
LL144 requires an annual bias audit of automated hiring tools. It says nothing about the 364 days of decisions in between. Here is the record layer for those.
System notes
NYC Local Law 144 requires that an automated employment decision tool used to screen NYC candidates or employees have a bias audit by an independent auditor within one year before use, with a summary of results — selection rates and impact ratios by sex and race/ethnicity categories — published, candidate notice at least ten business days before use, enforcement by DCWP, and civil penalties of $500 for a first violation and $500 to $1,500 for each subsequent one, each day of use counting separately.
An LL144 bias audit is point-in-time and aggregate: it establishes group-level impact ratios for a past period and says nothing about any individual decision the tool makes between audits.
This system does not compute selection rates or impact ratios and is not an LL144 bias audit; it is a per-decision record layer that would let an auditor or respondent reconstruct any individual decision after the fact.
A governed screening decision pins the rule set to a content hash, requires each of three seats across two model families to derive its verdict clause by clause in machine-readable form, and seals only when a deterministic comparison finds the derivations identical.
A unanimous verdict is refused and escalated to a named human when the seats derived it differently, so agreement that hides divergent reasoning cannot authorise a candidate outcome.
Every governed finding must declare the records that were absent and the evidence that would flip the conclusion, and a panel facing a deliberately withheld record abstained and sealed the abstention rather than deciding.
The same machinery audits the rule set itself: a governed critique of a case file found eight defects, the lead one a necessity-stated-as-sufficiency error that had caused every prior derivation divergence.
Per-seat error rates are measured under a fixed rule set, and a 30-case calibration study on synthetic determinate fixtures sealed zero wrongful authorisations, with seat verdict accuracy of 30/30 and 29/30.
In 72 controlled calls, the auditable structure — declared absences, flip conditions, rejected alternatives — appeared in zero of 48 calls without the governing constitution and only under it.
No employment-domain calibration exists: the measured rates come from synthetic determinate fixtures in other task classes, no study covers resume or candidate data, and no impact-ratio computation is performed anywhere in the system.
Evidence ledger 10 · tier-ranked · API
5 more ranked claims
Ask this article · 8 suggested prompts
Text the build (+14245134626) or WhatsApp — slug|question creates a question node. Paste evidence with ingest slug|q:NODE_ID|your paste.