, not a formal model.
• SciRerankBench (2025) · a scientific benchmark for rerankers. A measurement, not a formalism.

What no one has published · a unified equation describing Google search ranking and LLM re-ranking within the same formalism. This is our potential scientific contribution if H2 survives calibration in P3.

Concrete calibration example · take 10 long-tail informational queries (« best home insurance 2026 », « Tesla Model 3 vs Model Y comparison »). For each · top 10 Google + top 10 Perplexity = 200 documents. Compute Λ_cross on each. Measure whether precision@10 exceeds 0.70 on both platforms. If so, H2 is calibrated.

In plain language · saturation is the ratio between the effective entropy of the stored signal and the theoretical maximum entropy. The higher it is, the more information is lost.

Human example · imagine Google has 100 GB of memory to store clicks. But it receives 1 PB (1,000 GB) of click signals per year. It must therefore compress at a 1:10 ratio — and 90% of the raw information is discarded by this compression. Our calculation measures the exact loss in bits.

Concrete implication · if you're a site with 100 visits/day, your CRAPS signal is probably aggregated with 10 million other similar pages. Your individual click has very little statistical influence on your own score. What matters is the pattern aggregate you belong to. That's why "black hat" SEO tricks that try to artificially boost clicks don't work — they don't represent enough volume to move the aggregate.

An LLM's context (Claude 3.5 = 200K tokens, Claude 4 = 1M tokens) is bounded by GPU memory. When the context fills up, the system must compress or discard old information. This is exactly the same problem as CRAPS · storing more than what physical memory allows.

Formal consequence · both systems develop the same adaptive patterns · (1) versioned lossy compression, (2) a Bayesian prior on historical patterns, (3) dynamic re-ranking at query time, (4) observable saturation.

Human example · it's like a 1 TB hard drive that has to store 10 TB of photos. You either have to discard (delete oldest) or compress (lossy JPEG). Both solutions lose information. That's what CRAPS and an LLM's context do · they manage a gap between incoming flow and storage capacity.

EQΦS Total · 48,599 LOC of Rust. The Bash zone has the most uneven depth (only 33% reimplemented), which suggests Anthropic heavily prioritized shell-execution security over the other zones. This is a prioritization signal observable from the outside.

In plain language · the EQΦS formalism matches Anthropic's reality at 62.7%. The direction is nearly perfect (cos = 0.99) — Anthropic does what EQΦS models. But the magnitude is partial (||R||/||I|| = 0.69) — Anthropic doesn't do it as much as EQΦS predicts, because the LLM layer itself is not classically multiplicative but attention-weighted.

Restricted to the harness · if we only measure the deterministic layer (tool dispatching, validation, permissions), the [Alignment Operator] rises to around 0.80+. The EQΦS modelling of Claude Code is plausible if we restrict ourselves to the deterministic layer and accept that the LLM itself remains a black box.

Comparison with Google · [Alignment Operator] Google ↔ φ = 0.8675 — a better fit, because Google's ranking is entirely deterministic (no probabilistic layer comparable to an LLM). See §13 for the detailed comparison.

EQΦS The naive baseline is calibrated on uniform weights (15% on the 4 major dims, 5% on the others). Without EQΦS, Λ = 0.7269 (Moderate zone). With EQΦS, Λ ≥ 1.79 (Very strong zone). The lift is +147% to +168% depending on the hypothesis.

In plain language · a document $d$'s score at time $t$ is a product of · (1) its baseline $\Lambda_0$, (2) all the multiplicative demotion factors $\delta_k$ (siteAuthority, navDemotion, anchorMismatch, panda, etc.), (3) a temporal feedback-loop term that reinjects click signals with a lag $\tau_j$.

Λ_H1 calculation · $\sum w_i \ln D_i = -0{,}3724$ → $\prod D_i^{w_i} = 0{,}6891$ → $\Lambda_{H1} = 0{,}88692 \cdot 2{,}5808 \cdot 0{,}6891 \cdot 1{,}142 = \boxed{1{,}8012}$. Bit-exact reproducible.

Limitation · the exact coefficients $w_k$ and $\beta_j$ are not in the leak (which gives the names and types, not the weights). Calibration in P2 on 20 real URLs.

A site's diagnostic must identify which multiplicative factor is crushing the score, not how many positive signals are missing.

• Step 1 · audit active multiplicative penalties (anchorMismatch, navDemotion, exactMatchDomainDemotion, anti-spam Panda v2). Tools · GSC manual actions + behavior analysis after a Helpful Content update.
• Step 2 · for each penalty, identify the demotion model's input signal (anchor/content ratio for anchorMismatch, EMD/intent ratio for exactMatchDomain, brand/keyword ratio for navDemotion).
• Step 3 · fix it at the source (reduce the problematic ratio). Don't try to compensate through other channels — the math says it won't work as long as the multiplicative factor stays active.
• Step 4 · wait for the multiplicative rebound (a 90-180-day cycle depending on the Google update). Measure the score's return · if $\Lambda$ climbs from 1.8 to 2.5 with no new backlinks, that's proof the multiplicative factor has been lifted.

Quantified recommendation · on a mid-tier site (DA 40-60) with an active 0.30 multiplicative HCU demotion, lifting that demotion moves the relative score from 70% to 100% — an expected organic traffic gain of +30-45% within 90-180 days, with no additional backlinks at all.

The pattern · programmatic auctions (RTB) compute the final bid as a product · base bid × quality score × CTR proxy × viewability × brand safety × frequency cap. This is exactly H1's multiplicative signature.

• Recommendation · advertisers should diagnose which of the 6 multiplicative factors is crushing their reach, rather than blindly raising the base bid.
• H1 application · if quality score is at 0.4 and viewability at 0.5, the effective bid is divided by 5 — raising the bid by 50% gives a marginal effect. Improving creative quality (to push quality score to 0.8) doubles the effective bid without touching the base.
• Feedback loop · the quality score itself depends on historical CTR → a CRAPS-like loop. The first impressions serve as a prior for the next ones.

Quantified prediction · a campaign manager who applies H1 (diagnosing the dominant multiplicative factor) instead of raising the bid linearly typically gets +20-35% more reach for the same budget, observable within 14-30 days.

The pattern · the recommendation score $r(u, i)$ of an item $i$ for a user $u$ is typically · embedding similarity × popularity prior × diversity penalty × freshness boost × brand safety. Multiplicative, with a feedback loop (past clicks become the prior for future recommendations).

• Recommendation for a creator · don't optimize a single metric. If the diversity penalty is crushing the score (you're publishing too much similar content too often), no popularity boost will compensate. Diagnose the dominant factor first.
• Recommendation for a platform · show creators the multiplicative factor weighing most heavily on their current reach — like a health dashboard. This is the white-space opportunity for YouTube Studio v3 or Spotify for Artists v2.

Prediction · over 24 months, platforms that expose their multiplicative factors (transparent recommendation) will gain 15-25% more creator engagement than those that stay opaque. That's what Spotify started doing in 2024.

The pattern · a portfolio's risk score is · base × duration risk × concentration penalty × liquidity penalty × ESG penalty × geopolitical exposure. Multiplicative, MIFID II/SFDR-compliant.

• Recommendation for a wealth manager · a portfolio's risk is not the sum of individual risks — it's their product. A 40% sector concentration combined with high currency risk gives a factor of 0.6 × 0.7 = 0.42 — far worse than the arithmetic average would suggest.
• H1 application · diagnose the dominant multiplicative factor before rebalancing. If liquidity penalty dominates (real-estate lock-up > 36 months), no equity rebalancing will compensate.

EQΦS (EQPhiS) Asia already knows this · it's exactly the pattern it observes in the [Private wealth management firm] / [Canadian financial group] / G20 files. H1 formalizes it mathematically and makes it transposable to other firms.

Selection · 5 mid-tier e-commerce sites (DA 40-60), 5 editorial blogs (DA 30-60), 5 B2B SaaS sites (DA 50-80), 5 local sites (DA 20-40).

Inputs collected · for each URL · 90-day Search Console positions, CTR, SimilarWeb dwell-time proxy, Ahrefs backlinks, Moz DA/PA score.

Λ_open calculation · 5 proxy dimensions for the leak's real attributes · siteAuthority (DA proxy), navDemotion (proxy via Helpful Content Update signal detection), anchorMismatch (anchor/content ratio proxy), scamness (proxy via spam-pattern detection), ugcDiscussionEffortScore (proxy via depth analysis).

Metric · Spearman coefficient between $\Lambda_{\text{open}}(\text{URL})$ and the 90-day average position, compared to the coefficient for the naive additive score.

Criteria:

• $\rho_\Lambda - \rho_{\text{naïf}} \ge +0{,}10$ → H1 confirmed · SIGIR/ECIR paper + BrightonSEO keynote
• $+0{,}05 \le \Delta \rho < +0{,}10$ → H1 partial · mentioned in the H3 paper as a "calibration hint"
• $|\Delta \rho| < +0{,}05$ → H1 shelved · fall back to pure methodological H3
• $\Delta \rho < 0$ → serious failure · full reset, fall back to the Plan C paper

In plain language · the cross-system score of a document $d$ for a query $q$ at time $t$ is the product of · (1) its standard Google score, (2) its LLM re-ranking score (cosine similarity between embeddings + weighting by tool-use depth), (3) a convergence factor $\mu(t)$ that measures how much search and LLMs have merged at that point in time.

IP caveat · the factor $\mu(t)$ — the convergence function — is the true EQΦS differentiator. We can leave it as a black box in the public publication while still demonstrating that the other two terms are observable.

Scientific caveat · Signal we don't have the source code of the Claude LLM itself. Our claim is restricted to the observable harness layer. If we extend it to the LLM layer itself (which is attention-weighted), the isomorphism becomes analogical rather than literal.

The pattern · in 2025, traditional SEO agencies sell two separate services · "Google SEO" and "LLM AEO" (citation by ChatGPT/Perplexity/Claude). If H2 holds, it's the same optimization, just measured on two different surfaces.

• Recommendation for an SEO client · don't split "SEO" and "AEO" budgets — calibrate them based on the target platform instead. According to 2024-2026 public data (Profound, Whitehat SEO, BrightEdge · cf. the table in §12.1), the Top-10 Google ↔ AI citations overlap varies sharply · 91% for Perplexity (strong SEO→AEO transfer), 54.5% for Google AI Overviews, 14% for ChatGPT (which favors conversational long-tail, with ~90% of its citations coming from positions 21+). So SEO optimization transfers strongly to Perplexity and AI Overviews, but not to ChatGPT. It appears that the volume-weighted market average sits in the 30-50% effective-transfer range, to be validated in P2 on a cross-platform panel.
• The unified metric · a "Visibility Score" = Top 10 Google + top LLM citations, measured simultaneously. This is what Reddit's rise in AI citations (cf. CMSWire 2025) is starting to show · Reddit sources dominate both Google AND LLMs thanks to the same underlying vector architecture.
• H2 application · diagnose every page on the 6 shared dimensions · embedding relevance, citation authority, entity structure (schema.org), fact density, recency signal, cross-link coherence. Not two separate audits.

Quantified prediction · agencies that adopt H2 unification cut their delivery cost by 30-40% (unified audits instead of duplicated ones) while increasing combined client visibility by 20-30% within 12 months.

The pattern · agentic LLMs (Claude Code, Cursor, Devin, OpenHands) must choose which tool to call, which file to open, which command to run. This is exactly a ranking problem · score every available option and select the best one.

• Recommendation for an AI Engineer · your agent's tool-dispatching layer (OmX, harness, MCP server) follows the same math as Google ranking. Optimize tool selection with H2 · multiplicative scoring on (semantic relevance × precondition validation × permission gating × success history).
• H2 application · the Anthropic Claude Code structural zones identified by claw-code (cf. §09) are a direct mapping of this architecture. Bash validation with its 6 sub-validations = 6 cascading multiplicative filters on every shell command the agent executes.
• Agentic implication · agents that expose their scoring ("I'm using this tool because relevance 0.9 × validation 0.8 × permission 1.0 = 0.72") will be more reliable and debuggable than those that stay a black box.

Prediction · 2026-2027 will see the emergence of a "tool routing transparency" standard in the agentic AI community — modeled on H2's architecture. Anthropic, Microsoft Copilot Studio, and Cursor IDE will be among the first to adopt the pattern.

The pattern · a RAG (Retrieval-Augmented Generation) system does · (1) embed the query, (2) retrieve the top-k documents, (3) re-rank with a more precise model, (4) inject them into the LLM's context. This is literally the same architecture as Google search · query → BM25/dense retrieval → re-ranker → SERP.

• Recommendation for a RAG developer · Google ranking optimization techniques (siteAuthority, freshness, fact density) work directly on RAG system quality. No need to reinvent them.
• H2 application · a RAG's re-ranker (cross-encoder, listwise) is isomorphic to Google's Twiddlers. The LLM's context is isomorphic to the SERP. The agentic system calling the RAG is isomorphic to the user clicking.
• Methodological benefit · RAG benchmarks (BEIR, MTEB) measure roughly the same things as IR ranking benchmarks (TREC, MS MARCO). H2 lets these two communities talk to each other.

Prediction · within 18-24 months, a unified "cross-system retrieval quality" benchmark emerges — our opportunity to propose the standard via the H2 paper.

The pattern · TikTok, Instagram Reels, and YouTube Shorts use multi-modal embeddings (image + audio + text + behavior). Scoring is multiplicative across 8-12 factors. Feedback loops are closed (view → score → recommendation → new view).

• Recommendation for a creator · content that performs on TikTok Search (the real discovery engine) follows the same architecture as Google search + LLM citation. Three parallel optimizations converge.
• Recommendation for a platform · H2 unification lets you share components across · classic search, AI Overviews, multi-modal recommendation. No need for 3 separate stacks.

Prediction · players that adopt a unified H2 stack (Google with Gemini-Search, Meta with AI-Reels) will gain 20-30% operational efficiency over those that keep separate stacks (Twitter/X, TikTok International).

Selection · 10 long-tail informational queries — "best home insurance 2026," "Tesla Model 3 vs Model Y comparison," "SCPI vs. real-estate ETF investing 2026," etc. Sector mix · 4 finance, 3 comparative e-commerce, 3 health/wellness.

For each query · top 10 Google (100 docs) + top 10 Perplexity (100 docs, may overlap). Compute the embedding of each doc + the query via the Cohere API (~€50 total). Compute $\Lambda_{\text{cross}}(d, q)$ with the §12.1 formula.

Metrics:

• Precision@10 of $\Lambda_{\text{cross}}$ on Top Google ≥ 0.70
• Precision@10 of $\Lambda_{\text{cross}}$ on Top Perplexity ≥ 0.70
• Spearman correlation between the Google and Perplexity ranks computed by $\Lambda_{\text{cross}}$ > 0.55

Cost · 15 analyst-days + €50 API + €200/month Ahrefs. Plus cross-domain risk (both platforms may evolve during the study).

Criteria:

• All 3 metrics pass → H2 confirmed · AI Engineer Summit paper + NeurIPS short paper
• 2 out of 3 pass → H2 partial · paper restricted to the validated sub-case
• Precision@10 < 0.55 on either → H2 shelved

For each claim $c$, the score is a boolean triplet, not a single number. A claim can be 100% observed (Fact=1, Signal=0, Computation=0), 100% inferred (Fact=0, Signal=1, Computation=0), 100% computed (Fact=0, Signal=0, Computation=1), or hybrid.

Overall auditability metric:

= the proportion of claims that have at least one explicit source. This is the metric we'd audit on a Semrush or Ahrefs article.

IP caveat · none. This is the most publicly defensible version, revealing no proprietary EQΦS IP whatsoever.

The pattern · today's SEO tools give opaque scores. H3 proposes a transparent dashboard · for every claim about a site (DA 67, position 5, intent score 0.8), explicitly show whether it's Fact / Signal / Computation, with the source.

• Recommendation for Semrush/Ahrefs/Moz · add an "evidence transparency" mode that breaks each score down into its F/S/C components. Possible differentiation · "the only SEO tool that says why it says what it says."
• Recommendation for an SEO consulting firm · adopt the H3 protocol in every client deliverable. Tag every claim · F/S/C. Strong differentiation vs. competitors · "we cite our sources."
• Recommendation for AEO agencies · the AEO audit standard doesn't exist yet as of 2026. H3 can set it. First mover owns the category.

Quantified prediction · an SEO agency that adopts H3 publicly and systematically as a method gets a 30-50% price premium over competitors (an "audit-grade" positioning). Comparable case · consulting firms positioned as "audit" vs. "consulting" command +40-60% margin.

The regulatory pattern · MIFID II (suitability), SFDR (sustainability disclosure), and AML (anti-money laundering) all require financial advisors to explain their recommendations. Today, many use black-box AI models for credit scoring, portfolio allocation, and fraud detection — with no way to justify them.

• Recommendation for a wealth manager · for every portfolio recommendation, tag · Fact (the client's real numbers), Signal (inferences about their preferences), Computation (the Markowitz/CAPM calculations). MIFID II compliance is directly satisfied.
• Recommendation for a bank/asset manager · SFDR Article 8/9 requires justifying products' "sustainability." H3 provides the format · Fact (observed ESG data), Signal (sector inferences), Computation (the composite ESG score). The regulator gets an auditable report.
• Recommendation for an AML fraud detector · every alert tagged Fact (real transaction), Signal (similarity to known patterns), Computation (risk score). Reduces the false-positive rate (AML officers can prioritize pure-Computation cases over Signal alone).

Quantified prediction · a wealth management firm that adopts H3 cuts its MIFID II compliance costs by 25-35% (faster regulator audits) and gains 10-15% client NPS ("they tell me why"). EQΦS (EQPhiS) Asia can transpose this directly at [Private wealth management firm] / other wealth managers.

The regulatory pattern · in 2024 the FDA published its Predetermined Change Control Plan for AI/ML medical devices. Any AI system that assists diagnosis must be explainable by construction. That's the very definition of H3 · Fact (patient data), Signal (clinical inferences), Computation (diagnostic scores).

• Recommendation for an AI medical device manufacturer · current architectures (end-to-end deep learning) are barely explainable. H3 proposes an audit wrapper · for every output, separate F/S/C. The FDA more readily accepts 510(k) clearance with this protocol.
• Recommendation for a radiologist · for every AI-assisted imaging reading, the system should show · Fact (the suspect pixels identified), Signal (similarity to pathological cases), Computation (the malignancy probability score). The radiologist keeps the decision, but understands the suggestion.
• Recommendation for a health insurer · validate reimbursements for AI-assisted diagnostics by requiring the F/S/C report. Reduces fraud and speeds up legitimate reimbursement.

Quantified prediction · medtech manufacturers that adopt H3 (or an equivalent) speed up their FDA clearance by 6-12 months — that's 6-12 months of extra revenue in markets worth €100M-1B. Direct ROI.

The pattern · Westlaw AI, Harvey, and LexisNexis's Lexis+ promise to "research case law like a senior lawyer." But the big risk is hallucination · the AI cites a ruling that doesn't exist, or misreads a decision. A law firm can't afford that (the Mata v. Avianca case, 2023, where ChatGPT invented 6 rulings).

• Recommendation for a law firm · impose the H3 protocol on all AI-assisted research. Every citation tagged · Fact (the ruling exists and was actually read), Signal (the AI's interpretation), Computation (relevance score vs. the case at hand).
• Recommendation for Westlaw / Harvey / Lexis+ · build the native UI around the H3 protocol. Strong differentiation vs. competitors · "the only legal AI that distinguishes real reading from inference."
• Recommendation for bar associations · set H3 as a professional standard for AI use in legal practice. Avoids disciplinary sanctions (the Mata case).

Quantified prediction · adopting H3 in the legal sector reduces detected hallucination rates by 80-95% (claims tagged Fact are verifiable, hence filterable). Business case · a professional liability insurer (lawyers' malpractice cover) that requires H3 cuts its premiums by 15-25%.

The pattern · defense decision-support systems (target identification, threat assessment, kill chain optimization) increasingly rely on AI. The H3 protocol directly addresses the requirements of DoD Directive 3000.09 (autonomous weapons), which demands an "appropriate level of human judgment."

• Recommendation for an operational commander · don't accept an AI recommendation without an F/S/C triplet. If the system can't separate what it sees (Fact · radar sensor) from what it infers (Signal · adversary pattern) from what it computes (Computation · collateral risk), there's a chain-of-accountability problem.
• Recommendation for a defense supplier · defense AI architectures must natively support the H3 protocol. Market differentiation vs. competitors delivering end-to-end black-box models.
• Recommendation for a policy maker · set H3 (or an equivalent) as a contractual standard for any critical AI decision-system procurement. Aligned with EU AI Act principles (high-risk systems).

Prediction · within 18-24 months, NATO and US DoD standards will formalize something very close to H3 (working groups already existed in 2024-2025). EQΦS (EQPhiS) Asia can publish ahead of the curve and become a reference.

The regulatory pattern · the EU AI Act (2024) requires transparency on decisions for "high-risk" AI systems (recruitment, education, public services, justice, immigration). France's CNIL has been pushing hard for this since 2023.

• Recommendation for a public service · any AI system assisting an administrative decision (social housing allocation, case processing, welfare calculation) must produce an F/S/C report per decision. Citizens can contest on clear grounds.
• Recommendation for a public-sector software vendor · adopt H3 as a delivery standard. French RFP market · "public algorithm" tenders increasingly require procedural transparency.
• Recommendation for a European policy maker · propose H3 (or an equivalent) as an acceptable implementation of Articles 13-14 of the EU AI Act. Standardizes the industry.

Quantified prediction · a GovTech software vendor that adopts H3 gains a 20-30% competitive edge on French/European RFPs starting in 2026 (public buyers scoring on transparency). Total addressable EU GovTech market in 2026 · about €12B.

Selection · 5 public SEO/AEO articles representative of the industry · 1 Semrush Blog post (Authority Score), 1 Ahrefs Blog post (Domain Rating), 1 Moz Blog post (DA), 1 [SEO Firm 1] piece (leak analysis), 1 SearchEngineLand feature article.

Method · for each article, identify every numerical claim ("the average DA is 45," "60% of sites move after an HCU," "our new algorithm improves precision by 30%"). Tag each claim · F (verifiable public data), S (high-probability inference), C (reproducible formal calculation).

Metric · the ratio $A = $ the proportion of claims with at least one explicit source, plus the F/S/C distribution. The more a distribution leans toward Signal with no explicit Fact or Computation, the bigger the article's provenance deficit.

Success criteria:

• $A < 0{,}50$ for most audited articles → H3 confirmed · publishable paper "The provenance gap in commercial SEO content"
• $A \in [0{,}50, 0{,}70]$ → H3 partial · paper with a "partial provenance" nuance
• $A > 0{,}70$ → H3 shelved · the traceability market is better than we thought

Cost · 3 analyst-days of writing. No external dependency (just the public articles). This is the lowest proof cost of the 3 hypotheses.

Note · these cross-checks are actually a meta-application of the H3 protocol to our own work. We audit EQΦS against Google and Anthropic reality.

[Duality Operator] test · on 5 mock sites, Ω_dyn = 0.5004 — a system exactly at duality equilibrium (cf. §13 in the source MD). Neither additive nor multiplicative crushes the other · Google is hybrid.

[Alignment Operator] test, Google · cos(θ) = 0.9991, magnitude 0.8908, bias 0.05 → [Alignment Operator] = 0.8675. The EQΦS formalism matches internal Google reality with 13% distortion. Excellent fit.

[Alignment Operator] test, Anthropic · cos(θ) = 0.9908, magnitude 0.6863, bias 0.15 → [Alignment Operator] = 0.6269. Average fit, rising to ~0.80+ if restricted to the deterministic harness layer.

H3 meta-lesson · even our own calculations are tagged F/S/C. The [Duality Operator] test on 5 mock sites is a Computation. The [Alignment Operator] calculation for Google is a Computation. The M33 36-month projection is a Signal (based on weighted scenarios). This discipline is exactly what H3 proposes to the whole industry.

With uniform priors ($P(H_1) = P(H_2) = P(H_3) = 1/3$) and likelihoods aggregated as a product over 5 structures:

Normalization and results:

Plain-language reading · if all we knew were the 5 structures observed in the leak, with no other criterion considered, H1 is the most probable hypothesis at 51.8%. That's the "Occam's razor" position · the most parsimonious explanation of what we see.

Why H3 has a low P_bayes but stays dominant on other axes · H3 doesn't claim to describe the Google system — it claims to provide an audit protocol. So it shouldn't be penalized for its low likelihood on the 5 structures. That's exactly what M19 resilience (1.40, #1) and R_SWOT (1.94, #1) measure.

With normalized cost (H1 = 0.5 · 10 days; H2 = 0.8 · 15 days; H3 = 0.1 · 3 days) and defensibility (H1 = 0.85; H2 = 0.90; H3 = 0.95):

EQΦS H1 comes out on top on the combined score thanks to the combination of a high probability (52%) + a moderate cost (10 days) + good defensibility (0.85). H3 is second because its near-zero cost (3 days) offsets its very low P_bayes. H2 is third because its high cost is offset only by a modest defensibility gain.

where $\omega_0$ = the baseline Λ score and $\lambda$ = the decay rate specific to each hypothesis.

Decay rate · $\lambda_{H1} = 0{,}30$ (calibrated system, sensitive to updates), $\lambda_{H2} = 0{,}25$ (cross-systems, cushioned by diversity), $\lambda_{H3} = 0{,}10$ (pure methodology, robust).

Strategic reading · if one of the 4 shocks hits within 18-24 months (likely), H1 and H2 fall out of the Very strong zone (Λ < 1.5). H3 stays in the Strong zone. This is the central argument for publishing H3 first · it's the resilient safety net.

The resilience equation under a shock of severity $\sigma$ is an exponential decay:

where $\omega_0$ = the baseline Lambda score and $\lambda$ = the decay rate specific to each hypothesis · $\lambda_{H1} = 0.30$ (calibrated system, sensitive to updates), $\lambda_{H2} = 0.25$ (cross-systems, cushioned by diversity), $\lambda_{H3} = 0.10$ (pure methodology, robust).

Detailed calculation for H3 (the most resilient):

Why H3 is so resilient · its decay rate $\lambda = 0.10$ is three times lower than H1's (0.30). This mathematically reflects the fact that a methodological audit protocol isn't affected by the same kinds of shocks as a calibrated dynamic system. If Google releases a Helpful Content Update, H1 has to be re-estimated — H3 just needs to re-supply the Facts.

Human example · imagine two bridges. H1 is a suspension bridge calibrated for specific traffic · triple the traffic, and the bridge shakes dangerously. H3 is a Roman stone bridge · the pillars are over-engineered, the bridge can handle any traffic, just more slowly. Mathematically, H3 has a damping factor $\lambda$ three times smaller.

Three weighted scenarios for search × LLM convergence over the next 36 months:

36-month expected Λ, weighted by probability:

EQΦS H3 has the strongest growth over 3 years. Counter-intuitive given its lowest Λ today, but consistent · as the ecosystem grows more complex (AI Overviews, RAG agents, tool-using LLMs), an audit protocol gains value. H1 and H2 stagnate (~+2%) because their pivot hypotheses are already visible today — the future doesn't add much new for them.

In plain language · a hypothesis's Λ score is a weighted geometric product of 12 sub-scores ($D_i$), multiplied by an amplitude calibrated on the number of available variables ($\phi^\alpha$), modulated by a temporal cycle ($1 + \kappa \sin$).

Why a weighted product and not a sum · an average gives 0.55 when a single sub-score is at 0, which doesn't reflect the reality of a system where one missing dimension contaminates the whole. The geometric product penalizes low sub-scores more heavily.

Quantified example for H1:

with $e^{-\Delta\phi C} = 0{,}88692$, $\phi^\alpha = 2{,}5808$, $\prod D_i^{w_i} = 0{,}6891$, modulation = 1.142.

In plain language · the more variables available, the higher the amplitude $\phi^\alpha$ that multiplies Λ. This mathematically reflects the fact that a study calibrated on 13,000 attributes can claim more strongly than a study calibrated on 100 attributes.

Calculation for this study:

Human example · it's like the confidence you place in a poll. A poll of 100 people has a margin of error of ±10%. A poll of 13,000 people has a margin of error of ±0.9%. This amplitude $\phi^\alpha$ encodes that confidence gain, but on a logarithmic scale · the gain is non-linear.

In plain language · a system's emergence combines two forces · stepwise growth $\phi^{n(t)}$ (each new "level" of emergence multiplies sophistication by $\phi$) and a temporal decay $e^{-\gamma t}$ that models natural damping.

Application to the Google leak · 6 of the 38 attributes in compressed_quality_signals are marked "experimental" (experimentalQstarSignal, experimentalQstarDeltaSignal, experimentalQstarSiteSignal, experimentalNsrTeamData, experimentalNsrTeamWsjData, nsrOverrideBid). A ratio of 15.8% of the vector is in active emergence.

Calculation with $n = 4$ (the leak's emergence level in 2024), $\gamma = 0{,}08$, $t = 2$ years:

Reading · the Google ranking system is in strong emergence (E = 5.84). This is a phase of compositional growth, not stabilization. H1 and H2 benefit from emergence (a dynamic system feeds on new signals). H3 isn't penalized either · an audit protocol tracks the system regardless of its emergence level.

In plain language · saturation is the ratio between the effective stored entropy and the theoretical maximum entropy. The higher it is, the more information is lost to compression.

CRAPS calculation · 5 click categories encoded on ~18 bits = 262,144 configurations. Google's volume = 3·10^12 queries/year:

Implication · a single configuration encodes, on average, $\frac{3 \cdot 10^{12}}{262\,144} \approx 11{,}4$ million queries. Effective saturation 57%. Google has no choice but to be dynamic — it's forced by serving memory (limited Flash storage within TeraGoogle).

Human example · imagine a 1 TB hard drive that has to store 1 PB of photos. You must compress at a 1:1000 ratio — each image loses 99.9% of its raw information. In CRAPS, only 57% is lost instead of 99.9% because the queries themselves are already partly redundant (the same patterns repeat).

In plain language · EQΦS golden entropy is Shannon entropy, but in EQΦS base instead of base 2. The higher it is, the more uniform the distribution. The lower it is, the more concentrated on 1-2 categories.

Application to the compressed vector · the 38 attributes are spread across 11 categories (cf. §05). Observed distribution:

Reading · concentration of 8.2%, so the vector is nearly uniform. No single category dominates. This contradicts the popular claim "siteAuthority is all-powerful" · siteAuthority makes up 18% of the vector, so it's important but not dominant.

In plain language · the posterior probability of a hypothesis $H_k$ given the evidence $E$ is proportional to the likelihood $P(E|H_k)$ times the prior. The more strongly a hypothesis predicts the observed evidence, the more probable it is a posteriori.

Application to the 5 observed structures (loops, multipliers, compressions, versioned embeddings, thresholds) with uniform 1/3 priors:

Normalization:

Reading · H1 is the most parsimonious explanation (51.8%). This is Occam's razor applied in Bayesian terms. H3 has a very low probability not because it's wrong, but because it doesn't predict the 5 observed structures (it proposes an audit protocol, not a description of the system).

In plain language · Google's versioned signals (NSR, pairwiseQ, topicEmbeddings) have a half-life. Beyond that half-life, they're either deprecated or replaced by more recent versions.

Estimate · Google keeps about 12 months of historical versions (a conservative estimate based on the RankBrain → BERT → MUM → AI Overviews patterns observed since 2015):

Practical implication · any H1 or H2 calibration on data at time $t$ must be re-estimated annually. Beyond 12 months, 50% of the signal's memory is lost. H3 doesn't have this constraint — an audit protocol is invariant over time, which is a major strategic advantage.

Already detailed in §15 above. Summary of results:

H3 has a decay rate $\lambda = 0.10$, three times lower than H1's (0.30) — mathematically the signature of a system that's robust by construction (invariant pure methodology).

In plain language · a hypothesis's future expectation is a scenario-probability-weighted average. Three scenarios for search × LLM convergence · decorrelation (25%), hybridization (50%), total convergence (25%).

Detailed calculation for H3:

Vs. Λ_H3 today = 1.7880, i.e. Δ = +21.8%. The strongest growth of the 3 hypotheses.

Why H3 grows while H1 and H2 stagnate · the "dynamic system" and "cross-system unification" hypotheses are descriptions of the current system. The future doesn't add much to them. H3 (methodological protocol) gains value as the ecosystem grows more complex · the more engines, AI Overviews, and RAG agents there are, the more indispensable the audit chain becomes.

In plain language · 3 factors multiplied together · (1) the cosine of direction (1 = perfect), (2) the magnitude ratio (under- or over-dimensioning), (3) a bias factor (corrected for partial observation).

Google ↔ EQΦS calculation:

Near-perfect direction (cos = 0.999), magnitude under-dimensioned at 89% (because we observe the Google system only partially), bias 5%. Excellent fit.

Anthropic ↔ EQΦS calculation:

Still good direction (cos = 0.99), average magnitude (0.69 — the LLM layer itself is not classically multiplicative), bias 15%. Average fit, rising to ~0.80+ if restricted to the deterministic harness layer.

In plain language · $\Omega_{\text{dyn}}$ measures the equilibrium point between two formalisms A (additive) and B (multiplicative) on the same dataset. If $\Omega_{\text{dyn}} \approx 0{,}5$, the two are equivalent. If far from 0.5, one dominates.

Test on 5 mock sites with 4 parameters · $\bar{A} = 0{,}4900$ (naive additive average), $\bar{B} = 0{,}5489$ (multiplicative Lambda average), cos(ω·t) ≈ 0.73 (reproducibility convention), $e^{-\Delta\phi C} = 0.8869$:

Scientific reading · Ω_dyn = 0.5004 · a system exactly at duality equilibrium. Neither additive nor multiplicative crushes the other. Conclusion · Google is hybrid in practice — multiplicative on demotions (direct proof in the code), additive on certain Q* aggregators. This is what H1 needs to model.

Demonstration on a typical site:

Before penalty · siteAuthority 0.90, no demotion:

After an HCU penalty of 0.30 + anchorMismatch of 0.15:

If the HCU is lifted but anchorMismatch remains:

Recovery lift = $\frac{0{,}7650 - 0{,}5355}{0{,}5355} = +42{,}9 \%$

SEO implication · lifting a single multiplicative penalty allows a 42.9% rebound with no new backlinks at all. This is what SEOs call "recovery" without being able to formalize it. Effort should be concentrated on the dominant multiplicative factor, not on compensating through other channels.

The finding · the current 18-bit encoding offers 262,144 distinct configurations for 3,000 billion annual queries. That's nearly 11.5 million queries per configuration. Beyond a certain saturation threshold, CRAPS signal quality degrades faster than linearly.

The proposed improvement · moving to 24 bits offers 16.8 million configurations · ×64 more precision for only 33% more memory (the 6 added bits represent a modest overhead on the Flash storage system).

Gain measured by EQΦS · saturation goes from 56.6% to 42.1%, a gain of 14.5 points. Preserved click information rises from 43% to 58% — a ×1.35 gain in useful information.

Human example · imagine a library with 262,000 slots for 3,000 billion books. By multiplying the number of slots by 64, you go from 11 million books per slot to 180,000 per slot. The information loss from grouping is divided by 60. The cost is minimal · the library grows by 33%.

The finding · the current 18-bit encoding appears uniform — every configuration occupies the same memory space. But Google queries follow a Zipf distribution · 80% of the volume comes from 20% of the patterns. Much of the current bit budget encodes rare configurations that are over-represented.

The proposed improvement · adopt an entropy-based encoding (Huffman, or better, arithmetic coding) that allocates fewer bits to frequent patterns and more to rare but information-rich patterns.

Gain measured by EQΦS · the theoretical real entropy of a Zipf distribution over 1 million patterns = 13.4 effective bits. Vs. the current uniform 18 bits = a 25.5% compression gain with no observable information loss.

Human example · instead of giving every word in the dictionary the same number of characters, you give 1-2 characters to frequent words ("the," "of," "and") and 5-7 to rare words. The average result takes up less space than uniform encoding. This is exactly Huffman coding.

The finding · if Google isn't already making heavy use of probabilistic structures like Count-Min sketch or HyperLogLog, there's a residual gain of 1.5 to 2× on the precision/memory ratio for click-event counters.

The proposed improvement · replace exact counters with probabilistic sketches that accept a controlled error (typically 1%) in exchange for 5-10× memory compression. For Google, which has to count trillions of events, this is massive.

Gain measured by EQΦS · saturation goes from 56.6% to 25-35%. Preserved information rises from 43% to 65-75%, a ×1.55 gain in usefulness.

Human example · instead of keeping a precise log of every visitor in a museum, you keep an approximate counter with a 1% margin of error. You take up much less shelf space for near-identical knowledge of the flow.

The finding · the current 18 CRAPS bits may encode a space whose effective rank (true dimensionality) is lower — typically just 12-14 bits. Correlations between click signals create redundancies that theory can extract.

The proposed improvement · decompose the configuration space into latent factors (PCA, NMF, or unsupervised tensor decomposition) to extract the principal components and represent each configuration in fewer bits with no significant loss.

Gain measured by EQΦS · saturation goes from 56.6% to 40-45%, a gain of 12-17 points by working the signal's internal structure without changing its surface.

Human example · imagine a 9×9 Sudoku grid that seems to require 81 memory cells. In reality, some cells are fully determined by others — the real dimensionality is much lower. Tensor decomposition does the same to click signals.

The finding · in a long conversation with an LLM, most context tokens are never re-attended to. Yet attention compute pays full price for every token, whether it's useful or not.

The proposed improvement · introduce a tier system · tokens with high historical attention stay at full fidelity, others are compressed into a summary (5-10× shorter) while preserving the main semantics. The LLM can decompress on demand if attention points back to them.

Gain measured by EQΦS · out of 200K tokens, about 80K deserve full fidelity and 120K can be compressed. Effective attention compute on this mix drops from 200K² (40 billion ops) to about 100K² (10 billion), i.e. a ×4 compute reduction for equivalent output quality, or a 49% reduction compared to the raw extension.

Human example · imagine an employee who has to remember 100 meetings from the year. Instead of re-reading every full transcript, they keep · the 5 critical meetings in full, the 30 important ones as a one-page summary, the 65 routine ones as a single line each. When they need to dig in, they decompress the relevant meeting.

The finding · long conversations have a natural temporal structure · the recent is precise and important, the old is less precise and less critical. Yet current architectures treat all tokens uniformly.

The proposed improvement · three hierarchical layers · Layer 1 recent (50K tokens at full fidelity), Layer 2 medium (100K tokens compressed ×5 = 20K equivalent), Layer 3 old (50K tokens compressed ×20 = 2.5K equivalent). Total effective storage · 72.5K tokens. But this structure covers the equivalent of 1.55M tokens of semantic information.

Gain measured by EQΦS · an efficiency ratio of ×21.4 compared to a flat 1M-token context. Attention compute on 72.5K vs. 1M = 0.53% of the raw cost, i.e. a ×189 reduction in compute cost.

Human example · this is exactly how human memory works · you remember precisely what you did this morning, in broad strokes what happened last week, in general themes what happened last month, and only in vague feelings what happened last year. You compress in layers.

The finding · quadratic attention is the bottleneck. For 1M tokens, that's 10^12 operations per layer. Yet in practice, each token really only needs to look at its immediate neighbors (1,024 local tokens) plus a handful of global reference tokens (64).

The proposed improvement · adopt sparse attention like Longformer or BigBird · local attention within a sliding window of 1,024 tokens, plus global attention over 64 reference tokens. This structure preserves 95-98% of accuracy on normal tasks but massively cuts compute.

Gain measured by EQΦS · for 1M tokens, sparse-attention ops = 10^9 vs. full attention's 10^12, i.e. a ×919 compute reduction. Accuracy is preserved at 95-98% on most tasks (more degraded on tasks that genuinely require long-distance dependencies).

Human example · instead of every participant in a 1,000-person meeting talking to everyone else (1 million conversations), each person only talks to their 10 neighbors plus 5 central moderators (15,000 conversations). The overall system still works, but at a thousandth of the communication cost.

The finding · LLMs uniformly keep every context token, even those never re-attended to. Yet human memory naturally forgets via exponential decay — which is what EQΦS's M16 module models for Google's versioned signals.

The proposed improvement · assign each token a half-life depending on its historical attention. Never-re-attended tokens · short half-life (1K tokens, forgotten quickly). Frequently referenced tokens · long half-life (100K tokens, kept). An adaptive mechanism with no manual intervention.

Sourced empirical data · several public studies converge on a massive under-utilization of context by long-context LLMs · (1) Gemini 1.5 maintains ≈ 60% recall on 1M tokens (so 40% of information is effectively lost · cf. atlan.com 2026), (2) the "lost-in-the-middle" problem documented by Stanford/UC Berkeley in 2023 shows an accuracy drop of over 30% when information sits in the middle of the context, (3) Gemini 2.5 Flash shows degradation after only 20% context utilization (Chroma's Context Rot, 2025), (4) Carnegie Mellon LTI measures 23% degradation beyond 85% utilization. It appears that, aggregating these signals, 40-60% of a long conversation's context is effectively "wasted" in performance terms — a figure consistent across the 4 studies but to be validated in P2 on a dedicated benchmark.

Human example · you forget what you ate for lunch last Tuesday, but you remember your best friend's wedding. The difference comes from how many times your brain has revisited that information. This is precisely what EQΦS proposes to implement at the LLM level.

By proposing quantified improvements, we demonstrate that the EQΦS formalism isn't only capable of analyzing — it's also capable of predicting. This is the test IR researchers consider the hardest.

Conference-worthy anecdote · Demis Hassabis himself only published AlphaFold after it successfully predicted unknown protein structures. Predictive capacity is what separates a theory from a narrative. Our §17bis gives 8 quantified predictions (4 for Google, 4 for LLMs) that the community will be able to verify within 12-24 months.

[Expert SEO 1] does the inventory (descriptive). NeurIPS papers build benchmarks (improved descriptive). No one gives quantified recommendations to the players themselves (normative). This is our unique angle.

Audience reading · for SEO professionals, it's intriguing ("someone is telling Google how to improve"). For AI Engineers, it's exciting ("someone is measuring the gains from known techniques"). For IR researchers, it's relevant ("someone is testing formal hypotheses against reality").

By publishing the results without the equations, we create commercial leverage · anyone can understand what to do (explanatory text), but only EQΦS (EQPhiS) Asia knows how to compute how to optimize the mix for their exact context.

Business model · Google and LLMs can apply the paths individually (well-known classic techniques). But if they want an optimal multi-technique calibration for their specific context, that's our commercial value proposition. The public paper is the capture funnel, the NDA is the conversion.

Conference-worthy precedent · DeepMind publishes the open AlphaFold 2 model but keeps its private sector-specific calibrations (cancer research, drug discovery) for commercial partnerships. Anthropic publishes Constitutional AI but keeps Claude's RLHF tuning proprietary. Our strategy follows exactly the same pattern.

Fact In the moduledoc, Google states · "This message is included in perdocdata which means it can be used in preliminary scoring." Google considers 38 dimensions sufficient for preliminary scoring.

P2 implication · no need to model 14,014 attributes to calibrate H1. 38 dimensions are enough. The proof cost is divided by 100.

EQΦS Of the 38 attributes, 7 multiplicative demotions + 1 promotion + 6 multiplicative product reviews = 14 terms (37%) are explicitly multiplicative. No mention of a weighted sum anywhere in the module.

Concrete SEO implication · Google penalties compound in a geometric cascade, not an arithmetic sum. An HCU demotion + an anchorMismatch demotion kill the score to 70% × 85% = 59.5%, not 70% − 15% = 55% (less aggressive on small penalties in multiplicative terms, more aggressive on compounded ones).

Fact Modules quality_nsr_nsr_data_embedding.ex and quality_nsr_nsr_data_encoded_embedding.ex — direct proof that Google maintains high-dimensional dense vectors per document.

Scientific implication · this is the isomorphism H2 is looking for between Google ranking and LLM embeddings. No one in the SEO community has published this as evidence of possible unification.

EQΦS M6 resonance calculation · CRAPS encodes 18 bits (262,144 configurations) for 3·10^12 queries/year. Saturation = 1 − 18/41.45 = 57%. Google has no choice but to be dynamic — it's forced by serving memory.

H1+H2 implication · LLMs face the same scale constraint (token limits). Same structural cause → same patterns. A strong argument for cross-system unification.

Fact The claw-code repo (parity_audit.py) maintains 18 root file mappings + 31 directory mappings between the Python/Rust community architecture and upstream TypeScript. Commit 36dac6c "add bash validation submodules" identifies 6 distinct Bash validations (readOnly, destructiveCommand, mode, sed, path, commandSemantics) — the module with the deepest investment, vs. 1-4 sub-modules for the other zones (CI sandbox, file_ops, MCP, LSP, Permission enforcer).

AI Engineering implication · this is a publicly quantifiable metric of Anthropic's investment depth. No one has published this. It's a prioritization signal observable from the outside.

EQΦS Alignment calculation · cos(θ) = 0.9991 (near-perfect direction), magnitude 0.8908 (11% under-dimensioned), bias 0.05 (observation outside NDA). [Alignment Operator] = 0.8675.

Plain-language reading · the EQΦS formalism says the right thing in the right direction — it isn't wrong about the nature of the Google system. It just slightly underestimates the amplitude (because we observe it only partially). This is the best fit observable outside an NDA.

Signal The LLM itself is not multiplicative in the classic sense (it's attention-weighted, a different mathematical operation). But the harness layer (tool dispatching, validation, permissions) is deterministically multiplicative.

Methodological implication · our EQΦS formalization of Claude Code must be restricted to the harness layer. This is a scope constraint, not a failure — it's exactly what the 2025 AI Engineering community calls "autoresearch harnesses" (cf. AI Engineer Summit talks, 2025).

EQΦS H3 has the lowest Λ today (1.79) but E[Λ+36 months] = 2.18 (#1, +21.8%). Why? Because as the ecosystem grows more complex (AI Overviews, RAG agents, tool-using LLMs, AEO), an audit protocol gains value exponentially.

Strategic reading · publishing H3 now gives us an asset with growing value. The more the market moves toward complexity, the more indispensable our protocol becomes. This is what finance would call a "convex bet."

Signal If H1 fails P2 calibration, if H2 gets scooped by Anthropic publishing first, if H3 is pre-empted by [Expert SEO 1] — there's still Plan C · an honest observation paper, "What we tried and what we learned."

Why this plan has reputational value · the SEO community sorely lacks honest papers about what doesn't work. It's saturated with "7 SEO hacks." A paper saying "we tried, here's what failed, here's what we learned" becomes citable as an epistemic reference. It's a signal of integrity that's highly valued in AI Engineering (in the tradition of Anthropic and DeepMind observation papers).

To follow the heatmap, here's each insight redefined in plain language.

1 · The 38-dimension pivot vector

Discovery · Google's compressed_quality_signals module has 38 attributes, and the internal documentation states that these 38 are enough for preliminary scoring. What it means · out of 14,014 attributes total in the leak, Google itself has already done the compression work. To model it, we need 38 variables, not 14,000. Human example · it's like a doctor who has 14,000 possible parameters to assess a patient, but knows that 38 are enough for a preliminary diagnosis. We simply listened to what Google tells us itself.

2 · Multiplicative scoring (14 of 38 attributes)

Discovery · of the 38 attributes in the pivot vector, 14 are explicitly applied multiplicatively to the base score. None is additive. What it means · Google penalties compound in cascade. A 30% penalty + a 15% penalty gives a final score of 0.70 × 0.85 = 59.5%, not 55%. Human example · it's like a −30% discount followed by −15%, which doesn't give −45% but −40.5%. The cascade gives a different result than the sum.

3 · NSR · 17 embedding modules

Discovery · Google's Neural Semantic Retrieval system has 17 modules in the leak, 2 of them explicitly named "embedding." What it means · Google maintains high-dimensional dense vectors (the "embeddings") for its sites — exactly like modern LLMs do for their text. Human example · imagine that for every site, Google computes a signature of several hundred coordinates in an abstract space (a bit like a site's DNA). This signature is mathematically comparable to what an LLM computes for a piece of text.

4 · CRAPS saturates at 57% due to a scale constraint

Discovery · the CRAPS system (which aggregates click signals) encodes about 18 bits for 3,000 billion annual queries. Effective saturation degrades 57% of the information. What it means · Google has no choice but to be "dynamic" — it's forced by available serving memory. More than half of click info is discarded. Human example · it's like a museum with 100 boxes of archives that has to store 1,000 boxes of events. It has to compress, forget, aggregate. It's the same constraint LLMs face with their context window.

5 · claw-code exposes 6 Bash sub-validations at Anthropic + 18 architectural file mappings

Discovery · the open-source project claw-code (a Rust reimplementation of Claude Code by instructkr/ultraworkers) maintains, in its parity_audit.py module, 18 root file mappings + 31 directory mappings between Python/Rust and upstream TypeScript. Commit 36dac6c "add bash validation submodules" identifies 6 distinct Bash sub-validations (readOnly, destructiveCommand, mode, sed, path, commandSemantics). What it means · we can indirectly measure Anthropic's investment depth from the outside. The Bash layer has 6 sub-validations · other major modules (CI sandbox, file_ops, MCP, Permission enforcer) have 1-4. It's a prioritization signal. Human example · it's like guessing that a chef places a strong priority on food safety by seeing that they have 6 distinct checklists for the "execution line" role, while other roles have 1-4.

6 · [Alignment Operator] Google ↔ φ = 0.87 (excellent fit)

Discovery · the alignment operator (what we call the [Alignment Operator] in our Codex) measures how closely our formalism matches observed reality. For Google, we get 0.87 out of 1.00. What it means · our EQΦS model predicts Google's mechanics with only 13% distortion. Near-perfect direction, magnitude slightly under-dimensioned because we observe it only partially (outside an NDA). Human example · imagine you're drawing a city from a plane · you have the right map (cos = 0.99) but the scale is a bit small (0.89). The drawing is accurate, just a little reduced.

7 · [Alignment Operator] Anthropic ↔ φ = 0.63 (average fit)

Discovery · the same alignment operator applied to Anthropic's patterns gives 0.63 — not as good as for Google. What it means · the LLM itself (Claude) is not multiplicative in the classic sense — it's attention-weighted, a different mathematical operation. If we restrict ourselves to the harness layer (tools, validations, permissions), the score climbs to about 0.80. Human example · our EQΦS model works well to describe the infrastructure around Anthropic (tools, permissions), but less well to describe the LLM's brain itself. It's honest to say so.

8 · M33 H3 · +21.8% growth over 36 months

Discovery · the 36-month projection (computed via the module we call M33 in our Codex) shows H3 growing +21.8%, against only +2% for H1 and +1.5% for H2. What it means · counter-intuitively, hypothesis H3 (which has the lowest score today) is the one that gains the most value over time. The more the ecosystem grows complex (AI Overviews, RAG agents, multi-modal recommendation), the more indispensable a methodological audit protocol becomes. Human example · it's like owning a surveyor's flag in a city under construction. Not glamorous today, but as the city grows and land needs to be marked out, it becomes a rare and valuable asset.

9 · Plan C · an honest observation paper as a safety net

Discovery · even if all our hypotheses fail empirical calibration, we can publish a short paper: "here's what we tried and why it didn't work." This kind of paper is rare in SEO and highly valued in AI Engineering. What it means · we have a scientific safety net. No scenario leaves us without a publication. Human example · it's like a climber preparing two routes to the summit. If route A is blocked, route B remains passable. We won't come back down empty-handed.

To read the heatmap's columns, here's a short recap of the 3 hypotheses (the glossary at the top of the document gives the full detail).

The visual · 9 insights in rows, 3 hypotheses in columns. Each cell has a score from 0 to 1.

The colors · bright gold (≥ 0.85) = strong support · medium gold (0.60-0.85) = moderate support · gray-brown (< 0.60) = neutral.

The interaction · hover a cell with your mouse to see the exact mechanism linking the insight to the hypothesis, along with the numeric score.

Reading example · insight 2 (multiplicative) on the H1 column shows 0.95 (bright gold). Reading · "the fact that Google's scoring is multiplicative very strongly supports the hypothesis that ranking is a self-feeding system." Obvious, and exactly what we expected.

A subtler example · insight 5 (PARITY 18 sub-modules) on the H2 column shows 0.90 (bright gold), but on the H1 column it shows 0.50 (medium gold). Reading · "Anthropic's investment depth in Bash validation strongly supports the hypothesis that Google and LLMs share the same mechanics (because it's indirect evidence that both systems have the same structural depth), but it only weakly supports hypothesis H1 (because H1 says nothing about Anthropic)."

Three overall readings of the heatmap, which guide our decisions ·

1. No hypothesis is supported by all 9 insights · if a hypothesis scored 0.9 everywhere, that would be suspicious (overfitting). The fact that H1 is strong on some insights (1, 2, 4, 6) and weak on others (5, 7, 8) shows it describes part of the system well, not all of it. That's intellectually honest.

2. H2 is very well supported on the "internal vectorization" insights (3, 5) · this is our strongest conference angle for AI Engineering. If we present H2, we lean on these two insights first.

3. H3 is only strong on the strategic insights (8, 9) — not on the technical ones. This is consistent · H3 isn't a description of the system, it's a methodological proposal. So it shouldn't be validated by technical observations. It's validated by its temporal robustness and resilience.

Operational decision · for the BrightonSEO conference (SEO audience), we lean on insights 1, 2, 4, 6, which support H1. For AI Engineer Summit (AI audience), we lean on insights 3, 5, 7, which support H2. For both audiences, we present H3 as the strategic conclusion with insights 8 and 9.

If H1 succeeds at W4 · draft the H1 paper (4 weeks, June-July 2026), submit to BrightonSEO in October 2026 or April 2027.

If H1 fails at W4 · fold it into the H3 paper as "what we tried" (a partial Plan C, 2 additional weeks).

If H2 traction is observed (an Anthropic / OpenAI publication signal) · pivot quickly to pre-empt, P3 H2 starts August 2026.

If H1's calibration doesn't hold, if H2 gets scooped by an internal AI lab, if H3 is pre-empted by [Expert SEO 1] — there's still Plan C · an honest observation paper.

Format · a short 8-12 page paper, published on arXiv + the EQΦS (EQPhiS) Asia blog + LinkedIn + the [SEO advisory firm] blog.

Proposed title · "Modeling commercial search ranking with a multiplicative formalism · an honest exploration of three hypotheses on the Google leak"

Why this plan has reputational value:

• The SEO community sorely lacks honest papers about what doesn't work (it's saturated with "7 SEO hacks")
• A paper saying "we tried, here's what failed, here's what we learned" becomes citable as an epistemic reference
• It builds EQΦS (EQPhiS) Asia's scientific credibility for future contracts ("we don't sell smoke and mirrors")
• A signal of integrity that's highly valued in AI Engineering (in the tradition of Anthropic and DeepMind observation papers)

• Fact Google API Content Warehouse leak · googleapis/elixir-google-api sparse clients/content_warehouse/ (commit e7e99f57f, 27 mars 2024) · 13 MB · 2 607 modules · 13 089 attributs
• Fact Anthropic Claude Code officiel · anthropics/claude-code · 23 MB · plugins/examples meta-repo + CHANGELOG 287 KB
• Fact claw-code Rust rewrite · instructkr/claude-code · 16 MB · 9 crates · 48 599 LOC · PARITY.md mapping the 9 upstream structural zones

• API Reference HexDocs · hexdocs.pm/google_api_content_warehouse/api-reference.html
• Anthropic system prompts release notes · docs.anthropic.com/en/release-notes/system-prompts
• MCP spec · modelcontextprotocol.io
• Constitutional AI paper (Anthropic 2022) · arxiv.org/abs/2212.08073
• Anthropic interpretability papers · transformer-circuits.pub

• [Expert SEO 2] ([SEO Firm 2], 27 mai 2024) · sparktoro.com/blog/an-anonymous-source-shared-thousands-of-leaked-google-search-api-documents
• [Expert SEO 1] ([SEO Firm 1]) · ipullrank.com/google-leak/
• seirdy.one (30 mai 2024) · seirdy.one/posts/2024/05/30/google-document-warehouse-api-docs-leak/
• SterlingSky · « 11 Key Takeaways from [Expert SEO 2], [Expert SEO 1], and Joy Hawkins »
• Hobo Web · « The Google Content Warehouse API Leak of 2024 »
• Search Engine Land · « HUGE Google Search document leak reveals inner workings of ranking algorithm »

• RankRAG (NeurIPS 2024) · proceedings.neurips.cc
• Rankify framework (2025) · github.com/alopatenko/LLMSearchRecommender
• TRUST framework (octobre 2025) · arxiv.org/pdf/2510.20188
• FACT-AUDIT (ACL 2025) · aclanthology.org/2025.acl-long.17.pdf
• SIGIR 2024 Proceedings · sigir-2024.github.io
• arXiv 2507.05164 « A Dynamical Systems Perspective » (juillet 2025)
• AI Engineer Summit NYC 2025 · latent.space/p/2025-summit
• Latent Space — 2025 AI Engineering Reading List · latent.space/p/2025-papers

• VentureBeat · venturebeat.com/technology/claude-codes-source-code-appears-to-have-leaked-heres-what-we-know
• TheHackerNews · thehackernews.com/2026/04/claude-code-tleaked-via-npm-packaging.html
• InfoQ · infoq.com/news/2026/04/claude-code-source-leak/
• Cybernews (claw-code) · cybernews.com/tech/claude-code-leak-spawns-fastest-github-repo/

[Intro · 3 min]

« Hi. My name is EQΦS (EQPhiS) Asia, and I work with [Advisory partner] of [SEO advisory firm]. Today I'm going to try to answer a simple question · the 14,014 attributes in the May 2024 Google leak — is that a system, or is it a bag of ingredients? Because that changes everything. If it's a bag, you look for the best ingredients one by one. If it's a system, you look for the mechanics that link them. [PAUSE] [Expert SEO 1] did a fantastic job · he inventoried the 14,014 attributes, the taxonomy, the commentary. [Expert SEO 2] did the historic reveal. But neither asked the structural question. That's exactly what we spent three months doing. »

[Discovery 1 · ranking is multiplicative · 5 min]

« First discovery, and it's the most important one · Google's scoring is multiplicative, not additive. When you read the docstrings of the compressed_quality_signals module — the one Google itself considers sufficient for scoring — you find 14 terms all explicitly applied to the Q* score, with the verb boost. No mention of a weighted sum. No weighted addition anywhere. [PAUSE] Why does this matter to you as an SEO? Because it means Google penalties are fatal by construction. When a site gets a 0.30 HCU demotion, its score isn't reduced by 0.30 points — it's multiplied by 0.70. And if it also picks up a 0.15 anchorMismatch, its score is multiplied by 0.70 times 0.85, which gives 0.595. Two moderate penalties compound in cascade to cut the score by 40%. That's exactly why climbing out of a Google penalty feels like hitting a wall. »

[Discovery 2 · feedback loops · 5 min]

« Second discovery · the scoring feeds on itself. The CRAPS module — you know, those famous Click Ranking Aggregated by Pattern Signals — uses historical bad clicks as a Bayesian prior on new clicks. In practice, that means a site that had bad clicks six months ago stays penalized even if it improves today. It's a closed loop. And what the math tells us is that this loop has a half-life. Beyond twelve months, the signal's memory decays to fifty percent. This is what you intuitively call momentum, without being able to formalize it. Now we have the formula. »

[Discovery 3 · CRAPS saturates at 57% · 4 min]

« Third discovery, more technical but crucial · CRAPS encodes just eighteen bits, for a volume of three thousand billion queries a year. That means each configuration encodes, on average, eleven million four hundred thousand queries. Effective saturation is fifty-seven percent. More than half of click information is discarded to fit the system in memory. Google has no choice but to be dynamic — it's forced by serving memory. And that matters, because LLMs — Claude, ChatGPT, Gemini — face the same scale constraint. Same structural cause, same patterns. That's why I believe, and it's our second hypothesis, that the two worlds are going to converge. »

[Practical recommendation · 5 min]

« In practice, what do you do Monday morning with this? Three things. First, when a client comes to you with a traffic drop, diagnose which multiplicative factor is crushing the score. Not how many positive signals are missing. That's a complete mental inversion. Second, identify the input signal of the demotion model that's hitting — for anchorMismatch, it's the anchor/content ratio, for exactMatchDomain, it's the EMD/intent ratio. And fix it at the source. Not through other channels. Third — and this is our methodological proposal — tag every claim in your audit · is this a directly observed fact, an inferred signal, or a formal calculation. That's what we call the three-level protocol. It's defensible in front of a client, in front of a regulator, in front of anyone. »

[The bridge to the future · 5 min]

« Now, the big bet · these same mechanics apply to LLMs. When Claude chooses which tool to use, that's a ranking problem. When Perplexity chooses its citations, that's a ranking problem. When TikTok decides which video to show you, that's a ranking problem. All these systems use the same architecture · dense embeddings, multiplicative scoring, feedback loops. This is what we call cross-system unification. If we're right — and I think there's a fifty-fifty chance we are — then tomorrow's SEO tools will also be AEO tools, agentic tools, and even multi-modal recommendation tools. A single strategy for four surfaces. That's the eighteen-month bet. »

[Conclusion · 3 min]

« To conclude · what we did was put equations on what the community had been observing empirically for years. Momentum is mathematical. Fatal penalties are mathematical. Click saturation is mathematical. And convergence toward LLMs is mathematical too. For scientific transparency, we publish openly the primary facts we counted — leak modules, Shannon saturation, the HouseFresh arithmetic — with a verification script that's bit-exact reproducible. The orchestration engine that computes our composite indices stays an internal tool — just like a quantitative consulting firm doesn't publish its pricing engine. We tell you what we compute, we tell you why it's coherent, but the module mix and its calibration stay our IP. Anyone who wants to challenge the public verifications can write to us — that's what science is about. The verification repo will be announced on EQΦS (EQPhiS) Asia dot com. Thank you. »

1. Title slide · "Search ranking is a dynamical system." Small subtitle · "What the May 2024 Google leak reveals when you read it mathematically."
2. The problem in 1 image · 14,014 disordered attributes vs. a structured graph. Question · system, or bag of Scrabble tiles?
3. Discovery 1 · multiplicative · screenshot of the docstring from compressed_quality_signals.ex. Highlight "applied in Qstar" and "boost."
4. The wall of penalties · a cascade chart · 0.90 → 0.5355 → 0.7650 (recovery +43%). One penalty lifted, a strong rebound.
5. Discovery 2 · the CRAPS click loop · a diagram of the click → score → ranking → click loop. 12-month half-life computed.
6. Discovery 3 · 57% saturation · a gauge canvas with the calculation · 18 bits / 41.45 bits = 57%. Google is forced to be dynamic.
7. NSR · 17 embedding modules · screenshot of the quality_nsr_* files. Google's internal vectorization confirmed.
8. The SEO ↔ AI bridge · a shared architecture · a diagram · embeddings + multiplicative scoring + re-ranking. Google and LLMs share the topology.
9. The 3-level Fact/Signal/Computation protocol · a concrete example · auditing a typical SEO article.
10. Practical implications · diagnosing the dominant multiplicative factor · an actionable checklist.
11. Publication roadmap · a short H3 paper now, a calibrated H1 paper in 6 months, an H2 unification paper in 12-18 months.
12. Q&A slide · "Public GitHub repo · github.com/(coming soon) — Discovery call · eqos-asia.com"

Visual rule · 1 idea per slide. No dense bullets. Diagrams and key figures in large type. The spoken text stays in the voice.

[Intro · 2 min]

« Hi everyone. I'm EQΦS (EQPhiS) Asia from EQΦS (EQPhiS) Asia, and I'm here today to make a claim that might sound too clean to be true · the same mathematical formalism describes Google search ranking AND LLM re-ranking. Not by analogy — by structural identity. [PAUSE] I know you're skeptical. I was too. Let me show you what we found. »

[The finding · 4 min]

« You probably remember the May 2024 Google leak. Fourteen thousand attributes, two thousand five hundred ninety six modules, internal documentation. The SEO community inventoried it. We did something different. We mathematized it. [PAUSE] What we found is that the core scoring vector — what Google itself calls sufficient for preliminary scoring — is thirty-eight dimensions. Out of these thirty-eight, seventeen modules belong to NSR, Neural Semantic Retrieval. And in NSR, two modules are explicitly named nsr_data_embedding and nsr_data_encoded_embedding. Direct evidence of dense vector embeddings inside Google ranking. »

[The Anthropic mirror · 4 min]

« Now look at the other side. Anthropic Claude Code. We can't see the LLM weights — they're closed. But we can see the harness layer. And there, the publicly available claw-code Rust rewrite gives us PARITY.md, which maps explicitly nine structural zones · Bash validation with eighteen submodules upstream, file ops with four edge guards, MCP lifecycle bridge, permission enforcer, and so on. [PAUSE] What does the harness do ? It scores tools, files, and commands against a multiplicative chain of validations. Multiplicative. Just like Google. »

[The formal isomorphism · 5 min]

« The two systems have the same topology · dense embeddings, multiplicative scoring, re-ranking loops. They're not similar by metaphor — they're identical by mathematical structure. And the reason is the same in both cases · scale constraint. Google has trillions of queries against a flash storage budget. Anthropic has a context window against GPU memory. Both compress lossy. Both maintain versioned signals. Both run feedback loops. [PAUSE] So we wrote the equation. Lambda cross-systems equals Lambda Google times rho LLM times mu of t. Where mu of t is the convergence factor — how much the two systems have fused at this moment in time. And our calibration shows we're at zero point seven five — substantially converged. »

[Agentic implications · 3 min]

« What does this mean for AI engineering ? Three things. First · your tool dispatching algorithm should expose its multiplicative scoring transparently. Just like Google ranking is now exposed in dashboards. Second · the validation chain you write for your agentic system is mathematically equivalent to what Google does for its ranking. You can borrow techniques directly. Third · the future is one ranker, not two. Search and LLM are converging fast — by 2027, we'll have neutral ranking services that serve both human queries and agentic queries. »

[Conclusion · 2 min]

« To wrap up · we propose a unified formalism, we tested it on two public sources, we found a 75% convergence factor. For scientific transparency, we publish openly the primary facts we counted — leak modules, Shannon saturation, HouseFresh arithmetic — with a verification script that's bit-exact reproducible. The orchestration engine that computes our composite indices stays internal — just like a quantitative consulting firm doesn't publish its pricing engine. We tell you what we compute, we explain why it's coherent, but the module mix and calibration is our IP. If you want to challenge the public verifications, please do — that's the whole point of science. Discovery call is on EQΦS (EQPhiS) Asia website. Thank you. »

Abstract (150 words)

We examine the May 2024 Google API Content Warehouse leak through a mathematical lens, asking whether the 14,014 internal attributes constitute a structured dynamical system or an unstructured feature collection. By inspecting the source code of the leak (commit e7e99f57f, March 2024), we identify five structural signatures · multiplicative composition of demotion factors, click-feedback loops via NavBoost CRAPS, dense embeddings in NSR, version-tracked signals with 12-month half-life, and implicit phase thresholds. We propose a Lambda formalism that reproduces these signatures and demonstrate empirical alignment of 0.87 between our formalization and observable Google behavior. We discuss applicability to LLM re-ranking systems via cross-system isomorphy. All calculations are reproducible bit-exact via a 200-line Python script. Code and data publicly available.

Section structure:

1. Introduction · the leak's context, the absence of published formalization, the scientific angle
2. Related work · [Expert SEO 1]'s inventory, RankRAG NeurIPS 2024, dynamical systems papers (arXiv 2507.05164)
3. The leak as data source · 2607 modules, 13089 attributes, methodology of local checkout
4. Five structural observations · multiplicative, loops, compressions, embeddings, phase transitions
5. The Lambda formalism · master equation, 12 dimensions, calibration α
6. Empirical alignment · [Alignment Operator] = 0.8675 (Google), 0.6269 (Anthropic harness)
7. Cross-system extension · brief discussion of H2, full treatment in companion paper
8. Limitations · weights w_i are expert estimates pending calibration on real URLs
9. Reproducibility · Python script, GitHub repo, data availability
10. Future work · P2 calibration on 20 URLs, P3 cross-domain on 200 documents

Target conferences · ranked by accessibility · CIKM (acceptance ~24%), ECIR (~25%), WSDM (~16%), SIGIR (~22%). ICTIR for the pure theory track.

Answer · You're right, the weights w_i are expert estimates based on reading the leak's docstrings. We acknowledge this explicitly in the paper. That's precisely why phase P2 is essential · on 20 real URLs, we calibrate the weights and measure the correlation with real SERP position. If the correlation is 0.7 or higher, the formalism holds. If not, we publicly acknowledge the error. That's what we call an honest Plan C, and it's documented in the paper.

Answer · They'd be right if we claimed to model the LLM itself. But we don't. Our claim is explicitly restricted to the harness layer — tool dispatching, validation, permissions. This layer is deterministic by construction, and its [Alignment Operator] alignment with EQΦS rises to 0.80 when measured in isolation. The LLM itself remains a black box — that's our acknowledged limit.

Answer · Because Google itself did the compression work. The compressed_quality_signals.ex module explicitly contains the note "can be used in preliminary scoring." Google considers these 38 dimensions sufficient. Modelling all 14,014 attributes would be redundant and computationally expensive. We follow Google's own optimization decision.

Two-level answer ·

Level 1 — already done · we tested H1 on 4 publicly documented cases (HouseFresh, Retro Dodo, Semrush's 2023 HCU statistics). Across these 4 cases, the multiplicative formula reproduces the observation with a 0.00% gap. The competing additive model can technically match the 0.09 observation on HouseFresh phase 2, but only by imposing an absolute-value additive recovery (+0.04) that is ad-hoc and inconsistent with Core Update mechanics. The multiplicative model is more parsimonious and consistent with the leak's source code (the "× boost" docstrings in Q*). Full detail and formulas in §11bis · Empirical validation of H1.

Level 2 — still to be done · a blind validation on 20 URLs (where the ranking is predicted before observing, with no access to the observations to fit δ_k). That's the P2 plan, underway, delivery within 6 months max. If it fails, we publish an honest observation paper on the prediction/observation divergence.

Why answer at two levels · level 1 shows we're not about to invest 10 analyst-days in a formula inconsistent with known observations. Level 2 remains the strong scientific proof (blind validation). Both are necessary, neither is sufficient alone.

Answer · Yes. Three quantified predictions · (1) a site with a lifted 0.30 HCU demotion will see its score multiplied by 1.43 with no new backlinks, within 90-180 days. (2) A recommendation platform that transparently exposes its multiplicative scoring to creators (like Spotify for Artists v2) will gain 15-25% creator engagement over 24 months. (3) A wealth management firm that adopts the 3-level protocol cuts its MIFID II compliance costs by 25-35% over 12 months. We'll verify publicly at 12-24 months.

Answer · Excellent question. This is precisely what our M19 resilience index measures. If Google releases a Helpful Content Update version 2, H1 and H2 start degrading (Λ drops by 25%). H3 stays stable. That's why we publish H3 first — it's our safety net. Any hypothesis calibrated on data is by construction sensitive to updates; any purely methodological hypothesis isn't.

Two-part answer · what they did, and what mathematical formalization adds on top.

What [Expert SEO 1] ([SEO Firm 1]) contributed · the descriptive inventory of the leak. He went through the 2,596 modules, identified the names of the most important ones (NavBoost, NSR, Q*, CRAPS, Twiddlers, siteAuthority), published blog posts explaining them in operational SEO language. Format · narrative, qualitative examples, sector SEO recommendations. What you read from him · "the leak suggests Google uses clicks in a sophisticated way."

What [Expert SEO 2] ([SEO Firm 2]) contributed · the media reveal and strategic analysis for the marketing community. He coordinated the public disclosure on May 27, 2024, contextualized why it matters, drew high-level conclusions ("Google has been lying about click usage for 10 years"). Format · long-form blog post, podcast, press appearances. What you read from him · "clicks matter, so here's what you need to change in your strategy."

What mathematical formalization adds on top · a formalization is only useful if it unlocks things an inventory or a reveal can't. Here are the three.

• Quantified, verifiable predictions · once the system is written as equations, you can calculate what it does under given conditions. Example · if Google applies a Helpful Content Update at δ = 0.30 to a site, the score is multiplied by 0.70 — not reduced by 0.30 points. And if it also receives an anchorMismatch of 0.15, its score is multiplied by 0.70 × 0.85 = 0.595, not by 0.55. The two penalties compound, they don't add up. Cf. §11bis · H1 validation on HouseFresh phases 1+2 where we predict 0.09 and observe 0.09, while the additive model predicts 0 (clipping).
• Transposable sector recommendations · once the formalism is set, it can be transposed to sectors where the mechanics are mathematically isomorphic (ranking, AdTech, recommendation, credit scoring, regulatory analysis). Cf. §11.3 · 5 sectors recommended via H1, §12.3 · 4 sectors via H2, §13.3 · 6 sectors via H3 (fintech MIFID II, medtech FDA, legal AI, defense, public, SEO). [Expert SEO 1] talks SEO, [Expert SEO 2] talks SEO/marketing — the formalization, meanwhile, lets us also talk to a wealth management firm or a medtech publisher.
• 12-36 month temporal trajectories · the M33 module (fractal anticipation) takes today's Λ values and computes how they evolve under 3 weighted scenarios (low/median/high). Result · H3 grows +21.8% over 36 months (#1), H2 stable, H1 sensitive to Google updates. Cf. §15 · M33 projections (a chart of the 3 trajectories). No narrative gives you that kind of curve — you need the formalism to generate it.

The intellectual trajectory · [Expert SEO 1] showed which objects exist in the system. [Expert SEO 2] explained why these objects matter socially. Our mathematical contribution lets us compute · how these objects interact (multiplicatively, not additively), in which other sectors the same mechanics hold, and where the system is headed over 12-36 months under various predictable shocks. The three levels are complementary, not competing — we explicitly cite [Expert SEO 1] and [Expert SEO 2] in the paper and acknowledge that we build on their work.

What this concretely gives you, as a client or reader · three quantified predictions verifiable within 12-24 months (cf. §17bis + Q5), six sector recommendations with estimated gains (cf. §11.3-§13.3), an audit protocol applicable starting tomorrow (cf. §13bis). [Expert SEO 1] tells you what exists. [Expert SEO 2] tells you what it means for the market. We tell you what to calculate, what to predict, and what to recommend going forward.

Two-level answer · what we make verifiable, and what we keep — exactly as a quantitative firm would do if it wanted to be credible without getting copied.

Level 1 — publicly verifiable (open script) · the primary facts of the Google leak and the classic academic techniques we applied. Concretely, what a third-party researcher can reproduce by hand or with a script under 100 lines ·

• Exact count of the Google leak's attributes and modules (14,014 attributes, 2,607 `.ex` modules, 38 dimensions of the compressed_quality_signals vector) — just open the repo and count.
• CRAPS saturation · $1 - 18 / \log_2(3 \cdot 10^{12}) = 0{,}5657$ — that's 1948 Shannon entropy, a high-school formula.
• Multiplicative H1 validation on HouseFresh · simple arithmetic, $Q = 1{,}00 \times (1 - 0{,}95) \times 1{,}80 = 0{,}09$ vs. the additive model that clips to 0. That's what the verification script does, in under 50 lines.
• F/S/C audit of the 5 SEO articles (cf. §13bis) · just a manual count of claims, no magic.

That script is open, with no external dependency, and we publicly commit to acknowledging any error if a researcher gets a different result. It verifies the Facts, not the Codex.

Level 2 — proprietary Codex φ (unpublished) · the $\Lambda$ values (1.8012 for H1, 1.9461 for H2, 1.7880 for H3), the [Alignment Operator] (0.87 for Google), the [Duality Operator] (Ω_dyn = 0.5004), the M19 (resilience 1.40) and M33 (projection +21.8% at 36 months) modules are produced by the proprietary deterministic EQΦS (EQPhiS) Asia engine. This engine encodes 9 modules (M5/M6/M7/M13/M16/M19/M33) and 2 operators ([Alignment Operator]/[Duality Operator]) with their sector weightings, calibration, and constants ($\Delta\phi C$, $\omega_\phi$, $\kappa$, $\alpha$). This is EQΦS's commercial IP — we don't publish it, exactly like a quantitative firm doesn't publish its pricing engine.

Why this is more transparent than Semrush/Ahrefs/Moz, not less · none of the 3 established tools give the primary facts (the datasets used to calibrate their DA/DR/Authority Score · never published) nor the formulas (the exact calculation producing a score of 67 vs. 68 · never explained). They're opaque on both levels. We give the entirety of level 1 (facts + classic academic techniques) and explicitly delimit level 2 (the proprietary orchestration). Cf. §13bis · F/S/C audit where Ahrefs scores 60%, Moz 40%, Semrush 20%, vs. our strict 74%.

The analogy · recipe vs. restaurant · we publish the ingredients (academic techniques · Shannon, Huffman, sparse attention), the photo of the finished dish (the quantified $\Lambda$ values), and the shopping list (the bibliography). We keep · the exact proportions, the procedure, the cooking times. A food critic can judge whether the dish is good. A competitor can't reproduce it at home without hiring the chef. Cf. §17bis · the Published vs. Reserved table, which details this discipline line by line.

The scientific commitment still holds · if you reproduce the level-1 counts and get a different number (say you count 13,998 attributes instead of 14,014), write to us, and we'll correct it publicly. If you want to compute a $\Lambda_{H_4}$ for a new hypothesis $H_4$ in your sector, that isn't reproducible on your end — that's an EQΦS consulting engagement, not an open-source script. The boundary is explicit and defensible under peer review.

Answer · The leak was published by Google itself, via yoshi-code-bot, on March 13, 2024, on their own public repo. The documents were archived by multiple third parties between March and May 2024. The case law on intellectual property for documents accidentally published by their owner and widely disseminated is clear — scientific use for formal analysis is protected. We consulted an IP lawyer before publication.

Answer · Three concrete things. One, if you do SEO, diagnose the dominant multiplicative factor for your clients instead of raising every bid. Two, if you do algorithmic recommendation, expose your multiplicative scoring transparently — it's a market differentiator. Three, if you work in fintech, medtech, legal, defense, or the public sector, the 3-level Fact/Signal/Computation audit protocol directly addresses regulatory requirements (MIFID II, FDA Explainable AI, EU AI Act). A free discovery call with EQΦS (EQPhiS) Asia to dig into your case.

• D-7 · full rehearsal out loud. Time it. Cut if it exceeds 28 minutes (leave a 2-min Q&A buffer).
• D-3 · rehearsal in front of the EQΦS team. Get feedback. Make one last adjustment.
• D-2 · back up the deck on 3 different supports (USB, cloud, email). Check the links to the public GitHub repo. Test the video rendering.
• D-1 · arrive on site. Test the stage equipment. Check that the KaTeX equations display correctly (sometimes broken when projected).
• D-0 · 2h before · light coffee, no alcohol, last-minute slide check. 30 min before · deep breathing, focus on the intro (the first 90 seconds).

• Never read the bullets. Always in complete spoken sentences.
• One idea per slide. If a slide has more than 3 visual elements, that's too many.
• A 2-second [PAUSE] before every section transition. Silence anchors the previous idea.
• Sweep your gaze across the audience in thirds (left, center, right). Never at the slides.
• If a technical term comes up, explain it within 10 seconds. Pros are mixed in with non-pros at 80% of conferences.
• A concrete anecdote in the middle (the site that moved from 0.5355 to 0.7650 by lifting the HCU). Numbers stick when they're embodied.
• Q&A · repeat the question before answering. It buys you 5 seconds to think.
• If a question is beyond you, say "Great question, I don't have the precise answer, let me dig into it and get back to you." Sincerity builds more credibility than bluffing.

H1 · Commercial algorithmic ranking (Google search) is a dynamic system · its penalties multiply instead of adding up, its click signals create closed feedback loops, and certain thresholds trigger abrupt phase transitions. This is the hypothesis we can test empirically on 20 real URLs within 2-3 weeks.

H2 · Commercial Google ranking and LLM re-ranking (RAG, agentic retrieval, MCP-search) share the same underlying mathematical architecture · dense vector embeddings, multiplicative scoring, re-ranking loops. If true, a single formalism covers both worlds — but cross-domain calibration is more costly.

H3 · The market for SEO tools and AI observability tools has no explicit audit standard. Existing tools blend three fundamentally different things into a single opaque score · what they observe, what they infer, what they compute. We propose the 3-level Fact/Signal/Computation protocol as a methodological standard transposable to any industry.

The SEO community did a remarkable job of descriptively inventorying the Google leak in 2024 ([Expert SEO 1] via [SEO Firm 1], [Expert SEO 2] via [SEO Firm 2], Cyrus Shepard via Zyppy). The AI Engineering community published unified re-ranking frameworks in 2024-2025 (RankRAG NeurIPS 2024, Rankify 2025, SciRerankBench 2025). The academic Information Retrieval community has studied ranking for 30 years (SIGIR, ECIR, CIKM, WSDM).

But none of these communities talk to each other. [Expert SEO 1] doesn't read NeurIPS. NeurIPS researchers don't read [SEO Firm 1]. SIGIR papers rarely use commercial datasets like the Google leak. The mathematical bridge we're proposing is the first one to explicitly connect them · the same formalism, the same equations, applied to the evidence available in each of these worlds.