Specialist desks debate every name — then the system stress-tests its own verdict.

Every sealed board with a directional stance, graded on the realized EXCESS return of its name vs the benchmark (a long call in a rising market is beta, not a call). The entry is a close printed AFTER the seal — never one that already existed when the board was sealed — and both legs are read on the same entry and mark sessions. Boards on the same name entered on the same session are ONE call, and calls entered on the same session are discounted for sharing one tape: a rate needs both enough independent calls and enough EFFECTIVE observations, and it ships with a Wilson interval computed on the effective count and only as many decimals as that sample supports. Conviction buckets are cut on the figure re-derived from each row's own sealed desk stances under the rule the house stands behind today, with the sealed figure published beside it. Synthetic boards never enter and are counted as a stated exclusion, as is any name with no usable price history. This measures the desks' calls — it is separate from the self-falsification record, and it is published whichever way it comes out.

Independence. 31 sealed directional boards resolve to 20 independent calls (boards on the same name entered on the same session are ONE call), spread over 5 entry sessions and worth 4.55 effective observations. Calls entered on one session share one tape, so every rate below is floored on the EFFECTIVE count, not the call count. Calls entered on the same session are treated as perfectly correlated (they share one tape). That is the worst case, so the true effective count lies between this figure and the nominal call count: the discount can only under-claim. Computed as effective observations = 1 / Σ(share of calls per entry session)² — the Kish count for a size-weighted rate.

Conviction basis. Calibration is graded on the conviction RE-DERIVED from each sealed row's own desk stances under the rule the house stands behind today, not on the figure the row was sealed under — grading a rule the house has superseded would measure nothing anyone is standing behind. The sealed figure ships beside it, and the record counts how many rows moved (superseded), already agreed (current), or reconcile to neither rule (unreconciled). Sealed bytes are re-read and re-labeled, never rewritten.

The conviction scale. Revision 2 divides agreement by the desks ELIGIBLE to agree, not by the whole panel — abstention is priced once, in the net score, instead of twice — and publishes participation beside the figure instead of folding it in. Revision 1 figures are not comparable to revision 2 figures and are never mixed into one rate. A board whose revision cannot be determined from its stamp or its own sealed desk stances is reported unreconciled, not assigned one. Revision 2 was AUTHORED 2026-08-13, before every board in the graded record; it has itself priced all 31 graded boards forward, and 0 of 31 graded boards are superseded rows re-derived at the read. Computed as |net score| x (desks on side / desks eligible to agree) x mean on-side calibration weight.

Board and call. A call is every sealed board on this name entered on the same session, counted once. Its conviction is the arithmetic mean of those boards' current-rule figures — and so is the sealed figure printed beside it — so neither will equal any single board's number. The boards themselves are published unchanged. A call over a single board carries that board's figure exactly and is marked with nothing.

Names not called. Boards that took no direction on names the house did not otherwise call that session. Counted, never graded — an abstention is not a miss. The ledger is DISJOINT from the calls on the same (name, entry session) key: a neutral board on a name other boards called is booked to the call ledger only, so one market move never carries two labels; those names are listed as also-called rather than dropped. The mean is over INDEPENDENT abstentions (one name, one session = one abstention), the same denominator the hit rate uses, and the per-board figure ships beside it. It is a mean ABSOLUTE move — a magnitude, not a forgone gain — over a handful of correlated names, so it carries no interval and is never set against a signed return.

One basis. how far the names moved against the benchmark, unsigned — a magnitude, not a gain. Both sides are computed on ONE measure — mean absolute excess return vs the same benchmark over the same window. This record previously set the abstentions' mean ABSOLUTE move against the calls' mean SIGNED return and called the difference a cost; that comparison implies a direction accuracy of 1.0, which is precisely the figure this panel withholds. No cost is claimed here, and no gain is attributed to a move nobody positioned for.

Sized through the gate. Each call is sized through the SAME capital gate the enforcement path runs: the conviction-band cap scaled by the board's panel participation, averaged across the boards in the call. No falsification escalation and no calibration trim is applied — those need live state this record does not re-create, so the permitted size here is an UPPER bound on what the gate would have allowed. The breadth multiplier is a POLICY CHOICE, stated in full beside this figure; a different curve would move the weighted figure and nothing in this record can say which curve is right.

Breadth is policy, not a measurement. The breadth multiplier is LINEAR BY POLICY CHOICE. A 1-of-4 board is permitted exactly a quarter of what a 4-of-4 board is permitted at the same conviction because the rate is applied to the first power — not because anything measured that a quarter is right. A square, a square root or a step would all be defensible; calibrating between them needs realized outcomes bucketed by participation, and the graded record stands at 20 independent calls on 5 entry sessions. Treat the curve as policy, not as a finding. Applied as permitted = the conviction band cap x the share of the panel that took a direction.

Where the exponent came from. Chosen for continuity — it returns the thin boards to the caps they carried under the superseded conviction denominator. Calibrated to reproduce the caps the superseded whole-panel conviction denominator produced for the three 1-of-4 boards (0.5% of book).

Observation, not expectation. A rate is an inference and is withheld below the floor. The mean of the realized returns is an OBSERVATION, and every return it averages is published per call in this same record — so withholding the average would not take it out of circulation, it would hand a reader an unqualified figure computed in their own head with none of this beside it. What is withheld is the EXPECTATION reading: no interval is printed until the effective observation count clears the floor the hit rate clears, and until it does, the dispersion, the median and the leave-one-out mean ARE the qualification the figure ships with. Dispersion here is across the calls; the noise scale measures each call against its own window, and the two answer different questions.

The scale. the standard deviation of this call's daily excess return over its own graded window, scaled up to the length of that window. Sigma is measured on the SAME bars the return is measured on — realized, not modelled, not annualized from elsewhere. It is a scale for reading one return, never a significance test: 20 calls on 5 entry sessions cannot support one.

The floor. At the observed accrual (0.9524 independent calls and 0.2381 entry sessions per day) the floor is at least 63 days away — a LOWER bound, because effective observations can sit below the entry-session count.Effective observations can never exceed entry sessions, so clearing the 20-effective floor requires at least 20 distinct entry sessions. Any projection here is therefore a LOWER bound on the time to a publishable rate.

• conviction 0-24 — 3 of 11 right, mean excess −0.87%, rate withheld — this slice carries 4.84 effective observations of the 5 required — 11 calls spread over 5 entry sessions
• conviction 25-49 — 4 of 9 right, mean excess −2.37%, rate withheld — this slice carries 3.52 effective observations of the 5 required — 9 calls spread over 4 entry sessions

• excluded — GOOGL: no close has printed since the seal — the window has not been observed yet
• excluded — AMZN: no close has printed since the seal — the window has not been observed yet
• excluded — NVDA: no close has printed since the seal — the window has not been observed yet
• excluded — GOOGL: no close has printed since the seal — the window has not been observed yet
• excluded — AVGO: no close has printed since the seal — the window has not been observed yet

marked 2026-09-16 · benchmark SPY · first close printed strictly after the seal instant — never a price that existed when the board was sealed

Mark Rule

the latest session BOTH the name and the benchmark have finished — finished meaning the tape has stopped printing for it (20:00 New York), not merely that the bell has rung, because a day print keeps absorbing late trades after the close. A session still trading is never marked, so two reads inside one session return the same figures: a close does not move

Return Rule

excess = name return − benchmark return over the same sessions; a short is right when the excess is negative

Sample Rule

rates are computed over independent calls, keyed by (name, entry session)

Abstention Rule

the abstention ledger is DISJOINT from the call ledger on that same key — a neutral board on a name other boards called that session belongs to the calls, and is listed as also-called rather than counted twice

Comparison Rule

abstained and called names are compared only on ONE basis (mean ABSOLUTE excess). A magnitude is never set against a signed return and never called a cost: that would assert a direction accuracy this record withholds

Independence Rule

a rate needs 20 independent calls AND 20 effective observations — calls entered on one session share one tape and are discounted for it, so twenty names on one day never clear the floor

Interval Rule

every published rate carries a 95% Wilson score interval computed on the effective observation count; computing it on the nominal count would narrow the band by exactly the design effect

Precision Rule

a rate is printed to the decimals its sample supports (a 20-observation rate resolves to 5 percentage points, so it prints to whole percent) — hits and n always ship, so the exact ratio is recoverable

Conviction Rule

conviction buckets are cut on the figure RE-DERIVED from each row's own sealed desk stances under the rule the house stands behind today, never on a superseded sealed figure; the sealed figure ships beside it

Sizing Rule

the weighted return sizes each call through the capital gate — conviction-band cap x panel participation — and the equal-weight figure it is set against is recomputed over the SAME sized calls, never over a larger set

Noise Rule

every call carries the realized sigma of its own daily excess series over its own window; a return inside one sigma is a direction that landed, and is reported as such rather than as a magnitude

Mean Rule

the mean call return carries the same discipline as a rate: its cross-sectional dispersion, its median and the mean without the single call that moves it most all ship beside it, and reading it as an EXPECTED return is withheld until the effective observation count clears the same 20 floor the hit rate clears

Split guard

a session move above 1.8x or below 0.55x inside the window excludes the name — unadjusted bars would read a split as a return