Calibration

Management guidance graded against the filed number. Each row is one executive; the grade accrues to the company's management until the attribution is upgraded. A calibration profile is a SOURCE-accuracy ledger. It confers no tier, no weight, no conviction, no corroboration.

32 management rows carry a filed comparison; the house row above is not a management row and is counted apart. 9 of the 32 were graded on a band the house chose, not on a tolerance the speaker stated; they are shown as awareness and count toward no hit rate. Mean bias is computed over the speaker-tolerance rows only. No row on this page counts toward any threshold until canon has defined, for that record type, the sampling unit and cluster — calls sharing an event, source or date window count as one — the cluster outcome, the null (p₀) taken from the base rate for that criterion type, and a look rule saying when the rate may be checked. The null is never assumed to be 0.50; for a two-sided band it depends on the width of the band. Comparisons between speakers are paired on the claims both made. Crossing a threshold triggers a ruling; nothing moves automatically.

mean bias: mean, over graded claims, of each claim's signed bias in per cent of the guided midpoint — (actual − midpoint) / |midpoint| × 100, averaged across the claim's graded legs; computed separately, never graded (R-MGMT-VOICE-2). Read from voices/management/<TICKER>/<person>.json at build.

Corrected 2026-09-18. Originally read: “mean bias: mean of (actual − point) / (high − low) over graded RANGE claims; computed separately, never graded (R-MGMT-VOICE-2). Read from voices/management/<TICKER>/<person>.json at build.” The definition described a range-width statistic; the figure shown is signed bias in per cent of the guided midpoint.

The mentor's priors, graded

Before each first filing read, a mentor names the signature the filing should show. The read then rewrites the target. Running record: 3 for 8.

• hits — GEV prepayments, TSEM capacity_reservation, VRT lead_time
• misses — AMAT lead_time, CEG take_or_pay, MRVL capacity_reservation (inverted), AAOI prepayments (falsified), ASML lead_time (not supported)
• unscored — TSM and INTC capacity_reservation (signature yes, rung no) (right signature, wrong layer)

Read from scout/rulings.json. A prior is a hypothesis, never evidence; the record grows one line per first read and is never averaged.

Two catches from the last review

The most recent scoring rep was drafted unattended and reviewed before admission. Two corrections were made at the review, and both are recorded on the ledger rather than smoothed over:

• A score was held, not raised. The draft read a peer's filing as support; the review read the same filing as adverse for the name (the constraint is real, the lock-in is absent) and held the thesis factor where it was. (review of 2026-09-06)
• A catalyst was scored on the canon scale, not on enthusiasm. The draft gave a confirmed near-dated event the top of the range; the scale caps proximity, and the review applied the cap. A rounding rule was written the same day: R-ROUND-TRUNCATE: quality and conviction are INTEGERS; a fractional binding factor TRUNCATES to the integer below it, never rounds up; the fraction is recorded beside the integer; a raise requires the BINDING FACTOR ITSELF to move, never the rounding.

Pre-registered: the 2026-09-30 print

Written 2026-09-06, before the event. THE SYSTEM'S FIRST PRE-REGISTERED CALL. Three blocks, each with its grading rule; outcomes enter this record after the 2026-09-30 print (release + 10-K). Nothing here scores until graded. Outcomes are graded after the print and enter the same ledger.

The scoring rep block states the model's factor values and is not published on this site; it is graded on the same record.

Read from ledger/performance-ledger.json (record MU, preregistered_20260930).