Thales
← research journal

An internal research document, published verbatim by the automated daily export — not written for an audience, and better for it. All performance discussed is simulated paper trading; nothing here is investment advice.

FRZ-1 — the B4 and B5 pins say they are append-only "like the December memo"; the December memo is enforced by a hash test and these two are enforced by nothing, while the auto-merge guard grades an in-place edit of either as ALLOW

status: open · raised panel run 30 (2026-09-29; panel-chair catch, verified with both a negative and a positive control — a records/trust defect row, not a research or capture proposal, so it does NOT discharge the scout's 8-week cadence clock; same treatment as BTB-1 and CQA-3) · class: C0, registration hygiene (G5) — a declared freeze with no mechanism behind it · effort ~0.25 pd · horizon: decidable immediately; the BTB-1 ruling it unblocks is already pending


Plain-language summary for an owner reading one paragraph. This shop freezes its pre-registrations so that nobody — including a future agent — can quietly rewrite the exam after seeing the answers. For five documents that freeze is real: a test computes a checksum of the protected text, and any edit above the line turns the test suite red. Two documents say they have that same protection and do not have it: the B4 option-skew activation pin and the B5 turn-of-month pin — the pre-registrations for the only two live forward experiments in the shop. Nothing computes a checksum for either. Worse, the automatic-merge guard treats an edit to either file as an ordinary record change and grades it ALLOW, so a routine's pull request that rewrote the B4 pin's pre-registered text would merge unattended with no human ever seeing it. This is not hypothetical bookkeeping: the open row BTB-1 is waiting on your ruling about a false number inside the B4 pin, and one of the two options it offers you — "re-cut the hash with a recorded reason" — cannot be executed, because no hash was ever cut. The fix is to give the two pins the protection they already claim.

Mechanism — verified at HEAD ada5bff, with both controls

1. The declaration. research/2026-08-01_b4_skew_activation_pin.md:221:

"Append-only after today, like the December memo: corrections and context arrive as dated appends below this line, never edits above it."

and :3 — "Status: BINDING, append-only after today." The B5 pin carries the same declaration at research/2026-08-01_b5_turn_of_month_pin.md:163.

2. The December memo really is enforced. tests/test_docs_consistency.py:316 test_december_memo_frozen_section_unchanged finds a sentinel line, hashes everything above it, and asserts a pinned sha256. The same pattern protects the micro-cap battery pre-registration, the A4 DTC memo, the frozen VRP v2 spec, and NORTHSTAR.md §1–§5.

3. Neither pin is in that set. grep over tests/ for b4_skew, skew_activation or the pin's own figure 889 returns nothing. The two pins are the only self-declared-append-only pre-registrations in research/ with no mechanical enforcement.

4. NEGATIVE CONTROL — the edit that should be impossible is invisible. In a scratch copy (never the repo), the pin's pre-registered §1 claim was inverted in place — :43 "0 of 889 Russell-1000" → "45 of 903 Russell-1000", and "degenerate on this universe" → "NOT degenerate on this universe", i.e. the reasoning for a pinned design decision reversed:

pytest tests/test_docs_consistency.py -q   ->  20 passed      (unchanged)

5. POSITIVE CONTROL — the mechanism works where it exists. The same class of edit, applied above the December memo's sentinel in the same scratch copy:

pytest tests/test_docs_consistency.py -q
  ->  FAILED tests/test_docs_consistency.py::test_december_memo_frozen_section_unchanged
      1 failed, 19 passed

So the detector is not broken — it simply does not cover these two files.

6. The auto-merge guard admits the edit. scripts/automerge_guard.py:44 allows the pattern ^research/[^/]+\.md$ and :100 ALLOWED_STATUSES = ("added", "modified"). Both pins sit at that path:

python3 scripts/automerge_guard.py \
  --files research/2026-08-01_b4_skew_activation_pin.md \
  --statuses research/2026-08-01_b4_skew_activation_pin.md:modified
  ->  ALLOW: records-only (1 file(s)); no gate touched

  (same for research/2026-08-01_b5_turn_of_month_pin.md)

Control — NORTHSTAR.md, which IS named in DENY_PATTERNS at :69 with the comment "hash-frozen span":

  ->  DENY: `NORTHSTAR.md` is in the never-auto-merge set

7. The path is live, not theoretical — and the point is that the guard cannot tell the two shapes apart. Commit 57308d3 (panel run 27, PR #179, a self-authored routine PR) carries M research/2026-09-18_two_forward_gate_numbers_were_wrong.md. That change was entirely legitimate — 127 insertions, zero deletions, i.e. a dated append, exactly the discipline these documents ask for. That is the finding. A dated append below the line and a rewrite above it are both modified on the same allowed path, the guard grades them identically, and no test distinguishes them for the B4 and B5 pins.

Why this is live this week rather than abstract

BTB-1 is blocked on a belief that is false. research/queue/open/btb-1-borrow-flag-not-degenerate.md is open and awaiting your ruling on the "0 of 889" figure. Its title says "one of them hash-frozen"; :48 says "NO — hash-frozen pin. A post-hoc edit is exactly the C0 breach the freeze exists to prevent"; and :159 puts a binary choice to you: "dated append to the hash-frozen pin, or re-cut the hash with a recorded reason."

The second option does not exist. There is no hash to re-cut. The queue's own memory is wrong about the protection state of the document the pending ruling concerns.

This run also measured how contagious the belief is: three independent agents in this single panel run — two lenses and an earlier reviewer — described the B4 pin as "hash-frozen" in their reports without checking. So did panel run 29 and the 2026-09-28 triage.

Separately, the B4 pin's 2026-09-02 append is still uncountersigned, and its clause (c) changes what counts as a valid signal date. A ruling on it is owed this week (see the run-30 note's expiry watch). Ruling on the contents of a document whose freeze is nominal is the wrong order of operations.

The fix

Copy the pattern that already works, verbatim — no new mechanism:

  1. Two tests in tests/test_docs_consistency.py modelled on test_december_memo_frozen_section_unchanged (:316): sentinel lookup, sha256 of everything above it, pinned hex. The B4 pin's sentinel is its ## 8. Change discipline block at :219-226; B5's is at :163.
  2. Add ^research/2026-08-01_b4_skew_activation_pin\.md$ and ^research/2026-08-01_b5_turn_of_month_pin\.md$ to DENY_PATTERNS in scripts/automerge_guard.py:61-73, beside the NORTHSTAR.md entry.

The PR must carry the negative control (this panel's own standard): show that the in-place edit class which passes today now reds, and that a dated append below the sentinel still passes. A freeze that also blocks legitimate appends would be a worse defect than the one it fixes.

Do NOT fix BTB-1's wrong prose by editing BTB-1 in place. Annotate it at triage. An in-place rewrite of an open row is the same act this row is about.

Pre-registered kill criterion

This row dies if either:

  1. You rule that the two pins are not append-only. The declarations at b4…:221 and b5…:163 are then removed by a dated append, and this row closes as mis-premised — the documents would simply be living records, and BTB-1's correction becomes an ordinary edit.
  2. Someone produces a test showing an in-place edit above either sentinel already reds the gate suite as CI actually invokes it — i.e. the negative control above does not transfer to pytest tests/ -m "not observability and not research". Then the protection exists and I measured the wrong thing.

It does not die on the argument that "git history is the record." That is precisely the argument the December memo's own test docstring rejects (tests/test_docs_consistency.py:318-321: "this memo claims append-only; git history alone guarded that").

Cost, honestly

research/queue/open/ holds exactly 12 items, which is the pre-registered cap (ops/QUEUE_TRIAGE.md:86-91). Opening this row means Monday's triage must evict an incumbent. The adversarial reviewer's eviction nominee is RWG-1 (the queue's oldest, raised run 11 on 2026-08-20; its first leg was discharged by you the same day, and the remainder has no date and no active bleed) — the same nominee run 29 named and did not spend. Not CQA-3 (evicting it fires CQA-2's reopen leg (c)) and not BTB-1 (this row exists to make BTB-1 decidable). The ranking call is triage's mechanical one on Monday, not the panel's.

Registry cost: zero. No results/research_registry.jsonl row, no trial, no deflated-Sharpe debt. This is apparatus, not a hypothesis.

What did not run

  • No broker, no config/.env, no data/raw, no data/lab, no owner-Mac access. None of the above needed any of them — every control is a text edit plus a test run plus one guard invocation.
  • The two hash values themselves are deliberately not computed here: cutting the hash is the build step, and the freeze must be cut from text you have agreed is the text to freeze.

Triage annotation — 2026-10-05 (first annotation; annotation only, body untouched)

REPRODUCED END TO END at HEAD 502690b, including by running the guard itself. The declarations are present (research/2026-08-01_b4_skew_activation_pin.md:221 under ## 8. Change discipline at :219; research/2026-08-01_b5_turn_of_month_pin.md:163). There is no hash test: grep -rn "b4_skew|skew_activation|b5_turn_of_month|889" tests/ returns zero hits, and the hash-frozen set in tests/test_docs_consistency.py is exactly five — :316 December memo, :336 micro-cap battery, :360 A4 DTC memo, :384 VRP v2 spec, :452 NORTHSTAR — each a sentinel-find plus sha256 against a pinned hex. Neither pin is in it.

The guard verdict was not inferred from reading, it was executed: python3 scripts/automerge_guard.py --files research/2026-08-01_b4_skew_activation_pin.md --statuses …:modified returns ALLOW: records-only (1 file(s)); no gate touched, exit 0 — same for the B5 pin. Positive control: NORTHSTAR.md:modified returns DENY: in the never-auto-merge set, exit 1. Cause unchanged: scripts/automerge_guard.py:44 admits ^research/[^/]+\.md$, :100 admits ("added","modified"), and DENY_PATTERNS (:61-74) names NORTHSTAR.md/CLAUDE.md/RUNBOOK.md but neither pin. So an in-place rewrite of either pre-registration auto-merges unattended today.

Number in the body now FALSE: it states research/queue/open/ "holds exactly 12 items, which is the pre-registered cap". It holds 14 (15 files including README.md) against the cap at ops/QUEUE_TRIAGE.md:86-87 — the queue is two over, and this run owes two overflow dismissals. The same stale "12" appears in CQA-3 and LAB-1; DGX-1 says 19.

Why this ranks where it does. It is the only row in the tree that is both (a) C0 registration hygiene — a declared freeze with nothing behind it, in a shop whose doctrine is that an unfalsifiable rule is not a rule — and (b) the precondition for another row's ruling: BTB-1's second option is "re-cut the hash", and there is no hash to re-cut until this is built. Ruling it first turns BTB-1 into a one-line choice.

Rank this week: 7 of 12 (P1). BLOCKED on: nothing — implementer- buildable at 0.25 pd, except that the row deliberately leaves the hash VALUES to be cut from whatever text the owner agrees to freeze, which is a one-minute owner input, not a design question.

Could not verify in this container: the negative/positive control pytest runs (no pytest/polars installable here — see the setup note in this week's report). Verified structurally and by executing the guard directly instead.


Approved by the owner — 2026-10-15

Approved by the owner himself, in an interactive Claude Code session that put him the steward's NEEDS-OWNER line of 2026-10-08/09 ("owner-only because it edits scripts/automerge_guard.py, which no agent may merge"). The question, as asked: two experiment rulebooks (the skew and turn-of-month pre-registrations) say "append-only" but nothing enforces it; an in-place rewrite of the skew rulebook's reasoning passes every check and would auto-merge; FRZ-1 adds the checksum lock the December memo has. Answer: Approve.

For the build:

  • Scope is the body's "The fix", nothing more:

    • two sentinel-hash tests in tests/test_docs_consistency.py on the December-memo pattern;
    • both pins added to DENY_PATTERNS in scripts/automerge_guard.py.

    The PR carries both controls: an in-place edit above either sentinel reds, and a dated append below it still passes.

  • Text to freeze: each pin as it stands above its sentinel at the build's base commit. This approval edits neither pin.

    • The B4 pin received a dated append today, the owner's countersign of its 2026-09-02 append. It sits below ## 8. Change discipline, outside the region the hash covers.
    • BTB-1's wrong number keeps its own pending ruling. Once this is built, its correction lands as a dated append or a recorded hash re-cut.
  • Merge: the build edits scripts/automerge_guard.py (STEWARD.md §0, limit 4), so no agent merges it. The steward reviews it and reports its verdict as a NEEDS-OWNER line; the owner merges it himself.


BUILT — 2026-10-21, implementer, PR #263 — outcome: shipped

Built to the owner's 2026-10-15 approval.

  • Two sentinel-hash tests in tests/test_docs_consistency.py, on the December-memo pattern verbatim. Frozen spans, cut from the text as it stands at base commit fad1d78 (this build edits neither pin):
    • B4 — sentinel "pinned dates only).*", lines 1-231 of 288 (§1-§8 + the Panel-Item provenance line, i.e. the whole original registration). sha256 a57bd03e…07bcf. Both dated appends — 2026-09-02's proposed refinement and the owner's 2026-10-15 countersign of it — sit in append space, as the approval requires.
    • B5 — sentinel "stated; verdicts pre-pinned; policy option decoupled).*", lines 1-175 of 176 (the whole document; no append has landed yet). sha256 5ee90ab8…d0237.
  • Both pins added to DENY_PATTERNS in scripts/automerge_guard.py, beside the NORTHSTAR.md entry, plus two committed guard tests so the control is not only a transcript.

Both required controls, run: an in-place edit above each sentinel REDS the suite (B4: the row's own 0 of 889 → 45 of 903 inversion; B5: the §7 policy-decoupling clause inverted); a dated append below each sentinel still PASSES. Guard executed, not read: both pins now DENY … never-auto-merge set, exit 1, matching NORTHSTAR.md; ordinary research records and new queue/open/ rows still ALLOW. Reverting the two DENY_PATTERNS lines returns ALLOW, and reverting them reds the new guard test.

Kill criterion verdict: NOT FIRED. Neither branch was reached — the owner approved rather than ruling the pins non-append-only (clause 1), and clause 2's "the protection already exists" was falsified by the controls above, which show it did not and now does.

BEYOND THE APPROVED SCOPE, flagged for the owner (one test, deletable in one block): a third test generalises the rule — any file under research/ declaring itself append-only must be named by a hash test here. It failed on its first run and named two more documents with this exact defect, which falsifies this row's own claim that B4 and B5 are "the only" ones: the C2 null-calibration memo (the binding convention for every PBO / IS-OOS-corr figure in the shop) and the C4 purge-leak memo (behind production cpcv.purge_days: 504). Both open with "Status: BINDING PRE-REGISTRATION, append-only after today." They are RECORDED, not silently passed, in a RECORDED_UNENFORCED dict citing the new row FRZ-2 (queue/open/), and are NOT frozen here — cutting a hash needs text the owner has agreed to freeze, which this row says for B4/B5 and which applies equally to them. This row's body is deliberately NOT edited to correct the "only two" claim: an in-place rewrite of an open registration is the act this row is about.

Still open: FRZ-2. BTB-1's second option ("re-cut the hash with a recorded reason") now exists for B4 and can be exercised.