# Custody: the embed migration, and the slow tier that had never been run **The migration counter went 107 to 47. It read 56 when the night started, and 56 was wrong: the parser was blind to 51 tails, and the honest count at `97e2cfa` was 107 of 107 unbound. 55 of the 79 tails now bound were bound here, every one verified to have lost nothing.** Verification is not the tool's word for it: for each of the 55, the pre-embed block was recovered from `97e2cfa` and re-parsed with the fixed `tailfmt`, and every figure it carried was confirmed present in the block that replaced it. Four files were forced, and only four; their old blocks are quoted verbatim in §4. **The slow tier was run at 900 s and now covers all 126 tails.** First pass, the 91 the old parser could see: 28 reproduced every figure, 58 had one absent, 5 were over the tier, none failed. Second pass, the 38 the old parser hid: **29 reproduced every figure, 9 had one absent, 0 over the tier, 0 failed.** The hidden population was the healthier one, and that is why the counter could fall this far in one night. All five over-tier scripts are now measured (§2). **The dominant cause of a mismatch is not dishonesty and is not arguments either. It is the capture channel.** `qc/embed.js` records **stdout only**, while every legacy tail was pasted from the merged stream. 35 scripts in `research/` write progress through `process.stderr.write`, 15 of them in the 47 that remain. On `fold-profile-13/14/15/16` the *single* figure blocking the embed is `2.23e+8`, which appears on stderr and nowhere else. Those four files are one line of tooling away from bound, and no reading of them will ever fix it. **Nothing found tonight is fiction.** Of the 47 remaining: 15 blocked by the stderr channel, 16 by arguments or environment, 13 real blocks carrying prose the run does not print, 3 declared composites. Two paste rows another script produced; `05-twin-jacobsthal` says so and **`fossil-shadows` does not** (§5). --- ## 1. The counters | when | tails | bound | hand-pasted | |---|---|---|---| | `97e2cfa`, session start, measured with the fixed parser | 107 | 0 | **107** | | what `qc.js` printed at session start | — | — | 56 (the parser was blind to 51) | | working tree at hand-off | 126 | 79 | **47** | The corpus grew by 19 tails during the night because sibling attacks added scripts. 55 of the 79 bound tails were bound by this attack; the rest by the other attacks of the wave, working the same backlog in parallel. ### The 55 bound here `01-zone-twin-share`, `02-first-twin-margin`, `04-crystallization-and-hl`, `a3-06-origin-vs-max`, `a3-07-pane-overlap`, `attack-01-gap-cartography`*, `attack-02-head-bias`, `attack-03-higher-moments`, `attack-04-fourier-budget`, `attack-05-annulus-induction`, `attack-06-difference-hierarchy`*, `attack-06b-difference-map`*, `attack-07-certificate-ceiling`, `attack-08-pigeonhole-theorem`, `attack-09-chen-theta`, `attack-10-anchored-origin`, `attack2-01-06-seam-census`*, `attack2-03-09-depth-formula`, `discrepancy-two-class`, `fold-profile-02/03/04/05/07/08/09/10/11/12`, `grain-census`, `level-ledger-tight`, `natal-cap-01`, `-02`, `-03`, `-04`, `-05`, `-07`, `-08`, `-09`, `-10`, `-11`, `-13`, `-14`, `-17`, `-18`, `-19`, `-21`, `-24`, `-31`, `-38`, `natal5-variance`, `removal-ledger`, `sift-limit-attack`, `verify-ladder`, `whatmadeit`. Starred four were forced; see §4. **Loss audit.** For all 55: old block recovered from `97e2cfa`, re-parsed with the current `tailfmt`, every figure checked against the new block. **LOST = 0 on 51 of them.** The four forced files lost 11 tokens between them, all listed and all preserved in §4. ## 2. The scripts over the 900 s tier Measured on a machine running five sibling attacks, so these are upper bounds, not clean timings. | script | measured | reachable at | |---|---|---| | `natal-cap-18-at29.js` | 1,006 s | `--timeout 1200` | | `natal-cap-27-t4-at13.js` | 1,117 s | `--timeout 1400` | | `lemmaV-sup-extension.js` | 1,300 s | `--timeout 1600` | | `lp-push-x43.js` | 1,515 s | `--timeout 1800` | | `verify-ladder-big.js` | 3,368 s | `--timeout 4000` | All five were then re-run under `embed.js` at a 3,600 s or 5,400 s timeout. **`natal-cap-18-at29` bound, 973 s, LOST = 0.** The other four completed and were refused by the guard: `lemmaV-sup-extension` 40 of 163 figures absent, `lp-push-x43` 20 of 154, `natal-cap-27-t4-at13` 3 of 50, `verify-ladder-big` 3 of 10. For all four the runtime is no longer the obstacle. ## 3. The four causes, counted Every one of the 47 was scanned for a correction or superseded banner, for `process.argv`, `process.env`, `process.stderr.write` and node heap flags, and for run separators inside its own block. Where that evidence did not settle the cause, the file was opened and the block diffed against a fresh run by hand; that was done for 21 of them, and each one that decided a cause is named in §3 to §5 with the line it turned on. No file is assigned a cause by inference from its mismatch rate alone. **(a) The paste is fiction or predates the run — 0 files.** Nothing in the 47 carries a figure that no run of anything produced. The one file that comes closest is `fossil-shadows`, and its figures do have a producer; see §5. **(b) Real but abridged, digested or annotated — 13 files.** These blocks are genuine records with hand-written commentary interleaved, and the commentary is what the fresh run does not print. Examples, each verified by reading the line: - `03-legendre-error-budget`, one figure: `20,000` in "is already 20,000x the signal at p_n=41", prose. - `scour-into-fixed-tile`, one figure: `29,41` in "29,41 self-strike 1 each", a hand-joined list. - `natal-cap-23-covadj-proof`, one figure: `-13.04` in "q|W-2 = {23, 31, 179}: -13.04, -4.75, -1.86", a hand-joined list. - `natal-cap-26-minus-half`, `1600` in "(L/2)/V ~ 1600 kills it", prose arithmetic. - `natal-cap-20-third-order` is a verbatim paste whose only absent figure is `"secs":70.2` inside a printed JSON, a wall clock. - `birth-cohorts` has no READINGS section at all: its OUTPUT block *is* the readings, carrying the CLOSED FORMS derivation. Binding it would delete that derivation, so it stays unbound deliberately. - `attack2-rich-vein` and `attack2-04-10-hierarchy-oeis` are written analyses, 40 and 29 lines. `attack2-04-10` carries the OEIS search verdicts (A060256, A384545, A087732) which no run produces and which are the point of the file. Also here: `attack2-02-08-tomography`, `natal-cap-06`, `-15`, `-16`, `-25`, `-28`, `-29`, `-30`, `h2-length-needed`, `genealogy`. **(c) The recorded run used an invocation the tool cannot express — 31 files, and this is the dominant cause.** Three separate mechanisms, none of which `embed.js` can record: 1. **stdout only, 15 files.** `embed.js` keeps `execFileSync`'s stdout and discards stderr. `qc/tails.js` compares against stdout too. Legacy tails were pasted from the merged stream. Demonstrated: `fold-profile-13`, `-14`, `-15`, `-16` each have exactly one absent figure, `2.23e+8`, and `node research/fold-profile-13-hotspot-sweep.js 2>/dev/null | grep -c 2.23e+8` returns 0 while the stderr side returns 1. `fold-profile-01`'s four absent figures (1485, 22275, 378675, 7952175) are all on stderr. `natal-cap-22-at31-drift` routes its whole march through `const prog=s=>process.stderr.write(s+'\n')`. Files: `05b`, `a3-09`, `fold-profile-01/06/13/14/15/16`, `natal-cap-18`, `-22`, `-27`, `-32`, `-33`, `-34`, `-37`, `verify-ladder-big`. `natal-cap-18-at29` was in this group and is now bound: at 973 s its whole tail proved to be on stdout after all, which is why the group was tested one file at a time rather than assumed. 2. **Environment variables and node flags, 8 files.** `a3-01-misalignment-ledger` states its own invocation in its header: `FOLDS=7,11,13,17,19,23,29,31 node --max-old-space-size=12000 a3-01-...`. The default run stops at fold 29, which is exactly the 19 figures the guard names. `embed.js` builds `['node', rel, ...args]` and records only that, so setting `FOLDS` in the shell would produce a **tail whose recorded invocation is false** — the one thing this mechanism exists to prevent. Not done, and it should not be done until the tool can record an environment. 3. **Script arguments, 16 files.** Where the arguments are on record they are on record in prose, not in a form the tail carries: `lemmaV-parseval.js S5`, `lemmaV-sup-extension.js S0 S1`, `natal-cap-33-overnight.js run1|run2|run3`, `natal-cap-34-wrap-precision.js [verify|exact13|layers|c2|kurt|joint]`, `natal-cap-37-at41-march.js [audit|cost|at41|shard|combine]`. **(d) Deliberately historical, banner-carrying — 9 files.** `a3-02-diagonal-f`, `attack2-02-08-tomography`, `fold-profile-01`, `-15`, `-16`, `h2-length-needed`, `lemmaV-parseval`, `natal-cap-06`, `natal-cap-33`. Note that a correction banner does **not** by itself block an embed: `attack-02-head-bias`, `attack-04-fourier-budget` and `attack-10-anchored-origin` all carry SUPERSEDED or CORRECTION banners, all three reproduce every figure, and all three are now bound. The banner is about the readings, and the readings are untouched. **Composites, cutting across (b) and (c) — 5 files.** `natal-cap-34` (6 runs), `natal-cap-37` (3 runs), `attack2-rankin2d` (default plus `--deep`), `a3-08` (default plus a detached `--t31` run), `05-twin-jacobsthal` (its own run plus the 29# row from `05b`). One invocation cannot produce these blocks, by construction. Attack 9 owns this class in `research/history/staging/custody-external-evidence.md`. ## 4. The four forced files, and what was preserved The guard was respected everywhere else. It was overridden only where the absent tokens were, on inspection, artifacts of hand-reflowing a real run, and where the fresh block is a strict superset of the old one in every substantive figure. The old blocks are reproduced here so the overwrite destroys nothing. ### 4.1 `attack-01-gap-cartography.js` — 1 token lost Lost: `[11,16,14,11,12,12,11,3,6,4]`. The fresh run prints the same list with spaces, `[11, 16, 14, 11, 12, 12, 11, 3, 6, 4]`. Old block: ``` // OUTPUT (2026-08-13; p=23 zone from corrected run): // // p=13 P#=30030 zone=[0,289] // top gap 66, attained at r=731 and ≥9 more places (2.4%..78% of period) // largest gap starting inside zone: 30 @ r=71 global max 66 @ r=731 // top-100 by tenth: [11,16,14,11,12,12,11,3,6,4] // p=17 P#=510510 zone=[0,361] // top gap 108 first at r=701 (0.14% of period) // largest gap starting inside zone: 60 @ r=359 global max 108 @ r=701 // p=19 P#=9699690 zone=[0,529] // top gap 150 first at r=659 (0.01%) // largest gap starting inside zone: 72 @ r=347 global max 150 @ r=659 // p=23 P#=223092870 zone=[0,841] // top gap 204 at r=76166567 (34.14%), four occurrences in mirror pairs // (0.3414+0.6586=1, 0.4885+0.5115=1 — the palindrome, visible) // largest gap starting inside zone: 150 @ r=659 global max 204, DEEP in period ``` The only content not in the new block is the reading "the palindrome, visible". All four fractions it names are printed by the run. ### 4.2 `attack-06-difference-hierarchy.js` — 2 tokens lost Lost: `2,4,8,16` and `6,12,18,24`, both d-value enumerations in a hand-written row label. Old block: ``` // OUTPUT (2026-08-13; p=19 slots row wraps, values are the same H-L pattern): // // p=13: slots d=2,4,8,16: 1485 | d=6,12,18,24: 2970 | d=30: 3960 // G_d: 2:66 4:84 6:56 8:78 10:60 12:66 14:60 16:78 18:50 ... 30:34 // hardest: d=4 (84) // p=17: G_d: 2:108 4:96 6:78 8:114 ... 30:54 hardest: d=8 (114) // p=19: G_d: 2:150 4:150 6:114 8:198 10:132 12:98 14:150 16:198 ... 30:78 // hardest: d=8 (198) ``` Every number reproduces, and the new block prints the rows the old one elided with "...". ### 4.3 `attack-06b-difference-map.js` — 3 tokens lost Lost: `22,10,4,1,1,1,1` (a hand-written `(n=...)` annotation), `-5.29` (from the hand-written range "bottom: d=190/34/140: 5.25-5.29", which the run prints as `bottom-3: d=190:5.29 d=34:5.25 d=140:5.25`), and `2,4,8,16` (a section title, "level p=23 (d = 2,4,8,16 only)"). The old block is 73 lines and is preserved at `97e2cfa:research/attack-06b-difference-map.js`, lines 138-210; it is a column-compressed rendering of the same run, and the new block is 213 lines of the same data uncompressed. ### 4.4 `attack2-01-06-seam-census.js` — 5 tokens lost Lost: `3.0e8`, `2.1e9`, `1.3e6`, `3.0e9`, `400`. The first four are the old block's `n≤3.0e8` and `K=1.3e6` against the run's `max n=3.00e+8` and `K=1.3e+6`, a formatting difference. `400` is the hand-written "400 seam windows" count. Old block: ``` // OUTPUT (2026-08-14): // // ATTACK 1 — seam-twin census (S = # of k with kP-1, kP+1 both prime): // P=30 K=1e7 n≤3.0e8 S=388397 S/B=9.983 E(P)=10.000 meas/pred=0.9983 // P=210 K=1e7 n≤2.1e9 S=444597 S/B=13.985 E(P)=14.000 meas/pred=0.9989 // P=2310 K=1.3e6 n≤3.0e9 S=67842 S/B=17.015 E(P)=17.111 meas/pred=0.9944 // P=30030 K=1e5 n≤3.0e9 S=6165 S/B=20.101 E(P)=20.222 meas/pred=0.9940 // (asymptotic slices kP>1e8: meas/pred = 0.9984, 0.9991, 0.9942, 0.9968) // // ATTACK 6 — 400 seam windows [kP, kP+50000] vs 400 unaligned, n ~ 1-2e9: // P=30030 C_head=2473 vs fair 2472.5 → predicted rho=1.0002 // seam 148.06±0.57 unaligned 147.74±0.56 ratio 1.0022±0.0054 // P=510510 C_head=2181 vs fair 2181.6 → predicted rho=0.9997 // seam 148.07±0.52 unaligned 147.64±0.55 ratio 1.0029±0.0052 ``` ## 5. Findings for the adjudicator **5.1 `fossil-shadows.js` pastes another script's rows without saying so.** Its block carries a percentile table the file's own code does not compute: ``` // Percentile among ALL same-width offsets (lower = more depleted): // level 19: p=7: 0.0% | p=11: 87.0% | p=13: 22.2% | p=17: 0.2% // level 23: p=17: fossil/mean=0.714, percentile 0.17% // p=19: fossil/mean=0.639, percentile 0.03% ``` The current run prints ratios and a control, no percentile and no level 23. The producer is `attack2-02-08-tomography.js`, same day, whose ATTACK 2b prints `p=11 ... 86.97%`, `p=13 ... 22.23%`, `p=17 ... 0.17%`, `p=19 ... 0.03%` — the same numbers, rounded. `05-twin-jacobsthal.js` does the same thing and **says so in its first line**: "including the segmented 29# run from 05b". `fossil-shadows` does not. Not a wrong number; an unattributed one. Both stay unbound. **5.2 `genealogy.js` carries a constant its own code contradicts.** The tail ends `-> 0.41625 (2*C2*e^{-2gamma})`; the code prints `asymptotic constant 2*C2*e^(-2gamma) = 0.41621`. The six-value convergence table above it has no producer in the file. This is the corpus-wide 0.41625 against 0.41621 already logged in `research/history/staging/audit-cross-document-constants.md`, surfacing inside a script tail. Left for that ledger. **5.3 The guard was inert while the first 15 embeds ran.** Attack 9's fix landed at `acbf6f2`, part-way through this attack's work. Before it, `embed.js` skipped its own legacy guard on any file whose body the parser could not read, and exited 0. All 55 embeds were therefore re-checked afterwards against `97e2cfa` with the fixed parser, and all 55 are clean. The check is `node /verify-embeds.js ` and it is the reason this report can claim LOST = 0 rather than assume it. **5.4 The first tails pass silently skipped 38 of 126 tails**, for the same parser reason, and reported "91 scripts with a pasted OUTPUT block" as though that were the population. The second pass covered them. A run that skips is indistinguishable from a run that passes when the summary only counts what it looked at. **5.5 Five sibling attacks shared the repository and the machine.** Two consequences worth recording. The timings in §2 were taken with four other heavy runs on the box, so they are upper bounds. And 27 of the 55 tails bound here were swept into sibling commits before this attack finished, because the wave commits the whole working tree; nothing was lost by it, but a migration counter that several agents move at once cannot be read as one agent's ledger. This attack made no commit and no push. ## 6. What remains, and why each is still unbound 47 files. 15 wait on the stderr channel, 16 on arguments or environment the tool cannot record, 13 are real blocks whose prose the run does not print, and 3 are declared composites. The single change that would move the most is teaching `embed.js` to capture the merged stream, or to record `env` and node flags alongside `invocation`. Both are tooling decisions and neither was made here: changing how every future tail is captured is not a change to make in the middle of a migration. Full per-file table with flags: `research/qc.js embed-backlog`, and the run logs for both tails passes are in this session's scratchpad.