# Job #2987 — revision of `research/history/staging/attack-kstar-01.md` **One line:** the served document's §1 ladder table gains the row the reviewer asked for (19#→43#, K* = 20, cert = 21, C₂ = 618/150 = 4.1200, cert/C₂ = 5.10, route "pruned covering DP (`kstar_dfs.py`, #1144), period-free"), the "Full ladder (K*+1)" line is appended with 21 and its count corrected 16 → 17, and the four places that still treated the step as unrun now record it as reached — 16 anchored edits, every one of them named below, and nothing else in the file moved. Attempt `173f7d513da1d9de732dd0495fdb5552`, job #2987 (`audit`), session `99d9be65245bea7052b13091`, run `fb1-cd5ae88cb923bdba`, protocol `department-v2.research-2026-09-22.1`. Cites return **#1110**; the finding addressed is **#13** (`resolves`). The revision replaces nothing on the record: the served text is still at its own bytes (`1be1ded0…`), and this return carries the revised file plus a diff against it. | | sha256 | bytes | |---|---|---| | served base (what was edited) | `1be1ded08fc8ef1018df3019f1f24ad603505f8615a3184ec19d20c9df46e510` | 15813 | | this revision (`file`, in `files`) | `1200f6cac2d527e16f2ca2c39d19f3d0af383296d989e8750eca37ac1f3cff68` | 16864 | ## What changed, clause by clause against the finding The reviewer named four defects. Each is answered, in the document's own voice and formatting. **1. §1 ladder table — the missing row, and the enumeration it changes (7 edits).** The row is added verbatim in the table's columns, between 19#→41# and the "Full ladder" line. Adding it makes the document's own counts wrong unless they move with it, so they move: the title says *six* new steps (was five), the ledger verdict line counts 17 doubling steps (was 16), §0 says six new C₂ values and three period-free certificates (was five and two — 4.12 is inside the range the document already states, so the range itself is unchanged), §1 says three certificate-only steps, and the "Full ladder (K*+1)" line goes from 16 entries ending at 17 to **17 entries ending at 21**: `3 2 5 3 4 4 6 9 7 11 9 11 18 14 14 17 21`. §6's certificate list and §2's count of new C₂ values follow. The certificate's slack line is widened with it, 4.67 → **5.10** at 19#→43#. **2. The reviewer's "(s = 22, 23 by the P(2s) convention; check)" — checked, and correct.** The document uses `P(2s)` as *the largest prime ≤ 2s* (it is how §2 reads "cert = 18 at P(2s) = 31" and "the next chain step 31#→61#" at s = 32). Under that reading every row already in the table satisfies `P(2s)` = its target level: s = 15 → 29, s = 16 → 31, s = 17, 18 → 31, s = 19, 20 → 37, s = 21 → 41. The new row's s = 22, 23 → **43** falls out of the same rule (2s = 44, 46 → 43), so the reviewer's value is used as given rather than adjusted. `verify-revision.py` check **V5** re-derives this from the revised file itself — it does not trust this paragraph. **3. §3 "Where the engine's own reach ends" (1 edit).** The bullet now says 19#→43# *was reached by the pruned covering search in minutes* — K* = 20 over all 378,675 starts, `kstar_dfs.py`, #1144, rerun in review 2985 — where it previously listed the step among the ones this engine cannot afford. The rest of the bullet (23#→43#, 23#→47#, base 29+, still not run) is unchanged; those remain priced, not performed. **4. §7 NOT REACHED (3 edits).** Line 229 no longer says 19#→43# is "not run"; the bullet now names the instrument that ran it. Line 231 no longer rests 19#→41#'s K* = 16 on the I–E route alone — the pruned covering search returns 16 at that step too, which is a second, independent instrument — while keeping the honest half: the period 3.04·10¹⁴ still admits **no verifying walk**, so the confidence is the 14-step dual-engine validation plus that agreement, not a walk. §7's opening bullet also now scopes the fit to its 16 points (the new row is not in the drift fit) and says enumeration reaches 43. §2's I_eff bullet is scoped the same way. **5. §8 Sources and reproduction (2 edits).** The new row has a source row of its own (`kstar_dfs.py`, served, return #1144 — the pruned covering search, its K* = 20 and its certificate cells N_20 = 4, N_21 = 0) and a reproduction line: `python kstar_dfs.py 19 43` → `kstar: 20`, every start tested at length 21, so N_21 = 0; `--count 20` → `N_k: 4`. The line names `seconds` as wall time and as the only field that differs between runs — which is exactly what this run measured. **No embedded hash depends on the change.** The document's only hash-shaped content is the commit id `552143c`, which names the pre-registration commit, not the document's own text; nothing is re-embedded and nothing needed to be. The `node research/qc/embed.js --check` fingerprint belongs to `research/attack-kstar-01.js`, which this revision does not touch. ## Evidence, and how a reviewer re-derives it **`verify-revision.py` — 10 checks, exit 0, `verify-revision.out`.** Run from a fresh directory against the live server and the served instrument: * **V1** the document the server serves right now hashes to the `base` this revision names (`1be1ded0…`) — so the diff is against the text that was actually edited, not a stale copy; * **V2** re-applying `apply-revision.py` to those bytes reproduces the filed sha `1200f6ca…` byte for byte (exit 0, 16864 bytes); * **V3** the invariants the finding asked for hold: the ladder line has 17 entries and ends in 21, the 19#→43# row is present, and no falsified count survives (`five new …`, `16 doubling steps`, `P(2s) = 41 is inferred`); * **V4** the cited instrument still returns the frozen cells the document's other rows rest on: 13 29 → 10, 13 31 → 17, 17 31 → 13; * **V5** the `s` column against the document's own `P(2s)` rule, for every row (the check the reviewer asked for). **`reproduce-row.py` — the added row itself, exit 0, `rerun-kstar_dfs.out`.** Measured here, from a fresh directory, against the served `kstar_dfs.py` (`02104457…`, return #1144): `python kstar_dfs.py 19 43` → `{"D": 378675, "P": 19, "PP": 43, "Q": [23,29,31,37,41,43], "kstar": 20, "witness_start": 22352, "windows_tested": 378695}`, and the two certificate cells `--count 20` → `N_k: 4`, `--count 21` → `N_k: 0`. Field for field identical to #1144's published `dfs1943.json` and `dfs1943-counts.txt`; the frozen cells identical to its committed ones. The single field that differs is `seconds`: **175.6 s here against 329.2 s in #1144** — the document quotes #1144's own figure and attributes it to #1144, and does not restate a time this machine did not observe. The reproduction is deliberately byte-stable: `reproduce-row.py` drops `seconds` and never reads the progress lines (which carry elapsed seconds and go to stderr), so its stdout is what the assignment asked for — reproducible byte for byte elsewhere. `rerun-kstar_dfs.out` is that stdout, generated from the observed runs (`--from-dir`), so a reviewer's fresh `--full` run diffs against it. **Why the row is not this document's to re-derive, and why that is not a gap.** The 19#→43# row's substance — K* = 20, the witness, the certificate cells, the C₂ ratio 618/150, the 5.10 slack — is return #1144's result, already on the record and reviewed. This job's obligation is to carry it into the document *faithfully* and to say where it came from, not to recompute it; the recomputation above is corroboration, not a second claim. §1's "all five of this pass's new C₂ comparisons are custody-exact" sentence keeps its scope on the five the pass computed and attributes the sixth to #1144's own report on its own footing — the smallest change that keeps the sentence true. ## Scope, rung, and what is not claimed * Rung: the revision is **verified** in the finite, re-derivable sense — the diff is exact, the reproduction is mechanical, and the citations point at records that exist (#1110 the finding's return, #1144 the instrument and the row's source). Nothing here re-argues the mathematics of `attack-kstar-01.md`; the drift verdict, the HL identity and the trap grading are the author's and are not re-examined. * Not claimed: that 19#→43# is reachable by *this* document's I–E engine (it is not — §3 prices it at a ~2.5·10¹⁰-subset pass and says so); that the 175.6 s here reproduces #1144's 329.2 s (a covering search's wall time is machine-dependent, which is why the document names `seconds` as the field that differs); or that any other stage document needs the same row — this is scoped to the one file the finding names. * Left as the reviewer's own words rather than smoothed over: the `s = 22, 23` reading and the "period-free" route label are kept, because both check out against the document's own conventions (V5 and §3) rather than needing editorial repair. * **Accounting, stated rather than implied.** This assignment was issued *inside* the turn that then did it, and the harness writes a turn's usage only when that turn closes, so the log that can be posted at submission time carries the instruction and not yet the turn. Per the accounting rules the return is filed with the observed transcript and the usage left **explicitly pending** — no number is estimated from elapsed time or text length, and no second return claims anything — with the next action recorded: re-export the completed log for this thread's own turn range and post it once to `/return//transcript` with the per-turn counts, never re-submitting the scientific result. The turn that took, released and re-took this assignment is a single app-turn boundary, and the released `#3059` attempt's own turn is deliberately **not** carried here or claimed anywhere. * For my person, as the brief asks: **15 of @maxime-fleury's returns wait for a verdict**, the oldest since 2026-09-17. Only a trusted reviewer — a person granted trust, or a model trusted at a top thinking level, never this session's own handle or model — can decide them; no agent of the person's can, on any model.