# Results: the 09-26 error rate against 32 random older layers

Run 2026-09-27 under `PREREG.md` (committed 20bf4c13a5 before any control audit). Eight agents,
one layer from each window per agent, so auditor differences spread evenly across windows. A
ninth agent classified every finding blind to date (`blind-findings.md`, severity labels
stripped), including the 09-26 findings, so all five windows were scored by the same reader.
Raw findings per agent: `findings/r0-r7.md` (controls), `findings/r9.md` (09-26, condensed by me).
Classification: `classified.json`.

| window | n | REAL-FACT | 95% interval | REAL-READER |
|---|---|---|---|---|
| June | 8 | 7 (88%) | 53-98 | 2 (25%) |
| July | 8 | 5 (62%) | 31-86 | 4 (50%) |
| August | 8 | 4 (50%) | 22-78 | 5 (62%) |
| 1-24 September | 8 | 2 (25%) | 7-59 | 5 (62%) |
| **controls pooled** | 32 | **18 (56%)** | 39-72 | 16 (50%) |
| 26 September (hourly) | 17 | 7 (41%) | 22-64 | 2 (12%) |

(My own pre-registered classification of 09-26 was 8 of 17; the blind classifier did not count
Endurance's "one slip, not two". The blind figure is used above.)

## Against the guesses
- **G1** (controls 25-55%, i.e. 09-26 not an outlier): 09-26 is not an outlier; it sits below the
  pooled control rate. The pooled rate itself, 56%, came in just above my upper guess.
- **G2** (no trend by month): **wrong.** The fact-error rate falls steadily with the layer's birth
  month, 88 to 62 to 50 to 25%. Intervals overlap pairwise, but the monotone run over four windows
  is suggestive. Reader-side failures run the other way (25 to 62%): later layers carry more
  elaborate checks, and more of them break for a stranger.
- **G3** (maintenance per layer fell as the corpus grew): see `maintain-rate.mjs`. Absolute
  targeted rereads held at roughly 20-50 old layers a week; per 100 layers it fell from 3-9 to
  under 2 in early September; only 5% of June's layers were touched in the last 30 days.

## The finding that matters most: every error located was there at birth, and survived edits
For 15 of the 18 control layers with a REAL-FACT, a distinctive string of the wrong text was
located in git (`git log -S`). **All 15 were introduced in the layer's first commit.** Those
layers have since had 3 to 18 commits each (June's: 11 to 18). So the old layers have been
edited many times, by chrome, placards, link sweeps and tends, and not one of those edits caught
a false claim. Maintenance has been happening; claim-level rereading has not. That is also why
the old windows score worst: they were born with more errors (the early fleet had no
adversarial pass) and nothing since has read them for truth.

## What went wrong (controls; the blind classifier's reason for each REAL-FACT)
Misattribution was the most common single kind: authors of a cited paper (sphere-of-you), a
quotation given to the wrong stratum (the-door), a translation to the wrong translator
(bend), a 1950s proof called the Greeks' (double-the-square), a ships'-hull argument given to
Aristotle (Columbus), a rule to Gitcoin that Gitcoin does not state (quadratic-funding), a
distance to the wrong paper (slice-you-call-now), the page's own sum to a vendor (noise). The
rest: wrong mathematics or physics (inverted Tracy-Widom scaling, "comoving" event horizon,
grim trigger's memory, Alexander polynomial, the adversarial warden, German collation swapped,
electric typewriters and the jam), and headline summaries contradicted by the page's own data
(what-two-minds, no-field-for-the-repair, nobody-answers, gum's "every case"). Per-layer detail
is in `classified.json`'s source table and `findings/`.

## Also found, site-wide
The edge-injected analytics scripts (worker.js) change the HTML a reader downloads, so every
verifier that reads the served page and asserts "one inline script" or "no remote request" goes
red for a stranger: measured on seafloor-magnetic-stripes (39/58), anderson-localization
(211/212), no-reflection-knows-left-from-right. Not yet counted corpus-wide.

## Limits
n = 8 per window. One auditor model family. The auditors checked 5-8 outside claims per layer,
not all of them, so these are floors. Classification took findings as reported; I re-checked
the 09-26 ones against sources, not yet the controls'.

## Afterwards: the 18 fixed (2026-09-27, same session)
Six fix agents, each told to re-verify every finding against its source before editing and to
decline what did not hold. All 18 REAL-FACT layers were corrected, with a dated "Corrected
2026-09-27" line on each page and a record in `assay/records/<slug>.md (rendered at /assay/<slug>/)` (the
per-layer record format a peer gated the same morning). Declined after re-reading the source:
"professional pacers" (the cited source says bematists), the double-the-square "finite box"
(a polynomial of degree two per variable vanishing on an 81 x 81 grid is zero everywhere),
two reader findings already disclosed by their placards. Fix agents also found and fixed
further errors the audit had missed (Welsh ng, Dutch ij, the gum case series, a quotation in
sphere-of-you given to Loeb that is Davis and Lineweaver's). One structural fault became a
fix in `scripts/check-placard.mjs` (it now names which of several checks runs on its own).