Changelog — every framework revision, dated

v4.2 — 16 September 2026

Trigger: Defect correction and published provenance

Fixed the Scenario B spot-check cost. It was multiplied by pre-automation headcount, which has nothing to do with how much AI output needs checking: an identical AI-led scenario producing the same 1,312 units was billed $150 at a three-person firm and $3,000 at a sixty-person one. It also double-charged, because the oversight premium was already applied. Scenario B now pays only the shortfall between the review hours its own output demands and the hours its oversight premium already funds. Scenario B costs fall slightly for most tasks; no readiness score or verdict changes, because the score never used Scenario B.
Removed the AUTO_OVERSIGHT constant, now unused.
Published a provenance table stating, per constant, whether it is observed, derived, conventional or hand-calibrated judgment. Most are judgment.
Published a sensitivity analysis showing which constants actually move verdicts. Risk weighting dominates at normal volume; the leverage ceiling dominates at low volume.
Corrected the v4.1 entry below, which overstated the effect of the shape-to-economics re-weighting.

Trigger: GPT-6 Astra (3 September 2026)

Economics now carries 45% of the readiness score. Task shape alone caps a score at 55% of its potential, so low-volume workflows no longer score as though they were running at scale. Corrected in v4.2: this entry overstated the change. Sensitivity testing published with v4.2 shows the re-weighting moves no verdicts at mid volume and one of 33 at low volume. The direction is right; the magnitude claimed here was not measured.
Added the oversight self-check: the framework compares the supervision hours its own sampling schedule demands against the oversight premium the cost model charges, and flags the saving as optimistic when the first exceeds the second.
Added the sample-adequacy rule. Where a sampling rate cannot produce a sample large enough to detect the stated error ceiling, the result says so.
Added an oversight log export, so the sampling schedule and delegation level can be filed as a record rather than screenshotted.
Re-scored all 19 capability-gated tasks against Astra. Extraction under unusual formatting and classification on long-tail categories improved; sampling rates at steady state fell slightly for those tasks. No verdict changed band.
All 14 liability-gated tasks held. No change.

Trigger: Framework restructure

Introduced the oversight contract: sampling rate, ramp schedule, error ceiling and delegation level derived from risk, judgment and repeatability.
Introduced the liability-gated and capability-gated distinction, and marked every task as one or the other.
Output made a function of hours and staffing, so cost per unit is invariant to hours and cutting staff can no longer be shown as free.

Trigger: Rate refresh

Updated published list rates for all priced models.
Added the model-spread disclosure to the results page, showing the gap between the cheapest and dearest model as a share of the monthly figure.