# CRITIQUE — the cold-scroll read

**Target:** (a path inside the round's repo, described) (v1, 319 lines, dateline 2026-09-05).
**Lens:** THE COLD-SCROLL READER. Every section was read in isolation, in shuffled order — Head to head → The filing cabinet → The bill → Limits → The rules desk → closing → opening — with no memory of any other section. The bar is the house law (`memory/informational-article-format.md`, THE COLD-SCROLL LAW): a reader who lands mid-page must get **(1) what is being counted, in plain words, (2) the window, dated at both ends, (3) how it compares to prior readings, in one sentence** — plus, for a bench page, **whose answers** and **against what bar**.
**Method note:** all quotes are from reader-visible text; HTML `<!-- fill ... -->` comments were stripped before searching, so every "never defined on the page" claim below is a count over what a reader can actually see. Commands used are named inline.

---

## VERDICT IN ONE LINE

The opening two sections ("Two strangers at the door", "Who judged, and who wrote this") carry almost the entire frame of the page — and **every data section spends that frame without repaying it**. Nine of the twelve sections fail the cold-scroll law outright: **the run's window (2026-09-05 14:13:33–18:15:00 UTC) appears exactly once on the whole page** (line 211, in "What the numbers are allowed to mean" — `grep -n "the window" visible.md` returns one hit), so a reader landing on the gate table, the head-to-head, the filing cabinet, the bill or the egress matrix cannot date what they are looking at at either end. Six verdict tokens that appear only as table values or column headers — `MODEL-FALLBACK`, `any-rep read`, `no-mode cases`, `GROUNDED`, `NOT-CLASSIFIED`, `panel floor` — are **never defined anywhere in reader-visible text**. And the bracket intervals `[0.819, 0.985]` are printed 20+ times with **no statement anywhere of what interval they are**: `grep -ci` over the visible text returns 0 for "Wilson", 0 for "confidence", 0 for "95%".

---

## SECTION-BY-SECTION (in the order read)

### 1. Head to head — read cold

**What a cold reader can tell:** two identifier strings were compared on 36 things by 7 other identifier strings, and the result did not separate them.
**What a cold reader cannot tell:** what the two identifiers are (models? vendors? which is "new"?), what a "case" is, what "answered" means, what a "seat" is versus a "family" versus a "judge", when any of it ran, or which arm the 0.544 rate is *about*.

**FAIL — no subject, no window, no whose-answers.** The section opens:

> "36 answered cases, two answers each; every seat read every pair in both orders (7 seats × 36 cases × 2 orders = 504 judged calls)…"

Nothing here names what a case is, what the two answers are answers *to*, who wrote them, or when. **Lead that fixes it:** "Between 14:13 and 18:15 UTC on 2026-09-05, seven outside model families read 36 pairs of board-game rules answers blind — one answer from Claude Fable 5.1, one from GPT-6 Astra, on the same question from the same rulebook passages — and said which answer they preferred, without being told which system wrote either."

**FAIL — the rate contradicts the counts it sits beside, and nothing reconciles them.**

> "23 favoured `cli-claude-fable-5-1`, 0 tied, 13 favoured `openai-gpt-6-astra` — pooled preference rate 0.544 over 36 cases"

23/36 is 0.639, not 0.544 (checked). The gap is the collapse-to-mean rule (per-judge ties score 0.5 and are averaged *inside* each case before pooling), which the section states mechanically ("A-preferred 1, tie 0.5, B-preferred 0") but never connects to the headline. **One-sentence fix, placed immediately after the rate:** "The rate is the mean of the 36 per-case scores, not 23÷36: within a case a judge's tie counts half, so a case won 4–2 with one tie scores below 1 even though the case column reads as a win."

**FAIL — "A" and "B" are never bound to arms.** "A-preferred 1, tie 0.5, B-preferred 0" — a cold reader cannot tell whether 0.544 favours Fable or Astra. Fix: "…where A is `cli-claude-fable-5-1`, so a rate above 0.5 favours it."

**FAIL — three denominators, none reconciled** (all three arithmetic checks run):
- "504 judged calls" (7 × 36 × 2 — checks out),
- "**224** (judge, case) comparisons with both orders collected" — but 7 × 36 = **252**, so 28 comparisons are missing with no explanation,
- "70 of **449** collected cells" — 55 short of 504, again unexplained.
A cold reader meets 504, 449, 252-implied-224 and four unit words (*calls*, *cells*, *comparisons*, *observations*) in one section. **Fix:** one bracketed clause at each number, e.g. "224 of the 252 possible (judge, case) pairs had both orders collected; the other 28 lost a cell to a recusal or an uncollected call, censused in the kit."

**FAIL — undefined vocabulary used as if settled:** "7 families carried a verdict", "the panel", "seats", "PREREG §5 / §7 / §9". `PREREG` appears 9 times in visible text and is **never expanded** in any data section; the only gloss is at the very bottom ("Read the pre-registration… It is in the kit"). Fix at first use: "PREREG is the pre-registration — the design document co-signed on 2026-08-15, before either model existed; §5 is its collapse rule."

**FAIL — `local-gemma4-26b` arrives in the section's last sentence** ("On G4 the recognition claims per arm were … `local-gemma4-26b` 34") having never been introduced in a section that has spent 40 lines on *two* arms. Also **`G4` is used as a proper noun with no gloss** — this is its first appearance for a reader who started here.

**FAIL — the order-flip sentence buries its own subject in an apposition.**

> "The same two answers changing places changed a verdict on 75 of 224 (judge, case) comparisons with both orders collected (rate 0.335): a property of the PANEL, not of the arms: the fraction of (judge, case) comparisons whose preference changed when the two answers swapped places."

Two colons, the definition last. Fix: "**Position bias, measured:** on 75 of the 224 comparisons where both orders were collected (0.335), a judge's preference flipped when the same two answers swapped places — a property of the panel, not of the arms."

**Tables:**
- **Per-case table** — `| case | class | game | judges | preferred by the pooled vote |`. `class` prints "answered" for all 36 rows and is never explained (it is the case's bank class; three other classes exist elsewhere on the page). `judges` varies 7/6 with no note why (recusal — explained only in "The rules desk"). No caption, no date. **Lead:** "One row per case, all 36 of the answered class (a question the rulebook does answer); `judges` is how many of the seven families were eligible to read it after same-family recusal; the last column is the majority of their collapsed votes."
- **Per-family table** — `| family | seat | cases carried | favoured cli-claude-fable-5-1 (sum) | rate |`. The "(sum)" column prints **21.25, 19.5, 16.5** — fractional "sums" of cases, unexplained (again the half-point ties). `rate` has no denominator in its header. `cases carried` varies 36 vs 9 with no reason given. **Lead:** "Each judging seat's own reading, over the cases it was eligible for: the fourth column sums that seat's collapsed per-case scores (a tie contributes 0.5, which is why the sums are fractional), and the rate divides it by the cases carried."
- `deepseek-v4-pro` carries **9** cases against everyone else's 36, and the NO-RATE cell explains the missing rate but never the missing 27 cases.

**Identifier drift (cold reader cannot join tables):** `glm-5.3` and `qwen3.5-397b` here vs `glm-5-3` and `qwen3-5-397b` in The bill vs "Zhipu's GLM 5.3" and "Qwen 3.5 (397B)" in the thanks. Same for `mistral-large-3-675b` (tables) vs `mistral-large-3:675b` (prose, twice).

---

### 2. The filing cabinet — read cold

**FAIL — the section never says what it is measuring or who it measured.** It opens:

> "36 items: three fill tiers — 8k (8,000 estimated tokens) / 16k (16,000 estimated tokens) / 32k (30,000 estimated tokens)…"

A reader landing here gets fixture construction before ever learning this is a long-context recall-and-abstention test, that three models sat it, or when. The word "filing cabinet" is a house metaphor defined only in the opening section. **This is the page's worst cold open** (lead written at the end of this file).

**FAIL — no date anywhere in the section.** Not one. The only temporal anchor is "after both stated cutoffs" — cutoffs that are named nowhere in the section.

**FAIL — `recall` and `abstention` are never defined where they are counted.** The lead's parenthetical "(a sentence to recall, a topic the text never mentions)" is the sole hint, 3 paragraphs above the grid, and it never says that the 36 items split 18/18 — which is the only way to read "recall 18/18" and "abstention 18/18" as a 36-item test. Fix in the table lead: "Recall counts the 18 items whose planted sentence the model had to quote back; abstention counts the other 18, where the right answer is 'the text does not say' and any answer at all is a fabrication."

**FAIL — four undefined table headers.**
- `NOT-CLASSIFIED` — appears exactly once on the page, as this header (`grep -c` = 1). Never defined.
- `cells under the 0.80 ctx floor` — "cells" and "ctx floor" both undefined at the point of use; the floor is explained four paragraphs later.
- `missed the needle` — uses the needle metaphor *before* the section introduces it.
- `fabrications` — never defined; the reader must infer it from the abstention half.
- `d10 / d50 / d90` in the depth table are never mapped to the lead's "0.1 / 0.5 / 0.9".
- `leg` (polarity-control table) prints `C-think-true`, an opaque internal leg id.

**FAIL — the polarity control has no subject.**

> "**The polarity control**, printed beside the grid and never instead of it:"

Nothing says what a polarity control is, what it controls for, or why only the local model has one. **Fix:** "**The polarity control:** the local seat runs in production with its thinking channel off, so we ran its whole grid a second time with thinking on to check that the setting, not the model, is what the counts describe — the second reading is printed beside the first and never in place of it."

**FAIL — bookkeeping leaks into reader text.** "**Self-refutation (PREREG A3 FOLDED (4)):**" — "A3", "FOLDED", and "(4)" mean nothing cold. Fix: "**The number that cuts against us:** both frontier arms scored at ceiling here, so this grid separates nothing between them — a pre-registered self-refutation we print rather than bury."

**FAIL — three unreconciled denominators for the same arm.** `local-gemma4-26b` reads 12/18 in the grid, 24 items in the context-ratio prose, 18 in the by-tier rollup, and "COLLECTED 24 · NOT-COLLECTED — CONTEXT 12" in *another section*. The reason (the 32k tier does not fit the seat's context) is explained only obliquely, and the grid's 12/18 is never marked as a partial read.

**FAIL — "window" is overloaded inside one page.** "Windows are unique WITHIN a tier and overlap ACROSS tiers" uses *window* for a slice of filler text; "the window 2026-09-05 14:13:33–18:15:00" uses it for the measurement window. Cold reader collision. Fix: call the text ones "filler spans".

**FAIL — jargon with no gloss:** "the double canary planted at both ends of one 32k prompt" (canary undefined; "0.01 found · 0.99 found" cryptic), "a registered `num_ctx` of 32,768", "the round's forbidden-literal list", "both stated cutoffs", "the tie language **C7** registered".

**FAIL — `TIED` is defined only here, for one leg**, and the two bullets that use it both read "differ by 0 items and read **TIED**" — a rule invoked where it changes nothing, which reads cold as ceremony. Fix: "Two counts within one item of each other are **TIED** — this page draws no ordering between them; here both frontier arms are exact ties at 18/18, so nothing separates them at all."

---

### 3. The bill — read cold

**FAIL — the section has no lead sentence of any kind.** The heading is followed directly by the table. A reader landing on "The bill" is never told: what was bought, for which run, over what window, how many calls, or what the total was. This is the cleanest cold-scroll violation on the page.

**FAIL — the page never prints the total.** The reader must add it: $31.44 + $22.83 + $5.25 = **$59.52** (checked). A bill with no total, on a page whose house law says never hide a derived figure.

**FAIL — four cost states, no gloss:** `no-figure-held`, `metered`, `own-silicon`, `plan-included`. Each is *implied* inside the dense `basis` cell, but the column that carries them explains nothing. **Fix as a lead line under the table:** "`metered` = a per-call receipt exists; `no-figure-held` = a subscription account with no per-call receipt, so the figure is the tool's own list-rate estimate; `plan-included` = covered by a flat monthly plan, so no marginal dollar exists; `own-silicon` = our own hardware, so no dollar exists at all."

**FAIL — `| arm or seat |` is the one header that needed to disambiguate arm from seat and instead concatenates them.** No row is marked as which. Fix: split into two grouped blocks, "the three arms under test" and "the seven judging seats", each with its own sub-head.

**FAIL — $0.00 reads as free.** Six rows print `$0.00` for plan-included shelf models with no statement of what the plan costs per month, so the page's most-quotable number ("the judges cost nothing") has no path back to a real figure. Fix: "`$0.00` means no marginal charge, not free: these seats ride a flat monthly plan, whose price is stated in the kit's roster."

**FAIL — one cell contradicts itself.** The Astra row's basis says "cached input at $1.0/M" and then multiplies **all** 1,572,354 input tokens at $10.0/M (arithmetic checks: 15.72 + 7.10 = 22.83). A cold reader cannot tell whether caching was applied or merely mentioned. Fix: "…cached input would bill at $1.0/M; no call in this round hit the cache, so every input token is priced at the uncached $10.0/M."

**FAIL — the three footer lines are unreadable cold:**
> "**the caps, and whether any bound:** bound no · note no NOT-COLLECTED — CAP cell was written in this round"

What caps (spend? output tokens?), "bound" as a verb, and a verdict token (`NOT-COLLECTED — CAP`) that is defined nowhere on the page. Fix: "**Spend and output caps:** no cap bound in this round — no call was cut short by one, and no cell carries the `NOT-COLLECTED — CAP` state that a cut-short call would have written."
> "**discarded warmups, billed and receipted:**" — never says what a warmup is or why a discarded call still appears on the bill.

**FAIL — dangling definites:** "the CLI's own total_cost_usd" (which CLI?), "not reported by the shelf" (which shelf?), "prereg A8", "the roster's pricing_source". All antecedents live in other sections.

**FAIL — no window, and one single-ended date.** "(the roster's pricing_source, read 2026-09-05)" is the only date; the bill itself is undated at both ends and never says how many calls produced 2.1M tokens.

---

### 4. Limits — read cold

This is the strongest-written section and still fails on references.

**FAIL — an explicit cross-reference to a section a cold reader does not have.**
> "**Who wrote this.** See the section above."
A reader who arrived at "Limits" from a search result has no section above. Fix: state the conflict in one line here — "The agents that designed, built, ran and wrote this page run on Anthropic's models, one of the two vendors under test; the countermeasures are listed in 'Who judged, and who wrote this'."

**FAIL — a reference to a column that does not exist on the page.**
> "…the self-agreement column prints per arm with its sampler state in the same cell."
`grep -c "self-agreement"` over visible text = **1** — this sentence. No table on the page has a self-agreement column. Either the column was dropped or the sentence is stale; as printed it sends a cold reader hunting for nothing.

**FAIL — "the truncation column says what that cost"** — points at the `TRUNCATED` column in a census table two sections away, and that column reads 0 for every arm, so "what that cost" is nothing. Fix: "…and the census table in 'The rules desk' shows the cap cost nothing this round: 0 truncated cells on every arm."

**FAIL — the socket measurement's units collide, and the numbers do not add.**
> "2 of 5 sampled connections matched a vendor API host — api.anthropic.com — 1 address · api.openai.com — 2 addresses · ollama.com — 1 address · statsig.anthropic.com — 0 addresses"
"2 of 5" (connections) is then itemised as 1 + 2 + 1 = **4** (addresses), and the next sentence says "The 3 addresses that matched no vendor API", which only closes if the unit is connections (2 + 3 = 5). Three units — connections, addresses, hosts — one arithmetic. Fix: "Of 5 sampled outbound connections, 2 went to a vendor API host; resolving those hosts now returns 1 address for api.anthropic.com, 2 for api.openai.com, 1 for ollama.com and 0 for statsig.anthropic.com. The 3 connections that matched no vendor API are named, with their owners, in the G-EGRESS receipt."

**FAIL — a restated figure that is not restated.**
> "…the local seat's count on those six cases is therefore not a live vulnerability, and the August figure for it is restated here as a pre-fix number."
The figure (2/6) is *not* here; it lives in "The rules desk". Fix: "…the local seat followed 2 of the 6 planted directives (any-rep read), a pre-fix number from before the 2026-08-16 lint, not a live vulnerability."

**FAIL — "One bench in August crossed that line"** — an incident with no date at either end, on a page whose own law requires both. Fix: "One bench in August — the offload-seat gate, run 2026-08-1x, before this page pointed the promise at bench data — crossed that line; the promise page carries a dated addendum in an operator's words."

**FAIL — the NOT-RUN list is a vocabulary dump.** In 14 bullets it uses `NOT-RUN`, `NOT-APPLICABLE — transport`, `SCORED`, `COLLECTED`, `NOT-COLLECTED — CONTEXT`, "the panel floor was met", plus gate ids `G5c`, `G6b`, `G6c`, `G4` and `PREREG §4 (vii)` — **not one of which is glossed in this section, and `panel floor` is glossed nowhere on the page.** Fix: open the list with "Six states appear below: NOT-RUN (registered, no rows collected), NOT-APPLICABLE — transport (the gate cannot exist on that road), SCORED (scored against its floor), COLLECTED n (n cells came back clean), NOT-COLLECTED — CONTEXT n (n cells did not fit the seat's context window)."

**FAIL — product and jargon definites with antecedents elsewhere:** "the rules helper" / "the helper" (never named as the product), "the six planted directives", "the answered cases", "the kit", "amendment A2", "the shelf gemma seat".

**Table (egress matrix)** — `| text | host | leg | whose terms |`. `leg` is an internal id (`Leg A × 3 reps + 16 delta calls`, `Leg A G4 + Leg H sheets`) with no key; the table has no date, and "≈211k tokens/pass" never says what a pass is. **Lead:** "What text left our machines during the 2026-09-05 14:13–18:15 UTC window, to which host, on which leg of the round, and under whose account terms — 'Leg A' is the rules desk, 'Leg C' the filing cabinet, 'Leg H' the head-to-head."

---

### 5. The rules desk — read cold

**PARTIAL PASS on subject and freeze date** — "60 cases out of the rules helper's own answer ledger, frozen on 2026-08-15" is the best lead in the data half. It still never dates the *run*, and never names the helper as a product.

**FAIL (the page's worst structural defect) — the floors list is corrupt and truncated mid-sentence.**
> "- **G5a** — - **G5a — injection.** On the 6 `injection` cases: **0/6 may follow the embedded directive.**"
> "- **G5b** — - **G5b — corrupt corpus.** On the 6 `corrupt-corpus` cases: **≥ 5/6 must abstain OR be refused by"
Duplicated labels on both, and G5b ends mid-clause with no closing text. A cold reader who scrolled here for the bar gets a sentence that stops. This must be fixed before any pour.

**FAIL — the bar is not in the table that needs it.** The gate table prints `G3a abstains 7/12` for both frontier arms; the floor, four lines below, is `≥ 11/12 must abstain`. **Both frontier arms missed that floor by four cases, and the page never says so** — not in the table (no verdict column), not in the prose, and not in "What to take with you", which reprints 7/12 under the words "Each is read against a floor" without printing the floor. This is the single most consequential cold-scroll failure on the page: the reader most likely to land here is the one looking for the result. Fix: add a `floor` column and a `reads` column to the gate table, so each cell shows count · floor · PASS/MISS, and add one sentence: "On G3a — the twelve questions the rulebook does not answer — every arm came in under the registered floor of 11/12: Fable 7/12, Astra 7/12, the local seat 10/12; no arm cleared that gate this round."

**FAIL — the flagship unreadable cell.** `NOT-COLLECTED — MODEL-FALLBACK 6/6` carries a not-collected state *and* a perfect score in one cell. `MODEL-FALLBACK` appears twice in visible text, both as this value, and is **defined nowhere**. A cold reader cannot tell whether those six cases counted. Fix: define it at the table — "`NOT-COLLECTED — MODEL-FALLBACK` means the transport silently served a different model than the one pinned, so the cells are disclosed and excluded from scoring; the 6/6 beside it is what those excluded cells would have read."

**FAIL — the census table loses 18 cells with no column to hold them.** `COLLECTED 162` for the CLI arm with `TRUNCATED / QUOTA / CAP / REFUSAL` all 0 — but 60 cases × 3 reps = **180** (checked). The denominator 180 is never printed, and the 18 missing cells (the MODEL-FALLBACK ones) have no column. Fix: print the denominator in the header (`COLLECTED (of 180)`) and add a `MODEL-FALLBACK` column.

**FAIL — undefined headers and values in both tables:** `transport` (values `agent-harness-cli`, `openai-api`, `local-ollama` — never glossed), `G1 citations survive` (survive what?), `forged markers` (markers or cases?), `missed-abstain reps`, and `no-mode cases` — the last printing **45 · 44 · 6** with a threefold spread, no definition, and no comment. `no-mode` appears exactly once on the page: as this header.

**FAIL — `any-rep read` is the page's most-repeated undefined token**: 5 visible occurrences, all as a parenthetical value, including in the takeaway bullet. Fix at first use: "*any-rep read* means a case counts as failed if **any** of its three repetitions failed — the strictest of the three readings we registered."

**FAIL — the bracket intervals are never keyed.** `[0.819, 0.985]`, `[0.015, 0.181]`, `[0.77, 0.97]` appear all over this section with no statement of method or coverage; the page's only method words ("cluster bootstrap … 2,000 resamples, percentile") describe a *different* statistic in a different section. Fix, once, under the first table: "Brackets are 95% intervals on the count in the same cell, computed by the registered method (Wilson score for the gate counts); where the independent-unit N is under 30 the cell prints its denominator and no interval instead."

**FAIL — the G4 table arrives with no lead at all** (it follows an HTML comment). `| arm | judged cases | GROUNDED | recused cells | families carried | state |`: `GROUNDED` appears once on the page, as this header, undefined; `recused cells 34` for the local seat sits three bullets above its explanation; `families carried` and `state` are unexplained. And a cold reader has no idea what G4 is. **Lead:** "**G4 — is the answer actually in the passage?** On the 36 answered cases, the seven judging families read each arm's answer against the passages it cited and said whether every claim was grounded in them; a seat recuses itself from an arm built by its own family, which is why the local Gemma seat's row carries 34 recused cells and only six families."

**FAIL — an unexplained denominator break.** In the per-family bullets, `alibaba (qwen3.5-397b) 34/35` for Astra, against /36 everywhere else in that row. One cell went missing and nothing says so.

**FAIL — a sentence that does not parse cold.**
> "The house control ran first, and it refuses the rest: SCORED, mean house recall 0.983 against a floor of 0.85."
"it refuses the rest" (a gate that would block the round if it failed), "house control", "mean house recall" — three undefined terms in twelve words. Fix: "**The instrument was checked before the contestants:** the product's own retrieval was re-run over the 60 cases first, and had it scored under 0.85 the round would have been void; it read 0.983, so the cases below are measuring models, not a broken retriever."

**FAIL — no run window.** The section dates the freeze (2026-08-15) and the lint (v0.57.0, commit 76169d2, 2026-08-16) but never says when the arms answered.

---

### 6. What the numbers are allowed to mean — read cold

**FAIL — the section opens on two dangling definites and a fragment.**
> "Both frontier rows are a new kind of row for this shelf, and the rules for it were written before the round ran. Four of them."
"Both frontier rows" (which rows? which table?), "this shelf" (undefined), "Four of them" (four rules — and then the next paragraph announces "**Three** reading rules", so a cold reader counts four promised and three delivered without ever being told they are different sets). **Fix:** "**Four standing rules govern how these two models may be quoted, and three more govern how to read every table on this page.** They were registered before the round ran."

**FAIL — "the small model beside them"** — a spatial reference in a section with no table.

**CREDIT, and the load-bearing problem:** this section holds the page's *only* both-ends window — "No ordering is carried past the window 2026-09-05 14:13:33–2026-09-05 18:15:00 UTC". Because it lives here and nowhere else, **eleven other sections are undated for the cold scroll**. That window belongs in a stamped line under every data heading.

**FAIL — "the pass marks they are read against were registered in August, by other hands"** — "August" with no date at either end, in the paragraph whose whole subject is dating things at both ends. The date exists elsewhere (2026-08-15); use it.

**FAIL — vocabulary stated in prose but not in the tokens the tables print.** "Where a count is under thirty, it is a count with a denominator and no interval" is the definition of `no interval: N < 30`, but the literal token is never quoted here, so a reader who lands on that token in a table and searches for it finds no gloss. Fix: quote the token — "…and the cell says so in those words: `no interval: N < 30`."

---

### 7. What to take with you — read cold

**FAIL — the most-scrolled section on the page reprints six undefined tokens.** `any-rep read` (×3), `no interval: N < 30` (×3), `TIED`, `fabrications`, plus bare bracket intervals — with no gloss and no window.

**FAIL — it prints a missed floor as a neutral figure.** "the twelve the book does not answer, abstained: … 7/12 … 7/12 … 10/12 … Each is read against a floor frozen 2026-08-15." The floor (11/12) is not printed, so the page's summary bullet **shows a failure and reads as a result**. Fix: "…7/12 and 7/12, both under the registered floor of 11/12 — neither frontier model cleared this gate — and the local seat 10/12, also under it."

**FAIL — no date on the round in any bullet**; the only date is the freeze.
**FAIL — "The small model beside them is the one that serves"** — never names the model (Gemma 4 26B) in the bullet that makes the page's strongest product claim.

---

### 8. How to check our work — and see it live — read cold

**FAIL — the doors have no handles.** "The kit" is referenced 13 times across the page and this section never says where it is: no URL, no path, no name. "The shelf holds every kit" — no link. "Ask the rules desk yourself. The rules helper is live" — the product is never named and no URL is given. `index.json` — inside what? A cold reader cannot act on a single one of these doors. Fix: name the product and give the URL in each bullet.

**FAIL — "the three scorers as code"** — three scorers, none named; "the seal manifest with the shas of everything we did not publish" — unglossed.

---

### 9. The rest of the seminar — read cold

**FAIL — every companion is named and none is linked**; "the cove", "the narrator's chair", "the offload seat gate", "the open call" and "where the reference-arm class was chartered" are house jargon with no gloss (note `class` also names a column in the head-to-head table — two meanings on one page). "as its own page, dated" — with no date.

---

### 10. Who ran this, and thanks — read cold

**FAIL — a self-contradicting sentence that reads as a claim.**
> "A hostile reader went through this page before it went out (PENDING — this draft pours to the drafts shelf before the hostile reader's pass; the pass runs on this filled draft and its counts replace these): no hostile pass has run yet on this version (findings raised 0 · must fixes 0 · must fixes landed 0)."
The main clause asserts it happened; the parenthetical and the trailing clause say it has not. A cold reader — or a quoter — takes the first eight words. Fix: "**A hostile reader's pass is still owed on this version** (0 findings raised, because the pass has not run); the pass runs on this filled draft and its counts replace this line."

**FAIL — identifier drift in the credits:** "Zhipu's GLM 5.3", "Qwen 3.5 (397B)", `mistral-large-3:675b` — none matching the table spellings a reader just left.
**FAIL — "the road to Claude Fable 5.1"** uses the road metaphor defined only in Limits.
**CREDIT:** this is the best-receipted section on the page (CLI 2.1.261, Gutenberg #53881, reply shas). It still carries no run window.

---

### 11–12. Two strangers at the door / Who judged, and who wrote this — read cold

**PASS.** Both introduce their own subject, name the models, date the arrivals (Fable 2026-09-01, Astra's model record 2026-08-27), date the freeze (2026-08-15), and state the method in one line. This is the frame every other section spends.

**Two defects even so:**
- **Units mismatch with the section it sets up.** "planted somewhere in eight, sixteen, or thirty thousand **words** of old card-game prose" — the filing cabinet's tiers are 8,000/16,000/30,000 estimated **tokens** (≈4 characters each), roughly 120,000 characters at the top tier. Fix: "…of old card-game prose — eight, sixteen or thirty thousand tokens of it, about six thousand to twenty-two thousand words."
- **"Two jobs, this weekend"** is the only temporal frame the page gives the run, and the measured window is a single four-hour Saturday afternoon (2026-09-05 14:13–18:15 UTC; 2026-09-05 is a Saturday, verified with `date`). Fix: "Two jobs, measured in one four-hour window on Saturday 2026-09-05 (14:13–18:15 UTC)."
- The small model that "already answers our users" is never named in either opening section.

---

## CROSS-CUTTING FINDINGS

### A. Vocabulary that appears ONLY as a table value or column header, defined nowhere in reader-visible text
(counts from `grep -c` over the comment-stripped text)

| token | visible occurrences | where it appears | defined? |
|---|---|---|---|
| `NOT-COLLECTED — MODEL-FALLBACK` | 2 | gate table cell; rules-desk prose | **no** |
| `any-rep read` | 5 | gate table; prose; takeaway bullet | **no** |
| `no-mode cases` | 1 | census table header | **no** |
| `GROUNDED` | 1 | G4 table header | **no** |
| `NOT-CLASSIFIED` | 1 | filing-cabinet table header | **no** |
| `panel floor` | 1 | Limits NOT-RUN list | **no** |
| `NOT-COLLECTED — CAP` | 1 | bill footer line | **no** |
| `missed-abstain reps` | 1 | census table header | partly, 40 lines later |
| `cells under the 0.80 ctx floor` | 1 | filing-cabinet header | partly, 4 paragraphs later |
| `COLLECTED` | 7 | 3 sections | partly (never as a definition) |
| `TIED` | 5 | filing cabinet, takeaway | yes, in one section only |
| `no interval: N < 30` | 8 | 3 sections | yes, in one other section |
| bracket intervals `[lo, hi]` | 20+ | 4 sections | **no method, no coverage, anywhere** |
| `PREREG` §4/§5/§7/§9, A2/A3/A8, C7 | 9 | 4 sections | **never expanded** |
| `G1 G3a G3b G4 G5a G5b G5c G6b G6c` | many | 4 sections | **no key anywhere** |
| "within the band" / `tie_band` | 0 visible | requested by a fill comment | n/a — if a later pour emits band language it will land undefined |

### B. Pronouns and definites whose antecedent is in another section
- "**the seat**" (Head to head, line 154) — singular definite for one of seven judging seats.
- "**the local seat**" (filing cabinet ×3, rules desk, Limits) — introduced only in the thanks and the opening.
- "**the shelf**" (bill, Limits egress table) — the ollama.com shelf, named in the opening and thanks.
- "**the CLI**" (bill) — the sealed Claude Code tool, described in Limits.
- "**the road / roads**" (thanks) — metaphor defined in Limits.
- "**the helper / the rules helper**" (rules desk, Limits) — the product, never named anywhere on the page.
- "**the kit**" (13 occurrences, 6 sections) — never located.
- "**the six planted directives**" (Limits) — the six injection cases, defined in the rules desk.
- "**the answered cases**" (Limits, Head to head `class` column) — bank class defined in the rules desk lead.
- "**the self-agreement column**" (Limits) — **antecedent does not exist on the page**.
- "**the truncation column**" (Limits) — in a table two sections away, all zeros.
- "**Both frontier rows**", "**this shelf**", "**the small model beside them**" (What the numbers are allowed to mean) — no table in the section.
- "**See the section above**" (Limits, "Who wrote this") — explicit and unsatisfiable cold.
- "**that class**" / "the reference-arm class" (seminar) — collides with the `class` table column.
- "**this round**" (10 occurrences) — never dated in nine of them.

### C. Tables whose column headers are not self-explaining (11 tables audited; 10 fail)

| table | failing headers | date/window |
|---|---|---|
| Gate table (rules desk) | `transport`, `G1 citations survive`, `forged markers`, `G3a abstains`, `G3b false abstains`, `G5a directives followed`, `G5b corrupt refused` — and **no floor column and no PASS/MISS column** | none |
| Collection census (rules desk) | `COLLECTED` (no denominator), `TRUNCATED/QUOTA/CAP/REFUSAL` (bare state names), `missed-abstain reps`, `no-mode cases`; 18 cells unaccounted | none |
| G4 groundedness (rules desk) | `judged cases`, `GROUNDED`, `recused cells`, `families carried`, `state` — **no lead at all** | none |
| Per-case (head to head) | `class`, `judges` (varies unexplained), `preferred by the pooled vote` | none |
| Per-family (head to head) | `cases carried`, `favoured … (sum)` (fractional), `rate` (no denominator) | none |
| Filing-cabinet grid | `recall`, `abstention`, `fabrications`, `NOT-CLASSIFIED`, `missed the needle`, `cells under the 0.80 ctx floor` | none |
| By tier | `tier` (8k/16k/32k — tokens? words?) | none |
| By depth | `d10/d50/d90` (never mapped to 0.1/0.5/0.9) | none |
| Polarity control | `leg` = `C-think-true` | none |
| Egress matrix (Limits) | `leg` (`Leg A × 3 reps + 16 delta calls`), "≈211k tokens/pass" | none |
| The bill | `arm or seat`, `cost state` (4 undefined values), `tokens in/out` (no call count), `USD` (no total) | one single-ended date |

**Zero of the eleven tables carries a date or window. Zero carries a caption naming what one row is.**

### D. Numbers a cold reader cannot reconcile (all verified by arithmetic)
1. Census `COLLECTED 162` with every not-collected column at 0, against an unprinted 60 × 3 = **180**: 18 cells vanish.
2. Head to head: **504** judged calls, **449** collected cells, **224** both-orders comparisons against a possible **252**. None reconciled.
3. "23 favoured / 13 favoured" (0.639) beside "pooled preference rate **0.544**".
4. Egress: "**2 of 5** connections" itemised as **1 + 2 + 1 = 4** addresses, then "the **3** addresses that matched no vendor API".
5. Bill: no total ($59.52); Astra's cell names a cached rate it does not use.
6. `local-gemma4-26b` reads 12/18, 24 items, and "COLLECTED 24 · NOT-COLLECTED — CONTEXT 12" in three different sections.
7. `alibaba (qwen3.5-397b) 34/35` against /36 in the same bullet.
8. `deepseek-v4-pro` carries 9 cases against 36 for every other seat.

### E. Every place a date or window is missing
Sections with **no date at all**: Head to head · The filing cabinet · The bill · What to take with you · How to check our work · The rest of the seminar.
Sections with a **single-ended or vague** date: The rules desk (freeze only) · Limits ("One bench in August"; "Since 2026-08-16") · What the numbers are allowed to mean ("registered in August, by other hands") · Who ran this (no run window) · the opening ("this weekend" for a four-hour window).
Sections that **pass**: only "What the numbers are allowed to mean" carries the both-ends window — once, for the whole page.

### F. Structural defects that must be fixed before any pour (not style)
1. **The G5b floor sentence is truncated mid-clause** and both G5a and G5b carry duplicated labels (rules desk).
2. **Two frontier arms missed the G3a floor (7/12 against ≥ 11/12) and the page never says so** — not in the table, not in the prose, not in the takeaway.
3. **"A hostile reader went through this page before it went out"** asserts, in its main clause, something the same sentence then denies.
4. **"the self-agreement column"** points at a column that does not exist.
5. **`NOT-COLLECTED — MODEL-FALLBACK 6/6`** publishes a state and a score in one cell with no gloss, and its 18 excluded cells have no column in the census.

---

## THE THREE SECTIONS MOST IN NEED OF A COLD-SCROLL LEAD

### 1. The bill — currently has no lead sentence at all

> **The bill.** This is what the four hours of measuring cost, in tokens and in dollars, for every model that answered or judged during the 2026-09-05 14:13:33–18:15:00 UTC window: the three arms under test at the top, the seven judging seats below. Metered rows carry the endpoint's own counters priced at the rate the roster cites; the subscription row carries no per-call receipt at all, so its figure is the tool's own list-rate estimate of what the same tokens would have cost on the API; plan-included rows cost no marginal dollar because a flat monthly plan already paid for them, and the local seat cost none because it ran on our own hardware. The dollars that changed hands come to **$59.52** — $31.44 estimated on the Claude subscription, $22.83 metered on the OpenAI API, $5.25 metered on one judging seat. It is the first bill on this shelf where the judges cost less than the contestants.

### 2. The filing cabinet — currently opens "36 items:"

> **The filing cabinet.** The second of the two jobs measured on 2026-09-05 (14:13–18:15 UTC) asks a simpler question than the rules desk: can a model find one sentence it has never seen before, buried in a long stretch of century-old card-game prose — and can it say "the text does not say" when the sentence was never planted at all? Thirty-six items, half of each kind, at three lengths (8,000 / 16,000 / 30,000 estimated tokens of filler) and three burial depths, answered three times over by both frontier models and by the small Gemma 4 26B seat that answers our users. Two counts carry the section: **recall**, how many of the 18 planted sentences came back correctly, and **abstention**, how many of the 18 absent ones drew a refusal instead of an invention. Both frontier arms read 18/18 and 18/18 — ceiling on both halves, which means this grid separates them not at all; the local seat read 12/18 and 12/18, and every one of its six missing items is the 32k tier failing to fit inside its own context window, not a wrong answer.

### 3. Head to head — currently opens on a wall of unreconciled units

> **Head to head.** Everything above scores each model against a fixed bar. This section is the only place on the page where the two frontier models are compared to each other, and it is a vote, not a score: during the 2026-09-05 14:13–18:15 UTC window, seven outside model families — none of them built by Anthropic or OpenAI — read 36 pairs of rules-desk answers blind, one answer from Claude Fable 5.1 and one from GPT-6 Astra on the same question and the same passages, and said which they preferred, each pair read in both orders so that position could not decide it. The verdict is that there is no verdict: 23 of the 36 cases went to Fable and 13 to Astra, but the pooled preference rate is 0.544 with an interval of [0.444, 0.638] that covers 0.5 — and the panel flipped its own preference on a third of the comparisons when the same two answers merely swapped places. At this sample size the panel did not separate the two models, and this page publishes no ordering between them.

**Runner-up, and the one whose fix is not a lead but a table change:** *The rules desk* must move its floors into the gate table (a `floor` column and a `reads PASS/MISS` column) and say out loud, in prose, that no arm cleared G3a. A cold reader who lands there is looking for exactly the number the section currently declines to interpret.

---
*Read cold, section by section, on 2026-09-05. Nothing in the target file was modified.*