Not a summary.
A reading.

A real marketing page on the left — and on the right, what the best of five frontier AI models produced when handed the page's measured silence, under a prompt that forbids inserting, inventing, and padding. The only path left open was to read.

4 independent mathematics · every claim typed, cited & machine-audited against the source · generated 2026-07-10T19:00:23 · one demo document, every stage inspectable

yellow source-salient concepts every AI summary deleted (marked in the source) green void stems surfaced by ≥2 independent math classes, adopted by the winning reading (marked in the reading)

Caveat, auto-printed: this demo source splits into 1 paragraph(s), so citation granularity is effectively whole-document; multi-paragraph sources exercise the trace at full resolution.

The writer must ledger its own reading: F fact, traceable to cited paragraphs · I inference from structure · S speculation, drawn from the measured consequence field, hedged. Then the machine checks: facts traced lexically into their citations; absence claims searched across the whole document; every contest quoted. Confidence is earned here, never self-reported.

Across the whole panel this run: every positive fact traced — 30/30 — and the entire contested mass (34 claims) is absence claims whose negation vocabulary echoes the source's own words; 5 absence claims survived whole-document search. A contest here marks a judgment boundary, not a falsehood — the evidence is quoted in every contested row below.

Before any writer ran, a separate instrument asked ten models what this document raises but refuses to answer, clustered their questions deterministically, and reported consensus:

The winning reading opened on the same silence the census measured (lexical overlap 1/3 top clusters; the words differ, the gap is the same). Convergence between instruments that share no code is what “measured” means on this site.

Full judge × writer matrix (self-scores starred and excluded from standings; hover a dotted cell for the judge's stated reason):

Self-preference, measured on the anonymized diagonal: mean −0.25 across 5 judges (range +0.75 chatgpt … −1.25 grok). Labels are anonymized, so a nonzero diagonal is affinity for one's own style, not knowing self-favoritism — and this run the net lean was downward: the panel marked its own style down.

A divergence worth publishing: the panel's winner is not the audit's densest writer. gemini filed the largest ledger (26 claims, 15/25 traced, 0 bad refs) and placed fourth; claude won on insight with 25 claims, 11/19 traced and 8 contested — every contest an absence claim. Insight and provenance density are different measurements; this page publishes both and lets them disagree.

Local-panel coverage is partial this run: no local judge's parsed scores include C (deepseek), E (grok) — the column renders with per-label n and nothing is hung on it.

Send a URL or a text. One command runs the whole organism — harvest, the independent mathematics, both raycasts per void, the evidence-disciplined write by five frontier models, the ex-self panel, the claim audit — and you get back the unsaid report (deleted · unreached · unanswered · unsayable, four measured strata of your document's silence), the winning reading with its audited ledger, and the receipts: every artifact sha-stamped, identical inputs reproducing identical measurements, the two AI-written stages declared as such.

Full disclosure, standing policy: everything above is one demo document — thin marketing copy, and the pipeline itself flagged this document (FOREGROUND top 0.049 < 0.10 → THIN — thinness is a finding). The button grows the corpus. A position-matched placebo test is queued and will be published either way.

Measured: the census, the mathematics, the convergence law, the panel matrix, the claim audit, the self-preference number, the replications. Argued: that a page whose silence has been read wins visibility and revenue in the AI layer. Plausible, untested — nobody has that outcome data, because nobody else is reading the layer. Measure first.

All vectors are unit-normalized embeddings from bge-large-en-v1.5 (1024-dim), CPU-pinned for bitwise determinism. Per document: anchor h = E(title + prompt); per model group g ∈ {F, L, ALL}: response vectors e₁…eₙ, centroid c = normalize(mean eᵢ). Rulers: 50K curated-frequency clean vocabulary (readout + spiral leg), 253K Wikipedia-title vocabulary (flat leg), 184K global vocabulary (ring, optional; not run in this demo). There is no RNG anywhere in the geometric path; every artifact prints a RESULT sha — canonical JSON minus timestamp, sha256 → 12 hex — and identical corpus + anchor + parameters must reproduce it. The anchor is a load-bearing parameter of every method below except said/gap/lexcross.

I. Centroid arithmetic

said: v = c   ·   gap: v = ±normalize(c_L − c_F)   ·   centroid_surface: v = h

II. Gradient descent (both losses MINIMIZED; AdamW lr=0.05, wd=1e-4 — the weight decay leaks an undeclared pull-to-origin; 150 steps; sphere projection each step; init = c ⇒ determinism is structural)

logos_v10: L(x) = meanᵢ(1 − cos(x, eᵢ)) + 0.75·cos(x, c) − 0.30·cos(x, h)
logos_v9:  L(x) = LogosLossV9(x, E) + 0.15·cos(x, c) − 0.30·cos(x, h)

Sign semantics, stated because a prior analysis got them backwards (defect #49): minimizing +λ·cos(x,c) REPELS the consensus; minimizing −λ·cos(x,h) ATTRACTS the anchor. V10 in one line: the direction agreeing with every response individually while disagreeing with their mean — shared-but-not-central — tethered to the document.

III. Spectral (SVD)

E = UΣVᵀ (full_matrices=False); v = V[−1], the right singular vector of least σ;
flip sign if v·h < 0; n ≥ 3

The axis along which the group's answers cannot vary — what their span fails to express. Known gap: gate on σ_min (currently discarded); when σ_min is not small, this direction is noise.

IV. Counting

lexcross: stems said by the other model population, absent from this one, counted per model, alphabetic tiebreak. (History: a set-difference version made every count 1 and let the hash seed order the ranking — defect #39, the only RNG ever found in the organism.)

V. Ring geometry (optional)

A donut on the 184K ruler: inner ring excludes centroid-proximal words, outer ring gates on anchor relevance; hard fallback waives both gates; returns (results, centroid) — unpack at the call site.

Readout (all vector methods)

scan top-2000 ruler words by V·v; keep len≥4, not HARD_DROP;
dedupe by Porter stem; skip stems the group said; take K=8

flat leg — consequence raycast (253K ruler)

terminals T_λ = h + λ·d̂, λ ∈ {2, 3, 4}; kNN per terminal
density = mean pairwise cos(neighbors) · novelty = 1 − cos(T̄, v) · tether = cos(h, T̄)

Canonical example: a story avoiding “world war” whose consequence field contains “Hormuz”. Junk rule on this ruler (declared): leading non-letter | ellipsis | 3+ digit run | >4 tokens.

spiral leg — sentence-converged (50K ruler; C-spiral lineage, constants shared verbatim)

T = normalize(h + 2.0·d̂); pool = top-400 by cos(V, T)
keep w′: stem ≠ stem(w); stem unsaid by all models;
        convergence = #{source sentences s : cos(w′, s) > 0.45} ≥ 2
sort (−convergence, −cos); top 5

Both legs fire down the same ray: their agreement measures robustness across ruler + discipline, never independence — and it is priced (§E), never assumed.

CLASS-CONSENSUS: a void counts when independent mathematics land on the same stem. Distinct classes per stem is the metric; a word surfaced eight times by one class counts once (the said≡logos J=1.00 core proved within-class agreement is often structural: same centroid init, same responses). This run, computed over the full centipede JSON (every void per method, not the per-method-5 display feed handed to the writers): acura 2 classes (Centroid, Gradient) · automaker 2 classes (Centroid, Gradient) · delorean 2 classes (Centroid, Gradient) · drivetrain 2 classes (Centroid, Gradient) · infiniti 2 classes (Centroid, Gradient) · infotainment 2 classes (Centroid, Gradient) · precursor 2 classes (Centroid, Gradient) · prototypes 2 classes (Centroid, Gradient).

Disclosure, printed because the data demands it: every consensus pair this run is Centroid + Gradient, and the gradient stages initialize at the centroid (§B) — the two classes share a starting point, so their agreement is robustness evidence, not full independence. The convergence law is stated for the general case; this run it fired only on its weakest pair, and saying so is the point of this page.

shuffle null: cos(Āᵢ, B̄ⱼ) over ALL mismatched pairs with
different void stems — full enumeration, no sampling, no RNG.
Topic floor measured ≈ 0.615–0.67 across corpora to date.

Margins, session record (prior corpora): thin marketing copy +0.021 (correctly thin) · research story +0.058 · compliant-frame variant +0.058 (cone +0.123) · allegation dossier +0.080. This corpus, measured tonight: body median 0.636 vs shuffle null 0.6149 → margin +0.021.

Also printed per run: LEG-JACC (stem-Jaccard between leg void-sets, all pairs), ARM-JACC (~0.00 in ray mode — the rulers barely share surface forms; a disjointness document, not an agreement metric), per-leg consequence centroids vs body, and the RESULT sha.

void_freq(c) = stem-level TF salience of c in the SOURCE (max-normalized)
fidelity(c, summary) = max(cosine, lexical)
  cosine = max cos(c, any summary sentence), clamped [0,1]
  lexical = fraction of c's content stems literally in the summary's stem set
inv_fidelity(c) = 1 − max over summaries · VF-IDF(c) = void_freq × inv_fidelity

max() is the OR: the lexical channel is the false-void guard — a concept counts dropped only if every summary dropped it on both channels, so anything any summary carries lexically zeroes exactly. FOREGROUND takes only rows > 0.01 (zeros are preserved, not dropped); a top score < 0.10 prints THIN — thinness is a finding. This run's table:

COLLAPSED = the re-harvest now carries the concept: VF-IDF hits exact zero via the lexical channel. deepened = the rewrite made the concept more void, not less — a miss, published. The after-measurement treats the synthesis as the document under test: void_freq re-normalizes against the synthesis article, so before/after compare the concept's standing in two documents — source, then synthesis — under one metric.

Rerun record (the synthesis stage is a declared non-frozen stage): run-1 collapsed brand's; run-2 kept it and collapsed soon instead — a zero is a zero under any normalization. Both runs re-normalize the after void_freq against their own synthesis article — the synthesis becomes the document under test — so after-columns of different runs are not directly comparable (brand's 0.714 vs 0.111; show 0.143 vs 0.333; soon 0.143 vs 0.222). Run-1's stored after-rows additionally disagree with run-1's own console on two independent quantities (VF-IDF and survival), so run-1 numerics are withheld pending a serializer audit; run-2's stored values match its console exactly and are what renders above.

Carriage, measured. Tracked words in the 10 downstream AI summaries of the synthesis. A void concept that survives here has entered the model-mediated version of this topic — where, before, it did not exist.

the Summary Plus discipline (verbatim, sha 28f594d097b0)

You will write a SUMMARY PLUS of the source document: a summary that also reads the source's negative space.
Measured channels follow. Treat every channel item as a DIRECTION FOR ATTENTION, not a word to insert. Follow each direction into the source; where the source licenses it, surface what you find; if a direction finds nothing in the source, drop it silently. Invent nothing. Use zero analogies. Lead with what the piece leaves unresolved.
The source's paragraphs are numbered p1..pn. After your reading, output a line 'LEDGER:' and then one line per claim, strict format TYPE|claim|refs where TYPE is F (fact, traceable to the cited paragraphs), I (inference from structure), or S (speculation -- may draw on the consequence-field terms in Channel C, hedged). refs = comma-separated paragraph ids (e.g. p2,p4) for F and I; S may cite or leave refs empty. Every paragraph of your reading must be covered by at least one ledger line. Absence claims (things the source does NOT say) are F lines citing the paragraphs searched or 'all'.

the judge rubric (verbatim, sha 98f9385d5b3d; judges see prose only — the ledger is stripped before judging)

Score each summary 1-5 for INSIGHT:
5 = surfaces non-obvious, source-grounded structure a plain summary misses; 4 = several real observations beyond restatement; 3 = competent restatement plus one real observation; 2 = pure restatement; 1 = restatement with errors or invention.
Line 1 EXACTLY: A:n B:n C:n D:n E:n
Then one line per summary, A through E: its label and the single strongest reason for its score, grounded in its text.

Rubric amended 2026-07-10 (bakeoff v1.2, sha 98f9385d5b3d; previous one-line-only rubric sha cb7c7f362985): scores line plus one grounded reason per summary. The previous rubric commanded a ~19-character reply that the harvest floor (50 chars) destroyed — defect #52 below.

Writers anonymized A–E in deterministic (alphabetical) order; standings are EX-SELF means. Why the discipline forces improvement: the geometry supplies attention targets the writer could not self-generate, and the contract closes every cheap exit — insertion (directions, not words), invention (trace or omit), padding (drop silently; zero analogies). The one rewarded behavior left is reading.

parse: lines after 'LEDGER:' matching TYPE|claim|refs
F, positive:  ≥⌈stems/3⌉ of the claim's content stems present in the
              cited paragraphs (lexical trace) → TRACED
F, absence (claim contains no/not/none/never/absent/omits/lacks/…):
              search the WHOLE source for the claim's concept stems;
              none found → SUPPORTED-ABSENT; any found → CONTESTED,
              with the offending sentence quoted
I:            refs must be valid paragraph ids
S:            exempt; counted; must be hedged (style call, human)

Contests are surfaced, not auto-failed: “no performance figures” vs “strong off-the-line response” is a judgment boundary the audit's job is to make visible with evidence attached. Demo run: 5/5 frontier models emitted parseable ledgers cold (78 claims panel-wide). Across the whole panel this run: every positive fact traced — 30/30 — and the entire contested mass (34 claims) is absence claims whose negation vocabulary echoes the source's own words; 5 absence claims survived whole-document search. A contest here marks a judgment boundary, not a falsehood — the evidence is quoted in every contested row below.

closeness(w) = max(V·anchor, V·response-centroid)
plants = ascending-closeness words, len≥5, alphabetic, stem ∉ source ∪
said ∪ candidates; each camouflaged with its own top-3 vocab neighbors;
woven into sha-keyed slots, unlabeled

Measured, not curated: the vocabulary's farthest points from the document's field. This run's plants, measured min-closeness on this corpus: dreadfully (0.249), hysterically (0.255) — refused by the writer stage. Record across all runs: impossible decoys refused every time; one early ornamental decoy adopted twice, caught and evidence-quoted both times; rival brand names refused every run.

ten models independently list raised-but-unanswered questions;
lines ending '?' extracted (numbering stripped); embedded;
greedy deterministic clustering: fixed order (model, line index),
assign to best cluster with centroid-cos ≥ 0.80 else new cluster,
centroid renormalized per insert; report clusters with ≥ 3 models

One model's curiosity is an opinion; consensus is a measured gap. This run: How much will the 2026 Honda Prelude cost? 8/10 · What specific gasoline engine is paired with the hyb 8/10 · What is the curb weight of the 2026 Prelude? 6/10 — the same silence the winning reading opened on.

The bake-off is the selection instrument. Policy: run the ex-self panel per document class until the ordering stabilizes, then PIN one writer on three axes — panel score, audit cleanliness, price per run — record the pin in every artifact's provenance, and re-open the seat on major model releases. Current table:

Current pin: claude — panel 4.75 (n=4), ledger 25 claims, 11/19 traced. Provenance-density leader: gemini (15/25 traced, 0 bad refs). The seat re-opens on major model releases; the pin is recorded in every artifact's provenance.

#24 substring anchor-match (“lying” ∈ “identifying”) → stem-set comparison · #37 provenance hash globbed sibling corpora → loader whitelist, fix proven by restoring the original fingerprint · #39 hash-seed nondeterminism laundered through a set → per-model counts, caught by the RESULT sha on its maiden flight · #44 cosine-only fidelity made retained words look dropped → verbatim two-channel metric with exact zeros · #47 FOREGROUND sort/floor inversion → desc sort, >0.01 floor, THIN banner · #49 gravity sign error carried for hours → sign-semantics stated on this page, twice · #52 the judge rubric ordered a 19-character reply; the collection gate destroyed anything under 50 chars — for two runs the only surviving judges were the ones that disobeyed, and their stale files were then re-read as fresh by the next run → per-stage --min-chars, an mtime staleness gate on judge reads, a declared tie policy, and versioned output filenames.

1. Embedder: any bge-large-en-v1.5 checkpoint (or nearest
   successor); pin to CPU; unit-normalize all outputs.
2. Rulers: 50K = top-frequency English content words, stopworded,
   deduped by Porter stem; 253K = Wikipedia article titles; 184K =
   union corpus vocabulary. Exact historic rulers are fingerprinted
   by the shas below; the CONSTRUCTION is what matters.
3. Harvest: N models (5 frontier + 5 open-weight), one identical
   prompt, error-banner gate on refusal boilerplate, declared
   per-stage minimum-length floors.
4. Implement §B methods per group F/L/ALL; readout per §B; legs per
   §C; convergence law per §D; pricing per §E; VF-IDF per §F.
5. Writer stage: §G prompts verbatim; ledger contract verbatim;
   anonymize; ex-self panel; audit per §H; decoys per §I; census
   per §J.
6. Determinism check: run twice; RESULT shas must match. If they
   don't, you have an unfrozen stage you haven't declared — find it.
7. Honesty invariants: print raw metrics when gates saturate; price
   every agreement against its null; report what didn't work; the
   instrument that never delivers bad news cannot be trusted with
   good news.

• census: How much will the 2026 Honda Prelude cos 8/10 · What specific gasoline engine is paired 8/10 · What is the curb weight of the 2026 Prel 6/10 · Will there be any limited editions or sp 5/10 · When exactly will the official accessory 4/10.
• panel: spread 2.5 across writers (n=4 per mean); self-preference mean −0.25 (n=5), per-judge chatgpt +0.75, claude +0.25, deepseek +0.00, gemini −1.00, grok −1.25.
• plants: This run's plants, measured min-closeness on this corpus: dreadfully (0.249), hysterically (0.255) — refused by the writer stage.
• flat-leg honesty: clean gated terms for 3 of 136 void probes (raw raycast candidates for 53; the junk rule and quality gates account for the difference) — near-silence on thin copy is instrument behavior worth showing, not hiding.
• channel-A restoration by the winner: brand's not restored, show not restored, soon not restored — the discipline says directions, not words, and the winner chose structure over token restoration; the collapse table above is the corpus-level view of the same channel.
• pricing: This corpus, measured tonight: body median 0.636 vs shuffle null 0.6149 &rarr; margin +0.021.

Raw artifacts: bakeoff · centipede · census · synthesis