Israel Lee Armstead

AI agents can work for hours now. Nobody grades the work — and when an agent claims success for work that did not happen, nothing in the stack objects. The world it is built for is a working one: Chancellor already serves as the governance office of the entrant's ongoing multi-agent program — the discipline that program's own agents worked under while this entry was built — and its work continues after the hackathon. Chancellor is the governance-and-evaluation harness for long-running agents. Work runs as an explicit, durable task graph that persists to disk and resumes after a crash. Every consequential output is graded against a weighted rubric by an independent judge model — the worker never scores its own output. Chancellor's own instruments fail closed: if a check errors, the run halts; a broken gauge is a stop signal, never a pass. Authority escalates where capability doesn't: spend, external sends, publishing, irreversible acts, and credential use stop at a human gate, and every gate decision leaves an append-only, hash-chained receipt that anyone can verify offline. The shipped product (v1.2.2) also carries the surface the measurements pointed at. A built-in supervisor — `chancellor supervise`, the KEEP office — watches a governed agent's signed checkpoints: it tells dead from stalled, verifies the chain before any resume (a corrupted chain is refused, never restarted), and receipts every interruption with its cause graded proven, plausible, or unknown — never invented. That behavior is fault-injection verified: an agent killed mid-work is detected, verified, and resumed to completion, in 15 of 15 consecutive test pairs plus the pair run against the shipped package. The supervisor's measured trial (RUN-9) found K1 at 22/40 and K6 at 215/240 on watch 9104 against their pre-registered thresholds; the continuity claims are stated with that exception. The supervisor anchors on one machine; machine death is outside its watch, and RUN-9's termination cell demonstrated that boundary rather than claiming past it. Receipts now name the governing frame they were decided under, gate refusals leave receipts carrying named refusal states, checkpoints commit to the log's entry count as well as its root, and test runs are recorded in the same chain. The adversarial Suite C, the gate cap object, and production signer separation ship as design documents only, labeled as such in the repository. The shipped product also governs capability install: an agent gaining a new capability is a gated act class under the frame — proposed, checked before it moves, decided at the human gate, receipted, verified live by a reader who did not install it, watched through a probationary window, and reversible under receipt — proven in a pre-registered install hearing on file in the repository (docs/evidence/foundry-install-hearing/), every receipt verifiable offline. The grades stand as the records print them: the registry bindings are test-grade until the registrar custody act, and the decision chain is consistency-grade under the development signer — labeled on their faces, and labeled here. The install hearing's full record — run card, both chains' exports, verifier reports, and scored results — is filed at docs/evidence/foundry-install-hearing/ in the repository. Measured on the v1.1.0 tree; the install hearing ran on v1.2.0. The evidence is measured, and the bounds travel with the numbers. The Suite B reliability frontier was measured twice on the same tree — first at 5,600 runs (Oct 31), then re-measured at 10x: 56,000 runs on ten machines (Nov 1–2), every figure recounted from the raw outputs by an independent parser that imports no product code. Governed runs drifted 0 in 18,000 against 3,000 in 18,000 ungoverned on identical fault manifests — the zero carries a 95% upper bound of 0.0166% (n = 18,000), with manifest-level bounds of 0.99% (300 instances) and 1.10% (270 distinct) printed beside it in the record. A value-cell extension added 0 drift in all 200 cells — Part B is governed-only by design; its zero leans on Part A's paired governed/ungoverned control, and is reported that way. In forty endurance soaks — 19,200 segments, 1,920 injected faults — faults halted runs fail-closed and corrupted nothing: 18,403 audit-chain entries, 0 breaks, 2,320 signed checkpoints, all verified. The earlier 1x record (0 in 1,800 governed, bound 0.166% at the cell level and the weaker manifest-level bound of 25.9% (0 of 10 manifests), four 8-hour soaks) stands in the repository as the first measurement; the story is the same zero, re-measured at ten times the volume. One honest negative, printed on purpose: our Suite A judge-calibration gate has not passed — the calibration of record and one pre-registered remedy both failed their gates (the remedy on a single reference lost to transport errors, with all scored diagnostics above threshold). So Chancellor's judge verdicts ship labeled uncalibrated, and no Suite A score is claimed anywhere in this submission. The harness grades its own judge the way it grades everything else — and publishes the result either way. Project page (live): https://chancellor.iamgodiam.net