What high-reliability acceptance gates teach AI-agent acceptance contracts
The question
How do high-reliability fields (aviation certification, nuclear, clinical trials, shipbuilding sea-trials) structure staged acceptance gates, and what transfers to AI-agent acceptance contracts? Extends the commissioning brief + acceptance-contract verify-skill hardening toward the gates end of the spectrum, with the CAF restructure (106 skills → ~15-20 objective+acceptance-test skills, HITL gates, manifest as state owner) as the live application.
What we already know (from the vault)
- [[2026-06-09-commissioning-as-agent-spec-craft]] already maps the spectrum: pure outcome-commission + trust (prizes) vs tight spec + staged acceptance gates (warships, Apollo). It places naval commissioning at the disciplined extreme — keel laying → launch → builder's trials → INSURV acceptance trials by a board outside the builder → delivery and commissioning as two deliberately separate accountable events. Mark 7: scale verification to failure cost.
- [[2026-06-09-acceptance-contract-template]] codifies the contract block: pre-registered DONE (rule 1: "a contract authored after seeing the output rationalizes"), REAL-GOAL GUARD, STATE OWNER, and the failure-cost→gates ladder (reversible-cheap → 1 verifier; irreversible/expensive → majority-of-3 with distinct lenses).
- [[2026-06-09-caf-restructure-proposal]] is the live application: each collapsed objective skill carries OBJECTIVE + DONE test + load-bearing constraints;
manifest.yamlis the state owner holdingcontract_validationper phase andhitl_approvals(gate, approved_by, timestamp); orchestrators HALT at HITL gates; Meta-Council VETO/HALT stays un-collapsed per the DO-NOT-COLLAPSE list. - [[2026-05-20-verify-stack-two-gate-pass-fail-architecture]] deliberately replaced the 3-state PASS/ITERATE/SCRAP with two-gate pass/fail, keeping the iterate-vs-scrap call inside the fresh-eyes critic.
- [[2026-05-19-verification-as-independent-worker-pattern]] sets the structural rule: a producer never verifies its own work; verification routes through a skeptical fresh-eyes worker.
What the web says
- FAA type certification is a staged-gate chain with the criteria frozen first. The process runs certification basis (the exact rule set the design must meet, established and effectively frozen at application) → compliance planning → applicant analysis and testing → Type Inspection Authorization (official FAA flight tests begin only after the TIA issues, and conformity inspections first verify the test article actually matches the design) → Type Certificate (design approval) → Production Certificate (a separate approval that the organization can consistently reproduce the approved design) → per-aircraft airworthiness certificates. Design, reproduction, and instance are three distinct gates. (faa.gov "How it works: Aircraft Certification"; FAA Order 8110.4C; Wikipedia "Type certificate")
- NRC Part 52 ITAAC is the cleanest per-gate template: a three-part format of design commitment / verification method / acceptance criteria. The licensee performs all inspections, tests, and analyses and submits an ITAAC Closure Notification for each (10 CFR 52.99(c)(1)); NRC inspectors independently verify per Inspection Manual Chapter 2503; and before operation the Commission must make the aggregate 52.103(g) finding that all acceptance criteria are met. Performer, verifier, and authorizer are three different parties, and closure is a recorded artifact per gate. (nrc.gov ITAAC page; 10 CFR 52.99)
- Clinical trials gate by graduated exposure: IND clearance before first human dose, Phase 1→2→3 each enlarging the exposed population only after evidence at smaller scale, pre-registered endpoints and stopping rules, an independent DSMB with authority to halt mid-trial, an end-of-Phase-2 FDA meeting that pre-agrees the pivotal trial's acceptance test before it runs, NDA/BLA review, then Phase 4 post-market surveillance — acceptance doesn't end at approval. (Established FDA IND/NDA structure; stated from background, consistent with vault canon.)
- Navy sea trials sequence a cheap self-check before the expensive independent check: builder's trials (the producer's own shakedown) precede acceptance trials run by INSURV, a statutory board independent of both the shipbuilder and the acquisition program office. Delivery (custody transfer) and commissioning (authorization to operate) remain separate events. (Confirms the vault's commissioning-brief account.)
- Published agent-side work is converging on the same shape. arXiv 2603.15676 ("Automated Self-Testing as a Quality Gate") gates LLM-app releases on five dimensions (task success ≥80%, context preservation ≥90%, P95 latency <15s, safety ≥95%, evidence coverage ≥80%) with a PROMOTE/HOLD/ROLLBACK state machine — ROLLBACK below 70% of target, HOLD for marginal failures routed to manual triage — and grows its question bank from every HOLD/ROLLBACK to prevent teach-to-the-test. Over 38 runs it caught 2 ROLLBACK-grade builds. (arxiv.org/html/2603.15676v1)
- Practitioner consensus for 2026 agent deployment is golden datasets drawn from real failures, judges calibrated against human reviewers, and CI gates that block regressions — quality gates as release infrastructure, not one-off evals. (digitalapplied.com, adaline.ai eval-pipeline guides; arXiv AURA on autonomy risk assessment)
Convergences and contradictions
- Strong convergence on pre-registration: the FAA freezes the certification basis at application; ITAAC are written into the license before construction; FDA pre-agrees pivotal endpoints. This is exactly acceptance-contract rule 1 ("pre-registered, not post-hoc") — the web confirms it is THE load-bearing move in every regime, not an RDCO idiosyncrasy.
- One gap the web exposes in our template: ITAAC's three-part format declares a verification METHOD per criterion (inspection vs test vs analysis), not just the criterion. Our contract block has DONE + STATE OWNER but no per-criterion "how will this be checked" field — the verifier currently improvises the method.
- Mild contradiction on verdict states: the arXiv gate uses three states (PROMOTE/HOLD/ROLLBACK) where [[2026-05-20-verify-stack-two-gate-pass-fail-architecture]] deliberately collapsed to two. Reconciliation: their HOLD routes to a human, which in our architecture is the HITL escalation path, not a verifier verdict — but it suggests FAIL should distinguish "iterate autonomously" from "park for human triage." A second tension: aviation/nuclear gate the process heavily (conformity inspections, QA systems) while the vault canon says grade the end-state, not the path. Resolution: process gates exist where the end-state can't be fully tested; where DONE is fully observable, end-state grading suffices.
Synthesis for RDCO
Staged gates are per-stage closure events plus one aggregate finding — not a thicker single DONE test. All four regimes decompose acceptance into many small, individually-closable gates (an ITAAC, a TIA, a phase readout, a builder's trial card) and then require a separate aggregate finding before authority transfers (52.103(g), type certificate, NDA approval, commissioning). For CAF this validates the manifest design and sharpens it: each phase's contract_validation entry is an ITAAC-style closure, and the build-manifest handoff at Phase 7 should be an explicit 52.103(g)-analog — a recorded finding that every gate cleared, made by the account lead, not implied by the last phase completing. The concrete upgrade is to adopt ITAAC's three-part format per contract item: design commitment (what the phase promised) / verification method (how it gets checked — re-run, inspect artifact, cross-reference) / acceptance criterion (the threshold). Today's completeness-score-only validation is one-third of that structure.
Hold-point placement follows error-propagation cost, not uniform spacing. Clinical phases escalate exposed population only after evidence at smaller scale; the FAA requires conformity inspection before official flight test because testing the wrong article voids everything downstream. The CAF HITL gates land correctly by this logic: Phase-1 framing sign-off (cheapest point to fix the most expensive class of error — the Sydney lesson) and Phase-3 classification sign-off (the routing decision every downstream lane inherits, structurally a certification-basis freeze). The rule to write down: place a HITL gate wherever the next phase multiplies the blast radius of an undetected error or makes it irreversible; everywhere else, an autonomous verifier suffices. The same logic gives a graduated-exposure ladder for new agent capabilities generally — pilot one collapsed phase skill (Phase B) before the full 106→15 collapse is the clinical-phase move, and running one client before /run-batch over a roster is the type-certificate-before-production-certificate move: "it worked once" and "it reproduces reliably" are different gates.
Who signs off: three roles, never collapsed into fewer than two parties. Every regime splits performer (licensee, builder, sponsor, applicant) from verifier (NRC inspector, INSURV, DSMB, FAA flight test) from authorizer (the Commission's finding, commissioning authority, FDA approval). INSURV's lesson is the sharpest: the verifier sits outside the producing chain entirely, reporting around the program office. For CAF: phase skills perform; Meta-Council + fresh-eyes verification check (the DO-NOT-COLLAPSE list is exactly right — collapsing m-1/m-2/m-3 into the phases they police would be putting INSURV on the shipbuilder's payroll); the human at HITL gates authorizes. And every regime makes closure an evidence artifact — the ICN, the TIA, the trial master file. The manifest's hitl_approvals records who/when; it should also record what evidence the approver saw (paths into contracts/ and assets/), so a resumed session or auditor can reconstruct the basis for each cleared gate. One cheap sequencing trick transfers directly: builder's trials before acceptance trials means the worker runs its own DONE self-check before the fresh-eyes verifier is spawned — same independence guarantee, fewer wasted verifier cycles.
For the acceptance-contract template itself, two amendments earn their keep. First, add a per-criterion verification-method field (the ITAAC third part) so the contract names how each DONE item gets checked, not just what it is. Second, borrow the arXiv paper's anti-Goodhart mechanism for any standing gate (verify-pdf-output's 12 checks, a future CAF regression suite): every HOLD/ROLLBACK-class failure feeds a new scenario into the test bank, so the gate hardens with use instead of being memorized. The failure-cost→gate-count ladder in the template is confirmed by every regime studied — these fields are all at the catastrophic end and run maximum gates; the transfer is not "add more gates everywhere" but "make each gate a named, recorded, independently-verified closure event."
Open follow-ups
- Should the acceptance-contract template add an ITAAC-style per-criterion verification-method field (inspect / re-run / cross-reference), and what does that cost in dispatch overhead?
- Formalize a graduated-exposure ladder for agent autonomy (paper-trade → bounded-real → full) as written policy, mirroring clinical phase escalation?
- Should FAIL in the two-gate architecture split into "iterate autonomously" vs "park for human triage" (the arXiv HOLD state), and does that re-introduce the 3-state problem the 2026-05-20 SOP killed?
- Question-bank regression gate for skills: should /improve edits be blocked by a growing eval set seeded from past verify-* failures (arXiv 2603.15676's mechanism)?
- Who verifies the verifiers — is there an INSURV-style periodic independent audit of the verify-* family itself?
- What is Phase-4-style post-market surveillance for shipped agent skills — continuous in-production monitoring vs gate-at-release only?
Related
- [[2026-06-09-commissioning-as-agent-spec-craft]] — the spectrum this brief deepens at the gates end
- [[2026-06-09-acceptance-contract-template]] — the contract block the ITAAC three-part format would amend
- [[2026-06-09-caf-restructure-proposal]] — live application: manifest as state owner, HITL gates, Meta-Council independence
- [[2026-05-20-verify-stack-two-gate-pass-fail-architecture]] — the two-gate verdict design the HOLD-state question tests
- [[2026-05-19-verification-as-independent-worker-pattern]] — the performer/verifier split all four regimes confirm
Sources
- Vault: 06-reference/concepts/2026-06-09-commissioning-as-agent-spec-craft.md
- Vault: 02-sops/2026-06-09-acceptance-contract-template.md
- Vault: 01-projects/phdata/2026-06-09-caf-restructure-proposal.md
- Vault: 02-sops/2026-05-20-verify-stack-two-gate-pass-fail-architecture.md
- Vault: 02-sops/2026-05-19-verification-as-independent-worker-pattern.md
- https://www.faa.gov/newsroom/how-it-works-aircraft-certification — FAA type certification stages (via search; direct fetch 403'd)
- https://en.wikipedia.org/wiki/Type_certificate — staged TC/PC/airworthiness structure
- https://www.nrc.gov/reactors/new-reactors/how-we-regulate/oversight/itaac — ITAAC three-part format, ICN, 52.103(g) (via search; direct fetch failed)
- https://www.ecfr.gov/current/title-10/chapter-I/part-52/subpart-C/section-52.99 — ITAAC closure mechanics (via search)
- https://arxiv.org/html/2603.15676v1 — "Automated Self-Testing as a Quality Gate" (PROMOTE/HOLD/ROLLBACK, five dimensions, growing question bank)
- https://www.digitalapplied.com/blog/ai-agent-evaluation-pipeline-2026-testing-methodology and https://www.adaline.ai/blog/complete-guide-llm-ai-agent-evaluation-2026 — practitioner eval-gate consensus (via search)