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.

BTB-1 — "0 of 889 Russell-1000 names hard-to-borrow" is false at its own registration date (the true count was 2), and it is the stated basis for two pinned design decisions in three registered documents, one of them hash-frozen

status: open · raised panel run 22 (2026-09-14, frontier-WATCH catch; a records/trust correction, not a research or capture proposal — it does NOT reset the scout's 8-week cadence clock) · class: C0, registration hygiene (G5) — a pinned basis the shop's own store contradicts · effort ~0.25 pd (owner ruling) + ~0.1 pd (the two editable sites) · horizon: no deadline, but it should land before the shortability stream's own 60-trading-day test reads a panel everyone believes is all-zeros (~2026-10-20)


Plain-language summary for an owner reading one paragraph. Three registered documents say the binary hard-to-borrow flag is "degenerate" on the Russell-1000 universe — that 0 of 889 names were hard to borrow at the shortability capture's registration. That number is wrong. On the registration date itself, 2026-07-26, two names were hard to borrow (PSKY and TER), and across the 33 snapshots captured since, the flag fires on 24 of 33 days, one to eight names, mean about 1.5. This changes no conclusion — a one-to-two-name, non-persistent positive class still cannot power the hard-versus-easy separation test that was dropped, and the replacement (days-to-cover deciles) is the right instrument either way. What is wrong is the stated reason: "degenerate" and "empty positive class in-universe" are claims about the universe, and the universe does not support them. The shop's rule is that a decision traces to a criterion, and a criterion resting on a false measurement is the failure mode the external review's F1 finding named three months ago in a different file.

Ground truth — measured this run, twice, independently

The frontier lens swept the store; the adversarial reviewer re-derived it from scratch without reading that result first. Both agree.

Method: every data/shortability/*.parquet (33 files), inner-joined on data/universe.csv, counting borrow_status == "hard_to_borrow". No returns were touched; this is a universe-coverage count of the same class CAPTURES.md already publishes.

  • Registration date 2026-07-26: 889 in-universe names, 2 hard to borrow — PSKY (Paramount Skydance cl.B) and TER (Teradyne). Both also carry shortable=False and easy_to_borrow=False, so the count is 2 under every definition available in the schema, not an artifact of which column is read.
  • Across all 33 sessions 2026-07-26 → 2026-09-11: 24 days with at least one HTB name, range 1–8, mean ≈1.5, zero on 9 days. Peak 2026-09-08 (8 names). CPB recurs on 8 separate sessions, including 4 of the last 6 — so the positive class is not even uniformly transient.
  • Latest session 2026-09-11: 2 of 885 (CPB, OKE).

Where the false claim is load-bearing

sitetexteditable?
research/2026-08-01_b4_skew_activation_pin.md:43-46"That flag is degenerate on this universe — 0 of 889 Russell-1000 names were hard-to-borrow at the shortability registration" — the stated reason B4 replaced binary-flag conditioning with days-to-cover decilesNO — hash-frozen pin. A post-hoc edit is exactly the C0 breach the freeze exists to prevent. This needs a dated append or an owner ruling, never a correction in place.
CAPTURES.md:148"the binary borrow flag is degenerate here — 0/889 R1000 names HTB"yes
research/2026-08-01_quant_panel_suggestions.md:43B2's "the pinned hard-vs-easy separation test CANNOT RUN (empty positive class in-universe)"yes (a dated report; correct by append, not rewrite)
research/2026-08-01_quant_panel_suggestions.md:90"0/889 hard-to-borrow at rest"yes
research/2026-08-01_quant_panel_suggestions.md:121"the binary borrow flag is degenerate on the flagship universe (0/889)"yes
research/RESEARCH_QUEUE.md:2332a 2026-08-06 triple-kill of the weekly-option-availability candidate cites "the same degeneracy that killed the binary borrow flag" as one of its three legsNO — frozen pre-migration archive (read for prior-art, never written). Correction by pointer only.

The archive line is why this is worth an owner's ten minutes rather than a footnote: the false measurement has already been spent once as a kill justification for a different candidate. That kill survives on its other two legs, so nothing reopens — but a wrong number that has begun laundering into unrelated decisions is the compounding kind.

Note line 90's hedge — "at rest" — is the only place the original panel qualified the number, and it is the one phrasing that did not survive into either binding document. The pin and CAPTURES.md both carry it as a flat property of the universe. That is the specific way the claim hardened.

Why file it rather than fix it

Because the panel cannot fix the site that matters. 2026-08-01_b4_skew_activation_pin.md is hash-frozen and this row's whole point is that quietly editing a frozen registration is worse than the error it would correct. The owner's ruling is needed on one question: does this correction land as a dated append to the pin (leaving the frozen text intact and visibly wrong), or does the pin's hash get re-cut with a recorded reason? The two editable sites should be corrected either way, and should point at whatever the owner rules.

What must NOT happen, stated here so a later reader cannot mistake this row for an opening: this is not a reopening of B2 or B4. The separation test still cannot run at n≈1.5 non-persistent names; the days-to-cover replacement stands; nothing about the B4 activation pin's thresholds, dates or primary statistic is in question. Only the sentence explaining why.

Pre-registered kill criterion

This row is wrong and closes unbuilt if a re-run of the sweep above over data/shortability/*.parquet ∩ data/universe.csv returns 0 hard_to_borrow rows on 2026-07-26. One command, no judgement. (It returns 2 today: PSKY, TER.)

It is overtaken and closes if the owner rules that the frozen pin's sentence stands as written with the correction recorded elsewhere — that is a legitimate answer to the question this row asks, not a failure of it.

It is done when: the two editable sites state the measured figures, and the frozen pin either carries a dated append or is recorded as deliberately-left-standing with a pointer to the correction.

Effort · horizon

~0.25 person-days, almost all of it the owner's ruling on the frozen site. No code. No registry row and no DSR debt — this is a records correction, not a trial. No hard deadline; the natural one is the shortability stream's own pre-registered 60-trading-day read (~2026-10-20), which should not be interpreted by someone holding a false prior about the panel it reads.


Triage note — 2026-09-14 (annotation only; the body above is untouched)

Kill criterion re-run independently this run; it does not kill the row. Sweeping all 33 data/shortability/*.parquet inner-joined on data/universe.csv, counting borrow_status == "hard_to_borrow":

  • 2026-07-26 (registration date): 889 in-universe names, 2 hard to borrow — PSKY and TER. The criterion required 0.
  • 25 of 33 sessions carry at least one HTB name, mean 1.58, max 8. (The body says 24 and ~1.5; the one-day difference is a universe-snapshot artifact and is immaterial.) Latest session 2026-09-11: 2 of 885 — CPB, OKE.

Text sites confirmed verbatim: CAPTURES.md:147-148 and research/2026-08-01_b4_skew_activation_pin.md:42-46.

Why triage is ranking this above its own P4/P5 substance: it is the premise of a forward clock that binds in five weeks. CAPTURES.md:347-352 pins the shortability stream's minimum accrual at 60 trading days plus the pre-registered hypothesis "excluding HTB names from the momentum top-50 changes selection ≥1 name/month", measured BEFORE any return is looked at, with the explicit vacuity escape "if the screen never binds on R1000 names, the test is vacuous and the stream stays capture-only."

Counted from the accrued files this run: day zero 2026-07-27 (07-26 was a forced Saturday shakedown), 32 trading-day captures on the books through 09-11, so the 60th trading day lands 2026-10-19. This row's measurement is the input that vacuity escape turns on — a flag that fires on 25 of 33 days is not a flag that "never binds". Whoever reads that test in five weeks will read it against a panel three registered documents describe as all-zeros.

The ruling is the item. A routine must not edit the hash-frozen pin (2026-08-01_b4_skew_activation_pin.md) or the frozen archive (research/RESEARCH_QUEUE.md:2332, where the false figure was already spent once as one of three kill legs for an unrelated candidate). The single question is whether the correction lands as a dated append or the pin's hash is re-cut with a recorded reason. The two editable sites (CAPTURES.md, 2026-08-01_quant_panel_suggestions.md x3) are ~0.1 pd once that is answered.


TRIAGE ANNOTATION 2026-09-21 — the clock this row rides is now ~4 weeks out

data/shortability/ holds 38 capture files, 2026-07-26 → 2026-09-18 (two missing sessions, 08-07 and 08-28, both delayed-cron collisions). From day zero 2026-07-27 the pre-registered 60-trading-day read lands 2026-10-19/21 depending on whether day zero counts as day 0 or day 1 — call it four weeks. The false claim is unchanged at HEAD: CAPTURES.md:148 and research/2026-08-01_b4_skew_activation_pin.md:42-46 both still read "0 of 889".

The ruling has not moved and it is the whole item: dated append to the hash-frozen pin, or re-cut the hash with a recorded reason. No routine may edit frozen text either way. The two editable sites are ~10 minutes once the answer exists — and after 10-21 the correction lands after the test has been read against the wrong premise, which is the outcome this row was opened to prevent.


Triage annotation — 2026-09-28 (annotation only; the body above is untouched)

Still live at HEAD 864da4e. Nothing has corrected the premise. All four sites re-read verbatim this run: CAPTURES.md:148 ("0/889 R1000 names HTB"), research/2026-08-01_b4_skew_activation_pin.md:43-46 ("degenerate on this universe — 0 of 889"), research/2026-08-01_quant_panel_suggestions.md:43,:90, :121, and research/RESEARCH_QUEUE.md:2332 where the degeneracy is spent as a kill leg. The B4 pin has had no edit since 6701785 (2026-09-02) — no correction, no append.

The kill criterion is re-confirmed as NOT firing, from a new direction. No parquet reader is available in this container, so the universe join could not be re-run; instead the parquet footers were read directly. All 38/38 capture files carry a hard_to_borrow column, and the 2026-07-26 file's borrow_status column statistics carry min/max = hard_to_borrow / easy_to_borrow — i.e. at least one HTB row exists on the registration date. The row's pre-registered kill ("the check returns 0 on 2026-07-26") therefore does not fire, by a measurement independent of the one in the body.

THE MATERIAL CHANGE THIS WEEK — the 60-trading-day clock has split into two readings roughly a week and a half apart, and nothing rules between them. Measured: data/shortability/ holds 38 files, 2026-07-26 → 2026-09-18, with 08-07 and 08-28 as acknowledged holes, i.e. 37 accrued captures against the 60 required.

· calendar reading (day zero 2026-07-27 + 60 NYSE sessions) → 2026-10-19 · accrual reading (60 captures actually in the panel) → ~2026-10-28, and it slips one day per further dark session

Until the 2026-09-19 billing outage these two coincided, which is why every prior report quoted a single date. Six dark sessions (09-21…09-25, 09-28) have pulled them apart. CAPTURES.md:340-352 words the requirement as "Minimum accrual before ANY test: 60 trading days of panel" — which reads as accrual — but the wording has never been tested against a case where the two disagree. Choosing the reading after seeing either result is gate-shopping; the ruling belongs before 2026-10-19, the earlier of the two. This is carried as a decide item in the 2026-09-28 triage report, folded with the same ruling for options_iv_term.

Blocked on: the owner's ruling only — dated append to the hash-frozen B4 pin, or re-cut the hash with a recorded reason. No routine may edit frozen text either way, which is why the ruling is the item. Effort once ruled: ~10 minutes at two editable sites.


Triage annotation — 2026-10-05 (annotation only; the body above is untouched)

Both load-bearing sites unchanged at HEAD 502690b: CAPTURES.md:148 ("the binary borrow flag is degenerate here — 0/889 R1000 names HTB") and research/2026-08-01_b4_skew_activation_pin.md:43 ("0 of 889 Russell-1000 names were hard-to-borrow"), plus the three editable sites at research/2026-08-01_quant_panel_suggestions.md:43,90,121. Kill criterion does not fire and the row's own claim re-verifies exactly: on the registration date 2026-07-26 the snapshot carries 889 in-universe names, 2 hard-to-borrow (PSKY, TER).

THE CASE IS NOW STRONGER THAN WHEN THE ROW WAS WRITTEN, on a bigger sample. The store is 40 files, 2026-07-26 → 2026-10-02 — not 33 (body) or 38 (the 09-21 note). The flag fires on 32 of 40 sessions (body: 24/33; 09-21 note: 25/33), mean 1.95 names (body: ~1.5), max 8 on 2026-09-08. The latest session, 2026-10-02, carries 5 HTB (CHH, CPB, DG, GGG, MTN) — the 09-21 note's "latest: 09-11, 2 names (CPB, OKE)" is stale. CPB now recurs on 13 sessions. The vacuity escape hatch ("if the screen never binds") is further from firing than at any prior reading.

THE DEADLINE SPLIT IS NOW QUANTIFIED, AND THE TWO READINGS ARE 15 DAYS APART. CAPTURES.md:347-352 pins "60 trading days of panel". Measured from day zero 2026-07-27 through today:

  • elapsed NYSE trading days = 50 → the 60th is 2026-10-19 (the reading the row and all prior notes ride);
  • accrued panel = 39 sessions carry a capture; 11 are missing (2026-08-07, 08-28, and the whole outage block 09-21, 09-22, 09-23, 09-24, 09-25, 09-28, 09-29, 09-30, plus 10-05, for which nothing is committed yet) → 21 more sessions needed, i.e. ~2026-11-03 at the earliest, and later on any further gap.

The pin's literal words say panel, which is the accrual reading. Last week's note flagged that the two readings "coincided until the outage" and that the wording "has never met a disagreement"; it has now met one worth 15 days, and the disagreement grows with every dark session. Choosing the reading AFTER seeing which one gives a friendlier answer is gate-shopping — that is the whole reason the ruling is owed before 2026-10-19, the earlier of the two dates. Under the accrual reading the urgency is genuinely softer than the 09-21 note states; under the elapsed reading it is two weeks out. The row does not get to pick.

Rank this week: 6 of 12 (P1 — registration hygiene against a dated clock). BLOCKED on one owner ruling (dated append vs re-cut hash), and note that FRZ-1 is correct that the re-cut option is currently unexecutable: there is no hash test over this pin to re-cut. Rule FRZ-1 first and this becomes a one-line choice. No code, ~0.25 pd for the ruling plus ~0.1 pd for the two editable sites.

ANNOTATION 2026-10-09 (panel run 38; narrowed and corrected at adversarial review). THE WORD "NON-PERSISTENT" IN THIS ROW'S OWN BLOCKING REASON IS INACCURATE — AND CORRECTING IT STRENGTHENS THE ROW'S CONCLUSION, NOT WEAKENS IT. :14 and :78 both rest "the separation test still cannot run" on a "one-to-two-name, non-persistent positive class" / "n≈1.5 non-persistent names". The two prior annotations moved only the mean (1.58, then 1.95); neither touched persistence. Measured over all 44 sessions at HEAD c680133, restriction stated because it is a 4,302× trap — inner-joined on the 903-symbol data/universe.csv, since un-restricted the same column reads 8,994.5 HTB/day across ~14.4k Alpaca names: 92 in-universe HTB name-days across 57 distinct names, and CPB appears on 16 sessions (an 11-session maximal run, ended 2026-10-05) while MTN is HTB on six consecutive sessions (2026-10-01 … 10-08, unbroken and ongoing, consecutive in both the captured-session and the trading-day unit). So the absolute word is wrong. But 41 of those 57 names appear on exactly one session: the class is predominantly transient with two persistent outliers, which is a precision fix, not a falsification. And the fix cuts the row's way — 92 observations concentrated so that two names carry 22 of them have heavy within-name autocorrelation, so they are worth far fewer than 92 degrees of freedom. The "cannot run" conclusion becomes MORE robust, for a different reason: clustering rather than transience. This is explicitly not a reopening of B2 or B4 (:76 forbids that reading and the reading stands), and dismissed/intraday-shortability-sample stays shut — its legs are conjunctive and leg (b) is false twice over, at 44 of 60 accrual and with no runner written (grep -rn "hard_to_borrow|borrow_status|load_shortability|borrow_flags" src/ returns only the capture path data/borrow_flags.py / cli.py:2190 and the QA path data/captures.py; nothing computes the pinned hypothesis). One caveat on quoting the mean: 2026-10-08 carries 9 in-universe HTB names, 4.3× the 2.09 average — the trailing session is the panel's maximum, above the 8 of 2026-09-08. The third mean re-measurement this row would otherwise have received was dropped at review as padding.

Both readings of this row's 60-day deadline, on the record before the owner rules (research/steward/2026-10-08.md carries NEEDS-OWNER: pin the unit … captured sessions or elapsed trading days, and choosing after the result is visible is what C0 forbids):

streamcaptured sessionselapsed trading days60th, captured60th, elapsed
shortability44 of 60 (43 if the 2026-07-26 Saturday shakedown is not credited)53 of 602026-10-30 (44-basis) / 2026-11-02 (43-basis)2026-10-19
options_iv_term40 of 6048 of 602026-11-052026-10-26

This row's :192 (~10-28) and :241 (~11-03) are both superseded by the 2026-10-30 figure above. 2026-10-19 is the earliest date on the board. Two things a reader should know before the ruling: for options_iv_term it is substantively moot — the pinned statistic is 7.0 against a < 5 bar and fires on neither date, so that stream is not retired either way; for shortability it bites, because its 60-day gate opens a test rather than retiring a stream, and the earlier reading lands that completion date with no runner written. The panel takes no position on which unit is correct. Carried unchanged: this row's own ruling still waits on FRZ-1, which waits on an owner merge of scripts/automerge_guard.py.


ANNOTATION 2026-10-16 (panel run 39) — THE 2026-10-19 DEADLINE IN THIS FILE IS DEAD, and both of this row's stated blockers cleared on 2026-10-15. Read the header's "~2026-10-20" as history

Plain-language summary. This row is what the daily panel reads as its own expiry watch, and since Thursday it has carried a deadline that no longer exists plus two blockers that have both been removed. Monday 2026-10-19 is three days away; a reader of this file as it stood would treat that date as a fired clock and act on it. It is not.

What landed on 2026-10-15 (b927884, PR #240), and what it changes here:

  1. The 60-day unit was RULED, and this row's deadline moved 14 days. CAPTURES.md:375-383, verbatim: "RULED 2026-10-15 (owner): the 60-trading-day minimum counts CAPTURED sessions, meaning trading days with a file in data/shortability … So the earliest test moves from 2026-10-19 (the 60th elapsed session) to about 2026-11-02 (the 60th captured session; 48 were captured through 10-15). It was pinned before any test was run." Re-derived independently this run: data/shortability holds 49 files, less the 2026-07-26 Saturday shakedown the ruling does not credit = 48 session-dated captures. The ruling's arithmetic is correct. Therefore: this row's header ("it should land before the shortability stream's own 60-trading-day test … (~2026-10-20)") and its tail ("2026-10-19 is the earliest date on the board", "THE DEADLINE SPLIT IS NOW QUANTIFIED, AND THE TWO READINGS ARE 15 DAYS APART") are superseded. The split is resolved: captured sessions won. The live horizon is ~2026-11-02.

  2. FRZ-1, which this row was waiting on, is APPROVED. The tail carried: "this row's own ruling still waits on FRZ-1, which waits on an owner merge of scripts/automerge_guard.py." FRZ-1 moved open/ → approved/ by the owner in the same commit (research/queue/approved/frz-1-append-only-pins-not-enforced.md). The remaining wait is narrower and should be stated as such: not "FRZ-1 is unapproved" but "FRZ-1 is approved and awaiting the owner's own merge, because its build edits scripts/automerge_guard.py and no agent may merge that."

Why nothing updated this file. git log -1 -- research/queue/open/btb-1-*.md → d00429a, panel run 38, 2026-10-09 — no touch since the rulings. #240 annotated cqa-3 but not this row, and the steward's 10-15 "Carried forward" does not mention it either. The panel was dark 10-12 → 10-15, so this is the first run that could have caught it. That is the cost of the outage landing on the one routine that reads the clocks daily.

What is NOT changed by any of this. The row's substance stands untouched: the pinned "all-zeros" basis is still contradicted by the shop's own store, and the measured correction recorded on 10-09 (41-of-57 singletons, which strengthens the conclusion) is unaffected by a unit ruling about dates. The ruling settled when the shortability criterion may be read, not what it reads — CAPTURES.md:383 says exactly that of its sibling: "The ruling sets when the check runs, not what passes it."

Live state after this annotation: horizon ~2026-11-02 · blocker = the owner's merge of FRZ-1's build · substance unchanged · no deadline this week.

Expiry-watch instruction for the next reader, so this cannot recur silently: if any later panel or triage run is observed treating 2026-10-19 as this row's live deadline, that is evidence the annotation was not read and the header line itself should be struck rather than annotated.


Triage annotation — 2026-10-19 (annotation only; the body above is untouched)

Today is the date the header still names, and panel run 39 was right that it is dead. The deadline dissolved on 2026-10-15 when the owner ruled the unit (CAPTURES.md:375-383): the shortability test now reads at the 60th captured session, ~2026-11-02. Re-derived here: 50 files on disk, 49 session-dated (2026-07-26 is a Sunday), 48 through 10-15 — which is the pin's own number, exactly. The header at :3 and the table at :300 still read as live and should be struck, not annotated a fifth time, per this row's own expiry-watch instruction; triage has not struck them because a horizon line is row substance and editing it is not a triage act. Flagged to the steward instead.

Every number in this row is now LARGER, i.e. the row understates its own case. Measured over all 50 snapshots, inner-joined on the 903-symbol data/universe.csv, borrow_status == "hard_to_borrow":

as filed / last notedthis run
sessions on which the flag fires24 of 33 → 32 of 4042 of 50
mean HTB names per session~1.5 → 1.952.18
maximum in one session8 (09-08)9 (2026-10-08)
in-universe HTB name-days / distinct names / singletons92 / 57 / 41109 / 64 / 47

The registration-date claim reproduces exactly: 2026-07-26, 889 in-universe names, 2 hard to borrow (PSKY, TER) — against a pinned basis of zero. The singleton share holds at 73% (was 72%), so the "predominantly transient with persistent outliers" correction stands, and persistence is stronger than the 10-09 note recorded: MTN is HTB on 12 consecutive captured sessions, 2026-10-01 → 10-16, unbroken and still running (the note said 6); CPB 16 sessions with an 11-session maximal run. Latest session 2026-10-16: 2 of 883 (HIG, MTN).

All five text sites are verbatim at HEAD — CAPTURES.md:148, research/2026-08-01_b4_skew_activation_pin.md:43, research/2026-08-01_quant_panel_suggestions.md:43, :90, :121, plus the frozen archive at research/RESEARCH_QUEUE.md:2332, which still spends the degeneracy as a kill leg. #240 appended rulings to CAPTURES.md and touched no "0 of 889" sentence.

The blocker chain is now the actionable finding: one of this row's two options does not physically exist, and the other is available today. The "re-cut the hash" option requires a hash to re-cut, and there is none — grep -rln "b4_skew_activation_pin" tests/ scripts/ returns no hits, and scripts/automerge_guard.py carries no pin list. Creating one is FRZ-1, which the owner approved himself on 2026-10-15 and which has not been built: no commit or branch anywhere mentions a FRZ-1 build, and the implementer's only slot since approval (Wednesday 2026-10-14) was dark. So waiting for the hash option is waiting on a dark routine. The dated append is executable now, it is the option the steward recommended, and it is ~0.1 pd for the two editable sites. Recommendation to the steward: rule it as a dated append and stop holding this row behind FRZ-1.

Live state: horizon ~2026-11-02 · substance unchanged and strengthened · blocked on one ruling that no longer needs to wait for a build · no deadline this week.


ANNOTATION 2026-10-26 (panel run 44) — BOTH OF THIS ROW'S OPTIONS NOW EXIST; THE PARAGRAPH SAYING ONE CANNOT IS SIX DAYS STALE, AND THE HORIZON IS IN SEVEN DAYS

Still open; no new row. One fact, and it removes this row's stated blocker.

This file says, at :255-257, that "FRZ-1 is correct that the re-cut option is currently unexecutable: there is no hash test over this pin to re-cut. Rule FRZ-1 first and this becomes a [choice]", and at :419 that "one of this row's two options does not physically exist, and the other is available today."

FRZ-1 was built and merged as #280 on 2026-10-24 (fe81e8d). At HEAD 5c9d335, grep -rln "b4_skew_activation_pin" tests/ scripts/ — the check this file used to conclude no hash existed — returns tests/test_docs_consistency.py, tests/test_execution/test_automerge_guard.py and scripts/automerge_guard.py. The pin is tests/test_docs_consistency.py:413-445 (test_b4_skew_activation_pin_frozen, SHA a57bd03e454f97b10e2740b65c4ddcac123a7408ec459ef3e204c54436407bcf), and its own message says so in terms: a deliberate re-cut "updates this hash as a recorded act, which is the option BTB-1 was written assuming existed."

Checked rather than assumed: the two sentinel-hash tests are unmarked and tests/test_docs_consistency.py carries no module-level pytestmark, so they run inside the pre-trade gate's -m "not observability and not research" on both runners — the enforcement is real, not parked in a tier that is skipped. Only FRZ-1's generalising phrase-scan meta-test carries the observability marker.

So this row's remaining cost is one ruling, ~0.1 pd, and both endings are now available: a dated append under the B4 pin's sentinel stating the true hard-to-borrow count and whether the degeneracy-based design choice stands, or a recorded hash re-cut. Either closes the row. The steward may do this — the B4 pin is a research/ memo, not limit-4 text.

Why it should be ruled this week rather than next. This row's horizon is ~2026-11-02 (:341, restated at :367), which is also the day the shortability stream reaches its 60-captured-session minimum — the earliest permissible test of the HTB-screen hypothesis whose pinned basis ("0 of 889 R1000 names hard-to-borrow"; the true count on the registration date was 2) this row exists to correct. Two pinned design decisions in three registered documents, one of them hash-frozen, rest on that number. Ruling after the clock opens is ruling with the data on the table.

Kill criterion unchanged. This row dies when the dated append (or the recorded re-cut) states the true count and whether the design choice stands. It does not die on 2026-11-02 arriving: the horizon is the reason to rule, not an expiry of the defect.

Triage annotation — 2026-10-26 (annotation only; the body above is untouched) — MY 10-19 CLAIM IS WITHDRAWN: BOTH ENDINGS NOW EXIST

Correction to this routine's 2026-10-19 annotation, which said the "re-cut the hash" option "does not physically exist (there is no hash test over the B4 pin to re-cut)". It exists now. FRZ-1 shipped on 2026-10-24 (#280, fe81e8d): tests/test_docs_consistency.py:413 is test_b4_skew_activation_pin_frozen carrying SHA a57bd03e…, and scripts/automerge_guard.py:82 adds the file to DENY_PATTERNS. So both endings are now available, and the 10-19 inference that "the dated append is the only executable path" is withdrawn.

What that changes about who decides. A dated append below the sentinel is not frozen text and is within the steward's reach; a hash re-cut is owner-only (ops/STEWARD.md §0.5, "never edit hash-frozen text"). PR #287's annotation says "the steward may do this" — half right, and the half that matters is which option is chosen. Recommendation unchanged: the dated append, which needs one ruling and ~0.1 pd.

Not overtaken; nothing corrected. All five sites still carry the pinned-false claim verbatim at HEAD: CAPTURES.md:162, research/2026-08-01_b4_skew_activation_pin.md:43, and research/2026-08-01_quant_panel_suggestions.md:43, :90, :121.

The expiry in the header is superseded and should be struck, not annotated again. :3 says the ruling "should land before the shortability stream's own 60-trading-day test … (~2026-10-20)", and a later block names 2026-10-19 as "the earliest date on the board". Both are dead: the 2026-10-15 and 2026-10-24 rulings (CAPTURES.md:172-179, :377-396) made the unit captured sessions, shortability stands at 54 through 10-23, and the 60th lands 2026-11-02. This row's own expiry-watch instruction says to strike the stale line; #287 annotates it a sixth time instead. Striking body text is not this routine's to do — flagged.

Live-biting: records and trust, not trading. No code reads hard_to_borrow outside the capture (data/borrow_flags.py) and QA (data/captures.py:965), so the false premise bites on 2026-11-02, when the clock opens on it.


Triage annotation 2026-11-03 — THE HORIZON HAS ARRIVED. The clock this row was racing matured yesterday, with the false premise uncorrected.

The header's horizon reads: it should land "before the shortability stream's own 60-trading-day test reads a panel everyone believes is all-zeros (~2026-10-20)", superseded by the 2026-10-15 captured-session ruling to ~2026-11-02 (restated at :341 and :367).

Executed this run (thales capture-sessions, the pinned RAD-1 evaluator, installed via the ops/DAILY_AUDIT.md §0b fallback):

shortability — 60 captured sessions of panel before the HTB-screen
hypothesis may be tested at all
  files on disk: 61
  captured sessions: 60/60 (clauses 1+2 — the ruling's own arithmetic;
last counted 2026-11-02)
  clock MATURE — the criterion is due
  2026-07-26  14,137 rows  excluded by clause1_not_a_trading_day
  clause (3) is vacuous here: the stream records no capture instant

So the date is no longer a horizon to beat; it has passed in the confirming direction. The frozen premise is still on the page at every site: CAPTURES.md:162 ("0/889 R1000 names HTB"), research/2026-08-01_b4_skew_activation_pin.md:43 ("0 of 889 Russell-1000"), and research/2026-08-01_quant_panel_suggestions.md:43, :90, :121.

This does not kill the row — its own text says so, and the kill criterion (re-run the sweep against the 2026-07-26 file and find 0 HTB rows) is still live and unfired. What it does is remove the argument for waiting: the reason to rule was to get ahead of the test, and the test is now due.

Two mitigations, both verified, and they are the reason this stays in open/ rather than moving up the decide-list on its own:

  1. The 60-session mark is a FLOOR, not an appointment (CAPTURES.md:442-448, established by panel run 45 against two independent misreadings of :449-453 alone). Nothing is computed by the date arriving.
  2. The HTB-exclusion test is gated on "a future hash-frozen battery pre-registration" (CAPTURES.md:440-452) which does not exist. So no test can read the false premise yet — but writing that pre-registration is owner-only frozen text (ops/STEWARD.md §0 limit 5), and it would be written against these numbers.

Both endings exist (correcting my own 2026-10-19 note, which said the hash re-cut did not physically exist): tests/test_docs_consistency.py:433 test_b4_skew_activation_pin_frozen and scripts/automerge_guard.py:82. A dated append below the sentinel is within the steward's reach — a research memo append is not limit-4 text. A hash re-cut is owner-only. ~0.1 pd once ruled; no code either way.

The header's dead ~2026-10-20 date is now annotated for the seventh time rather than struck. Striking it edits the row's substance, which is not triage's move — recorded here so the next reader does not spend the same minute on it.


ANNOTATION 2026-11-09 (panel run 49) — the flag is not degenerate AND the screen is still vacuous; both are true, and a reader should not take the first to imply the second

Status unchanged, still open. This row is right and this annotation does not reopen or narrow it. It adds two measured things the body does not have, and separates a claim that a future reader could easily run together.

1. The flag CHURNS; it is not a fixed small number. The body establishes that 2 of 889 in-universe names were hard-to-borrow at the 2026-07-26 registration date, against the "0 of 889" written into five places. Measured here from the committed data/shortability/ panel, intersected with the 903-name data/universe.csv:

sessionHTB within the universe
2026-08-031 (RTX)
2026-11-024 (EQH, MTN, PTC, SATS)
2026-11-052 (MTN, SATS)

So the count moves between 1 and 4 across the window, and the names turn over completely. "Degenerate" is wrong in a second way the body does not claim: the flag is not just nonzero, it is time-varying. (The ~9,000 HTB names Alpaca reports are across its whole ~14k tradable list; only the intersection with the flagship universe is relevant to the pinned design decisions this row is about.)

2. The SCREEN is nonetheless vacuous on this universe, and that is not a contradiction. CAPTURES.md § shortability pins a separate hypothesis before any HTB test may run: "excluding HTB names from the momentum top-50 changes selection >=1 name/month." Measured here, returns-free (no return, equity curve or performance statistic was computed — only set intersections), by pairing each data/shortability/<d>.parquet with the same day's data/pit_snapshots/<d>.parquet and taking rank <= 50: 59 paired sessions, 2026-07-27 → 2026-11-05, and on all four monthly selection dates in that window — 2026-08-03, 09-01, 10-01, 11-02 — zero hard-to-borrow names sat in the momentum top-50. Independently recomputed twice in this run with the same result.

The two findings do not fight: the borrow flag is live and non-degenerate across the universe, and the handful of names it marks have simply never been among the 50 the strategy selects. A reader who takes this row's "not degenerate" as grounds to expect the exclusion screen to bind would be wrong, and the registered hypothesis is the thing that decides it.

What this does and does not discharge. The steward's 2026-11-05 carried-forward item — "Correcting BTB-1's false premise (a dated append) carries forward" — is not discharged by this annotation: the sites that need correcting are listed in the body's own table, one of them (research/2026-08-01_b4_skew_activation_pin.md:43-46) is hash-frozen and needs an owner ruling or a dated append there, and this panel may write neither. This annotation only supplies the fresher measurement that such an append would quote, so nobody has to re-derive it.