Lookback
Repo
Judge modeEach section shows the criterion it speaks to and a link to the evidence.Full evidence map

For judges

If you have three minutes: watch the video (2:50), then open work item #2. Its thread is the whole product: the appeal, the agent's triage, the merge request, the pipeline, and the restoration summary.

Path A, Supervised. A person approves the outcome, not the steps: one merge click after every gate has passed (agents have no merge tool on this platform). Everything before and after it runs on its own, except anything adverse to a household, which always waits for a caseworker by design. Special prize target: Most Stages Covered.

Criteria and evidence

CriterionOur claimEvidence
Technological implementationNine lifecycle stages with real work in each; two custom Duo flows (five agents) published to the AI Catalog and started by triggers; a chat agent and an MCP server (six read-only tools, 2025-11-25 transport, 12 conformance tests plus a 14-check live probe, exercised end to end by the GitLab Duo CLI); deterministic replay over signed container images and git history; hash-chained ledgers; replayability and replay gates; keyless Cloud Run deploys; 82 tests.
DesignOne thread from a hearing officer's comment to money restored, every artifact linked in order on the work item. Unhappy paths handled: not a software error, test does not reproduce, adverse findings, 12-month limit, tampered ledger.
Potential impactA legal duty that already exists (7 CFR 273.17) made executable. Real audience: state SNAP agencies and vendors, legal aid, any team deciding money about people. Measured on 42,327 real USDA audit households; in the live run $25,861 to 413 households from one appeal.
InnovationAn appeal becomes a collective remedy; remediation becomes a CI/CD stage; asymmetric autonomy (acts alone only in the person's favor) as a design principle. Prior art stated plainly: bank remediation tools and payroll retro exist but are not wired to the code fix.
PresentationThe video shows the real run end to end, with the refusal moment (overpayments held for a human) and the 12-month limit applied. This site mirrors it, with Simple and Technical modes and an explanation on every number.

Requirements

Things we would rather you hear from us

  • The hearing decision in the live run is constructed and labeled. The households are real (USDA FY2024 QC public use file). Numbers that use constructed timelines are reported separately from numbers that use real sampled months.
  • The demo ledger's months are shifted forward 21 months so they fall inside the 12-month restoration window; the shift is recorded in each record. The first dry run, without the shift, restored $0, because the law says so.
  • Replay recall and precision are 1.0 by construction: replay re-runs the same decisions exactly. The claim is the wiring, the gates and the guarantees, not detection cleverness.
  • The merge-request replay gate did not catch MR !3 (a household exactly at the income limit) because no real household in the window sits on the limit. A unit test caught it. Both layers are needed, and the MR says so.
  • Posting the final summary on the work item needs a token with API scope that group policy would not let the pipeline create; the release is published by the pipeline and the note was posted from a local token. Documented in SPIKES.md.
  • Federal SNAP rules only. State options are out of scope and listed.