# Evidence for job #3397 (finding #566) on `research/a3-08-adjacent-pairs.js` Served a1a074f34be79cbb6c20df1ab797e0c76bded2d4b290982f8afe16a852de185d, 65935 B, 1100 lines; revision eabc5c89a8ea2c7fcee018d41d126c6e47fbcfae9cde8c696923088db1d3477b, 66521 B. The file carries a bound OUTPUT block, so the fix was installed through the embedding tool, not by hand. ## What was changed 1. Header clause (c), l.66: `the L-1 gaps inside a run have mean at least 3p - p/(L-1)` becomes the split bound `mean >= 3p when L-1 is even, and >= 3p - (3p - s)/(L-1) when L-1 is odd, s = 2p-2 for p = 1 (mod 6) and s = 2p+2 for p = 5 (mod 6)`, with the fold-31 counterexample named (the T29 four-runs 60+126+60 have mean 82 < 82.67). This is the code region, so it moves `code-sha256`. 2. The same clause inside the READINGS section, l.935-944: the closing sentence `Their MEAN is therefore at least 3p - p/(L-1)` becomes the split bound and the tightness note for fold 31 (L = 4, p = 29). ## Re-embedding (the file's embedded hashes depend on change 1) `node research/qc/embed.js research/a3-08-adjacent-pairs.js` was run with the served tree laid out at the paths it expects (`embed.js` sha256 ... see the tool line below): - it re-ran the producer: 215 lines of output, 107.6 s; - `code-sha256` 15c7b93cf9616d17... -> 2670cf470bcbbbea... (the code moved, as it must); - `out-sha256` 5088283144d2d7772729487a11666894167b6bc5bb0e614399c9ecc87f7eba9a **unchanged** -- the stdout still reproduces the pasted block byte for byte; - header lines rewritten by the tool: `node: v22.23.3`, `embedded: 2026-09-25`, `elapsed: 107.6 s` (were v22.21.0, 2026-08-19, 89.6 s); - the tool's advisory is expected and is not an error: `50 figure(s) in the existing READINGS are not in this output` -- the READINGS section is curated (the detached `--t31` run, 6.2e9 slots, the 200.56e9 primorial), and embed.js preserves it while rewriting the bound output above it. `node research/qc/embed.js --check research/a3-08-adjacent-pairs.js` then reported, on the written file: `code-sha256 matches` / `body matches out-sha256 -- the pasted block is bit-honest` / `out-sha256 matches`, with the same 50-figure READINGS advisory. ## Unguarded reading of the defect Clause (c) as served claims an even-case mean bound for all L-1. The file's own data contradicts it where L-1 is odd: at fold 31 the T29 four-runs are 60+126+60 (recorded in this file, l.942-943), mean 82 against 3p = 87 and 3p - p/(L-1) = 82.67 for p = 29, L = 4. The sum bound this file already states two lines above (`3p(L-1)` for even L-1, `3p(L-2) + s` for odd) gives exactly the split form now written, and the odd-case value 3p - (3p - s)/(L-1) = (3p(L-2) + s)/(L-1) is the tight one. ## Checked but not changed Clauses (a), (b), the alternation statement and the 3-window ceiling are untouched, and no other figure, table or section moved (`diff -u` is confined to the two clauses plus the tool's four header/`sha` lines).