Measured claims

Every public headline number

One row per number the lab has published, newest first, with the interval where the piece gives one. Quote the number with its campaign and link to the piece. 60 claims.

Date	Claim	Number	Interval	Campaign	Source
2026-10-07	First round: open head decision-class accuracy, 800 eval cases, zero-shot	0.545	none	The open decision head	Executive piece
2026-10-07	First round: rule table decision-class accuracy, 800 eval cases	0.575	none	The open decision head	Executive piece
2026-10-07	First round: open head minus rule table, 800 eval cases (inconclusive)	-0.029	+0.001 to -0.060	The open decision head	Executive piece
2026-10-07	Jev (hosted) decision-class accuracy on the same 800 eval cases	0.540	none	The open decision head	Executive piece
2026-10-07	Jev decision-class accuracy on the 200-case contradictory-evidence set	0.536	none	The open decision head	Executive piece
2026-10-07	Second round: retrained head decision-class accuracy, 800 fresh cases (eval2)	0.728	none	The open decision head	Executive piece
2026-10-07	Second round: rule table decision-class accuracy, 800 fresh cases (eval2)	0.566	none	The open decision head	Executive piece
2026-10-07	Second round: retrained head decision-class accuracy, harder set	0.739	none	The open decision head	Executive piece
2026-10-07	Second round: rule table decision-class accuracy, harder set	0.565	none	The open decision head	Executive piece
2026-10-07	Retrained head ahead of Jev's 0.536 on the 199 cases both scored, harder set (gap excludes zero)	0.199	none	The open decision head	Executive piece
2026-10-07	Smaller fine-tuned head decision-class accuracy, fresh set (experimental label)	0.701	none	The open decision head	Executive piece
2026-10-07	Wrong-claim rate, retrained head, fresh set, before the evidence-pack fix (above the registered trigger)	0.047	none	The open decision head	Executive piece
2026-10-07	Wrong-claim rate, retrained head, after the evidence-pack fix	0.0113	none	The open decision head	Executive piece
2026-10-07	Wrong-claim rate, rule table, after the evidence-pack fix	0.0088	none	The open decision head	Executive piece
2026-10-07	Wrong-claim rate, learned head, after the evidence-pack fix	0.0163	none	The open decision head	Executive piece
2026-10-07	Expected calibration error of the first head (registered bar 0.08)	about 0.52	none	The open decision head	Executive piece
2026-10-07	Item-level calibration error of the later judging head, full eval2: uncalibrated 0.0178, refit	0.0143	none	The open decision head	Executive piece
2026-10-07	Jev single-question response time at $0.0000128 for 304 input tokens, measured 24 September 2026	0.35 seconds	none	The open decision head	Executive piece
2026-10-07	Open head fast-mode median response time, four benchmark streams	3.1 seconds	none	The open decision head	Executive piece
2026-10-07	Open head fast-mode p90 response time, four benchmark streams	4.8 seconds	none	The open decision head	Executive piece
2026-10-07	Checkpoints logged in the J5 eval2 run, across 794 cases (about 3.4 per case)	2,681	none	The open decision head	Executive piece
2026-10-07	Share of repeats that disagree, Jev, first run (open head 4.95% eval, 5.85% challenge, with a second job on the endpoint)	0.05%	none	The open decision head	Executive piece
2026-10-07	Share of repeats that disagree, open head alone at one call in flight (both processes)	0.05%	none	The open decision head	Executive piece
2026-10-07	Share of repeats that disagree, open head at four calls in flight, onboarding (accounts payable 1.95%)	3.4%	none	The open decision head	Executive piece
2026-10-07	Second process (accounts payable): open head accuracy on ap-eval, against rule table 0.648	0.650	none	The open decision head	Executive piece
2026-10-07	Second process: learned head trained on the first process, no retraining, ap-eval	0.522	none	The open decision head	Executive piece
2026-10-07	Second process: learned head retrained on 100 of the new process's labelled cases (271 decisions), ap-eval	0.663	none	The open decision head	Executive piece
2026-10-07	Laya fine-tuned on 1,567 labelled decisions, onboarding harder set (rule table 0.565)	0.815	none	The open decision head	Executive piece
2026-10-07	Laya fine-tuned on 201 labelled decisions from 58 accounts-payable cases, ap-challenge, below the rule table	0.554	none	The open decision head	Executive piece
2026-10-07	AUC of a vote across several readings as a wrong-claim guard (not in use)	0.56	none	The open decision head	Executive piece
2026-10-07	AUC of a guard that read the cited document itself (not in use)	0.52	none	The open decision head	Executive piece
2026-10-01	Prompt tokens at which imbalance-first audit reaches 0.80 omission detection, eight requisitions	745	740 to 749	A Trial Balance for Agent Omissions	Paper
2026-10-01	Best omission detection by plain log review at any budget or trace length (0.80 never reached)	0.58	none	A Trial Balance for Agent Omissions	Paper
2026-10-01	Correct omission findings per thousand tokens read, imbalance-first over log review at its best budget, eight requisitions	about 21 times (1.34 against 0.063)	none	A Trial Balance for Agent Omissions	Paper
2026-10-01	Omission detection by plain log review reading the whole diary	0.25	0.19 to 0.31	A Trial Balance for Agent Omissions	Paper
2026-10-01	Omission detection reading the whole diary with a per-step checklist instruction	0.58	0.48 to 0.67	A Trial Balance for Agent Omissions	Paper
2026-10-01	Wrong amounts carried on both sides of the book found by the balance check	0 of 1,044	none	A Trial Balance for Agent Omissions	Paper
2026-10-01	Full-read log review: wrong amounts found minus claimed-but-skipped steps found (0.976 against 0.179)	0.798	0.738 to 0.857	A Trial Balance for Agent Omissions	Paper
2026-10-01	Omission detection, imbalance-first over a random sample of the same reading budget	+0.907	0.891 to 0.925	A Trial Balance for Agent Omissions	Paper
2026-09-17	Sprawl cost exponent in fleet size, monitoring budget held flat (cluster-robust)	1.4665	1.4101 to 1.5230	Sprawl Cost Under a Fixed Monitoring Budget	Paper
2026-09-17	Sprawl cost exponent in fleet size, per-agent monitoring with review that scales	0.9001	0.8664 to 0.9337	Sprawl Cost Under a Fixed Monitoring Budget	Paper
2026-09-17	Cost multiple per doubling of the fleet, flat monitoring budget against per-agent monitoring	2.76× against 1.87×	none	Sprawl Cost Under a Fixed Monitoring Budget	Executive piece
2026-09-17	Matched cases where changing review capacity under diluted monitoring left the simulation bit-identical	55 of 64	none	Sprawl Cost Under a Fixed Monitoring Budget	Paper
2026-09-17	Mean ticks (one tick is one week) from an agent going bad to detection, fixed inspection budget, 57 to 787 agents	2.78 to 22.49	none	Sprawl Cost Under a Fixed Monitoring Budget	Paper
2026-09-17	Exponent on the four smallest against the four largest portfolio sizes (registered ceiling test refuted: intervals overlap by 0.0145)	1.4949 against 1.3238	1.4037 to 1.5862; 1.2294 to 1.4182	Sprawl Cost Under a Fixed Monitoring Budget	Paper
2026-09-17	Banked runs the independent verification found to be duplicates	95 duplicate pairs (190 of 256 rows; 161 distinct simulations)	none	Sprawl Cost Under a Fixed Monitoring Budget	Paper
2026-09-10	Registered series showing the predicted zero crossing in the value of review	0 of 12	none	SIGIL: does a liability ledger change what an overseer authorizes?	Paper
2026-09-10	Value per tier-5 decision against no oversight, liability ledger against plain review (convex cost model)	−3.81 against −2.64	none	SIGIL: does a liability ledger change what an overseer authorizes?	Paper
2026-09-10	Tier-5 decisions wrongfully blocked, plain review against review with a liability ledger	30.1% against 42.7%	none	SIGIL: does a liability ledger change what an overseer authorizes?	Paper
2026-09-10	Share of worker errors that were wrongful approvals, the only kind a blocker can fix, tier 1 to tier 5	93% to 1.4%	none	SIGIL: does a liability ledger change what an overseer authorizes?	Paper
2026-09-10	Mismatches in the independent recompute of the scored episodes, convex and linear cost models (sum 29,902, derived)	0 of 14,950 convex and 0 of 14,952 linear	none	SIGIL: does a liability ledger change what an overseer authorizes?	Paper
2026-09-03	Change in quality margin per increment of delegation depth (Sobol band-only fit, n = 4,948)	−1.50	none	Delegation Cliff	Paper
2026-09-03	Change in quality margin per increment of reviewer capacity, the largest positive lever	+1.46	none	Delegation Cliff	Paper
2026-09-03	Model capability against agent self-check calibration (read as tied)	+0.95 against +0.92	none	Delegation Cliff	Paper
2026-09-03	Share of the design space in a broad transition band at a 0.95 quality floor (no cliff)	60%	none	Delegation Cliff	Paper
2026-09-03	Share of the design space that is fragile at a 0.99 quality floor (one 256-unit sweep)	79%	none	Delegation Cliff	Paper
2026-09-03	Revalidation certificate: 95th percentile difference on 200 rerun configurations, tolerance 0.05	0.0117	none	Delegation Cliff	Paper
2026-09-02	Error correlation (phi), identical weights on different silicon, contested items only	+0.447	none	Error Independence: does different hardware buy a second opinion?	Paper
2026-09-02	Mean error correlation (phi) across five cross-family pairs, contested items only (range −0.469 to +0.280)	−0.102	none	Error Independence: does different hardware buy a second opinion?	Paper
2026-09-02	Error correlation (phi) across all pairs before conditioning on item difficulty	0.556 to 0.827	none	Error Independence: does different hardware buy a second opinion?	Paper

Intervals are 95% unless the campaign page says otherwise; Trial Balance reports 90% cluster-bootstrap intervals. A number that a later correction changes keeps its row, and the correction is added beside it.