The Taskmaster · Deployed on Google Cloud
Sixty Days reads a synthetic disaster-assistance decision letter, preserves the exact stated reason, organizes supporting evidence, prepares applicant-controlled requests, tracks the deadline, and builds a reviewable draft packet. It never sends or submits.
Public proof: the shared-substrate exit test passes 10/10, separately from the product-specific 24-step flow. Open the live product or inspect the public repository.
Start the guided preset, inspect its eight registered wakes, compare the close and wide evidence fixtures, then advance time and watch the scheduled safeguard build a visibly partial packet automatically. Open the conformance map for the official-source-to-code-to-test chain.
Evaluation map
Claims below point to executable behavior, measured recordings, or an explicit boundary.
| Criterion | What is distinctive | Evidence |
|---|---|---|
| Innovation | A standing 60-day watch converts one letter into an eight-wake plan and evidence routes. When the packet safeguard becomes due, it executes PacketBuilder against current accepted evidence and stores a partial-safe snapshot without sending it. | Run the four-stage console; inspect sixty_days/wake_actions.py, deadline.py, and packet.py. |
| Taskmaster fit | The workflow completes connected work: understand, gather, track, and review, while preserving applicant control. | The public acceptance script verifies 23 linked behaviors. |
| Architecture | Gemini readers and evidence checks sit behind typed routes, deterministic clocks, durable Firestore state, redaction, and independent verification. | Architecture below, conformance report, and standalone test suite. |
| Production readiness | Deployed service, bounded autoscaling, health endpoint, structured observability, measured fixtures, accessibility gate, and replayable demo. | Live URL, Cloud Run facts, recording reports, and reproduction commands. |
Mandatory technology
Vertex AI powers decision-letter transcription and observable evidence-photo checks. Four letter calls and two evidence calls are recorded and graded against adjacent truth.
One of the four accepted agent frameworks (ADK, GenAI SDK, Antigravity SDK, GenKit). The reader, evidence checker, and recording scripts call it with schema-constrained structured outputs at temperature zero. Deterministic fixture replay keeps public judging repeatable.
Cloud Run serves the application and scales to zero; Firestore holds structured cases, wakes, and durable counters; a dedicated sixty-days-wake-scan Cloud Scheduler job drives this service’s own worker over OIDC; Secret Manager holds key hashes; Cloud Trace and Cloud Logging receive OpenTelemetry spans; Cloud Build and Artifact Registry produce the image. Live scheduler proof.
Model Armor: the de-identified API screens untrusted letter text for prompt injection before any model call, in front of the deterministic redaction and quarantine rather than instead of them. Measured live it blocked a bare injection and missed the same injection diluted inside a long letter, which the quarantine caught. An unavailable screen is recorded as unavailable, never as a pass.
Prove the model is live: the console replays recordings so judging is repeatable, and also carries one button that calls Gemini 3.5 Flash on Vertex AI for real, on a letter image with no transcript supplied, graded against the same truth file by the same fields as the published run. Bounded per visitor and per day.
Three additional Google models: Gemma 4 reviews already-pattern-redacted text for remaining person names; Gemini 3.1 Flash Image creates the non-data-bearing first-use illustration; and Veo 3.1 Fast creates the optional four-second motion briefing. Media prompts, model IDs, byte counts, and hashes are public in the provenance manifest. Generated media never becomes case evidence, enters a packet, or predicts an outcome.
Deployment proof
This service now owns its scheduler job outright. It claims only the wakes it can execute, so a worker on the shared Firestore substrate cannot retire this project's safeguards.
sixty-days in us-central1, bounded to three instances and running as sa-reason.sixty-days-wake-scan posts to /internal/scan-due every minute with an OIDC token scoped to this service. The worker claims only wakes owned by sixty-days and sixty-days-beta, so a neighbouring worker on the shared Firestore substrate cannot retire this project's safeguards. Anonymous calls to the worker route return 401.scripts/sixty_days_demo_flow.py performs 23 public HTTP checks; /exit-test exercises the shared 10-stage substrate.Verify it yourself
Every button below runs against the deployed service, right now, in this browser. Nothing is pre-recorded and no terminal is required.
Creates a run, registers a reminder sixty days out, proves nothing fires early, advances a simulated clock, and checks that the agent wakes, records its action, and completes. Then it tries to break it: hidden instructions inside a document, a fabricated number, a worker that dies mid-run.
Runs the appeal workflow against the public API from this browser and checks what the README claims: reasons quoted from the letter, eight reminders registered, a bad photo refused, a request prepared but never sent, a draft that still lists its gaps, and no send or submission route anywhere.
The demo clock is simulated and says so. This is the wall clock: a Cloud Scheduler job calls this service every minute whether or not anyone has the page open.
Each published rule, the source it came from, the module that implements it, and the test that holds it. Read as data rather than as prose.
cd app
python -m venv .venv
.venv/bin/pip install -e ".[dev]"
python -m pytest -q
python scripts/check_a11y.py
python scripts/sixty_days_demo_flow.py --url https://SERVICE_URL
python scripts/verify_rules_compliance.py --url https://SERVICE_URLThe expected standalone baseline is 365 tests. Deployment and credentials are
documented in the repository README and deploy.sh.
Findings and learnings
Keeping the exact decision reason beside every action prevents a polished packet from drifting away from what the letter actually said.
An audit found all-caps applicant names and FEMA registration references missing from an older pattern set. Regression tests now enforce PERSON and CASE_REF coverage.
A packet may be generated with missing items visibly listed. The verifier refuses to turn absence into a confident claim.
A parallel rehearsal exposed a shared simulation clock: one submission could move the other's timeline. Demo clocks are now namespaced, simulated wake claims check run ownership, and the production wall-clock worker remains shared.
Research-to-design trace
Official examples are treated as routing options, never as a promise that one document will change an outcome.
| Official finding | Implemented decision | Boundary |
|---|---|---|
| The decision letter identifies the documents relevant to that determination. | Every requirement begins with an exact quoted reason; unknown reasons stop for caseworker review. | The catalogue cannot override the applicant's letter. |
| FEMA lists several ownership and occupancy records, including alternatives when public-record matching fails. | The planner names multiple accepted examples and their date context instead of prescribing one rigid document. | No listed record is called sufficient or outcome-determinative. |
| FEMA's insurance guide distinguishes settlements, denials, and policy evidence of exclusions or missing coverage. | The insurer route requests those records and no longer treats an unsupported signed statement as equivalent. | The applicant sends the request; no insurer is contacted by the service. |
| A September 2025 FEMA IHP explainer tells applicants to put application and disaster numbers on every submitted page. | The PDF repeats synthetic application and disaster references in every page footer and asks the applicant to verify them. | The PDF remains optional, unofficial, and not submitted. |
FEMA ownership and occupancy fact sheet · FEMA insurance quick reference · FEMA IHP application, eligibility, registration, and appeals explainer
Limits
Policy basis: FEMA appeal quick reference and GAO-20-503.
No practitioner or applicant study is claimed. Public research establishes the workflow context; executable tests, recorded fixtures, adversarial probes, live acceptance, and explicit limitations establish what this build actually does. Legal accuracy, eligibility, and applicant usability remain unproven.