Build order step 1 satisfied — backend/scripts/check_credentials.py makes one real call per service and all three pass: Parallel Search (10 results, top hit dor.georgia.gov), gemini-2.5-pro via Vertex in us-central1, and a Maps distance call (LA -> Atlanta, 3,498 km / 31.6 h).

backend/app/seed_jurisdictions.py — 4 real jurisdictions (Georgia, New Mexico, Louisiana, Texas) hand-curated from those live Parallel searches, standing in for Layer 1 until Vertex AI auth is done. Served via GET /jurisdictions/seed, golden-file tested in backend/tests/test_seed_jurisdictions.py. Delete or fold into real Layer 1 output once extraction.agent works for these jurisdictions.

Gemini via Vertex AI — auth done. gcloud installed, ADC configured, billing active on zeta-structure-437412-v7, and aiplatform enabled; a live gemini-2.5-pro call in us-central1 returns. agent.py's extraction path is now unblocked but has not been run end to end.

Google Maps distance client (backend/app/maps_client.py) — verified live (LA -> Atlanta, 3,498 km / 31.6 h). Uses a dedicated API key restricted to the Distance Matrix API alone. The Routes API is not available on this project, so Distance Matrix is the only path.

Layer 1 end to end, live. POST /jurisdictions/search calls extract_jurisdiction_rule for real (was a hardcoded 501 stub) and GET /distance exposes maps_client.get_distance. Run against Oklahoma, Illinois, and New York — real base rates, real dated citations, correct is_discretionary/pool_status handling, all through verify_rule. Two live bugs this surfaced and fixed: - sources/tiers/uplifts came back from the function call as plain dicts, not SourceRef/Tier/Uplift — would have broken on first attribute access (verify_rule's assess_confidence reads s.retrieved/s.is_primary). Now converted properly in agent.py. - The qualifying sub-object had no properties/required in RECORD_JURISDICTION_RULE_SCHEMA, so Oklahoma came back with qualifying: {} — a silent $0 credit (calculator.py's q.get(key) treats every missing key as not-qualifying), not an error. Now all six keys are explicitly required in the schema. Frontend: Results.tsx has a "search another jurisdiction" box that calls this live and folds the result into the ranked comparison — same compute/distance pipeline as the seed jurisdictions, no separate code path.

Map tab (frontend/src/screens/MapView.tsx) — real geography: Natural Earth world countries (110m) + US Census states (10m), served from frontend/public/geo and fetched only when the tab opens (~220KB, kept out of the main bundle). d3-geo Mercator fitted to whichever jurisdictions are in the comparison. Vector, not tiles: no second billed API and no map key in the browser for a supporting view.

Breakeven line + sparkline under the hero card (frontend/src/lib/breakeven.ts, BreakevenLine in Results.tsx). Scans ATL spend via a new POST /compute/batch endpoint (same pure compute_benefit, batched to avoid ~50 individual HTTP calls per scan). Honest caveat: with the current 3 ranked seed jurisdictions (Georgia/New Mexico/Louisiana), all flat-rate with no caps, ranking never actually flips as spend scales, so it always correctly reports "leads across the whole range tested" — the crossover-found branch is implemented and reasoned through carefully but not exercised by real data yet. It'll show up once a tiered or capped jurisdiction is added to the ranked set.

Constraint-based greying — JurisdictionRule.constraint_gaps: dict[str, str], computed server-side by backend/app/constraints.py and applied uniformly to seed and searched jurisdictions alike (main.py's _verify_and_annotate). coastline is a fully enumerable state list; spring_only derives from the rule's own sunset_date/pool_status; large_soundstage is deliberately left unevaluated (no confident facilities data source — same "don't guess" policy as extraction). Frontend: toggle chips in Results.tsx grey out (never hide) jurisdictions that fail an enabled constraint, with the reason on hover.

Deployed, live, publicly reachable, no auth wall: - Frontend: https://zeta-structure-437412-v7.web.app (Firebase Hosting, on the zeta-structure-437412-v7 GCP project) - Backend: https://incentive-verifier-backend-559874048514.us-central1.run.app (Cloud Run, built from backend/Dockerfile via gcloud run deploy --source) - Both API keys live in Secret Manager (PARALLEL_API_KEY, GOOGLE_MAPS_API_KEY), mounted into Cloud Run via --set-secrets; the Cloud Run runtime service account has roles/secretmanager.secretAccessor and roles/aiplatform.user (the latter for Gemini via Vertex — verified with a live extraction in prod, e.g. Michigan correctly came back "No Active Program" rather than hallucinating one). - CORS (main.py) is narrowed to the real frontend origins (was *), overridable via FRONTEND_ORIGINS env var without a redeploy. - Redeploy: cd backend && gcloud run deploy incentive-verifier-backend --source . --region us-central1 ... (see Setup for the full secrets/env-var flags), and for the frontend, npm run build then publish dist/ to the zeta-structure-437412-v7 Hosting site via the Firebase Hosting REST API (versions -> populateFiles -> upload -> finalize -> release) — done this way instead of the firebase CLI because the CLI's own login wasn't set up for this account/project and creating a new deploy-only service-account key was (correctly) blocked by this environment's safety policy; the REST calls reuse the same gcloud auth print-access-token credential already established.

PDF export — the memo as a real PDF via the browser's own print-to-PDF (selectable text, live links, correct pagination, no PDF library). Collapsed panels are force-expanded before printing and reset on afterprint, so an export always carries every jurisdiction, the full assumptions table, the can't-verify list, all footnotes and retrieval dates, and the film office contact — regardless of what was collapsed on screen.

PDF upload (backend/app/extraction/budget_parser.py, POST /budget/parse) — Gemini reads an uploaded budget PDF with the same forced-function-call discipline as jurisdiction extraction: it quotes the figures printed on the topsheet and says which line each came from, and is never asked to total or reconcile anything. Lands on the form, pre-filled, each field annotated with its source line and anything missing raised as a warning — never straight to results.

Automated testing pipeline — .github/workflows/ci.yml runs backend pytest and frontend lint/test/build on every push and PR, with no credentials needed. backend/scripts/smoke_test.py drives the real deployed stack and asserts the invariants that must hold whatever Parallel/Gemini/Maps return (the full face-value-to-cash chain reconciling, components summing, credit monotonic in spend, canonical jurisdiction names, retrieved stamped today, coastal states not flagged landlocked); it's a manual/scheduled CI job since it spends real quota. 267 backend tests, 68 frontend.

`net == gross − relocation` was the invariant until a credit stopped
being priced as cash on wrap day; it's now
`gross + discount − audit + timing_loss − relocation`, and the smoke
test reconciles that chain against pre-timing and pre-monetisation
deployments too, so it stays meaningful across a version skew.

What a credit is actually worth — three corrections that each moved the answer by roughly the size of the entire relocation calculation. Payout mechanism (credit_type): a transferable credit is sold at a discount, a refundable one isn't, and treating them alike flattered transferable states. Fringes: 22-35% of wages, counted only where a statute says they qualify. And present value — a credit is a claim on future money, so it's discounted from wrap to payment, with a required audit charged before discounting. Timing defaults live in CreditTimingAssumptions, visible and editable, because payment timing is administrative practice rather than statute; a rule's own stated timeline overrides them.

Funding availability gates the ranking (_availability_block) — a closed pool, a passed sunset or a shut application window makes a program non-computable rather than merely lower-scoring. This is the product's stated wedge and it wasn't wired until now.

Adversarial verification (backend/app/extraction/challenge.py, POST /jurisdictions/challenge) — a second Parallel + Gemini pass whose only job is to disprove the first, searching for suspensions, exhausted pools and pending amendments. Reports contradictions, never overwrites a figure; code rather than the model decides whether a disagreement is material. Makes confidence: "conflicting" reachable from our own code for the first time.

Refusal on impossible inputs (_impossible_inputs) — found by firing adversarial payloads at a running server, which answered all of them with HTTP 200 and a confident number. base_rate: 5.0 returned a $9,520,000 credit on a $2M film. Checked, not clamped: clamping 30.0 to 1.0 would invent a 100% credit and rank it first.

Statutes hand-verified for Georgia, New Mexico and Louisiana against O.C.G.A. § 48-7-40.26, Rule 560-7-8-.45, NMSA 7-2F-15 and Louisiana Entertainment. Three errors in three jurisdictions, and the top recommendation changed. See test_seed_jurisdictions.py.

ADK agent (backend/app/agent/, scripts/run_agent.py) — an LlmAgent over six tools, satisfying BUILD_BRIEF §4's "the agent must invoke the calculator tool". Tools address jurisdictions by name, never by value, so the model never carries a figure between steps. Additive: the REST pipeline is untouched and is still what the frontend uses.

Custom domain (currently the default *.web.app / *.run.app URLs)

Statutes unverified beyond those three. Texas and every live-extracted jurisdiction are unchecked. Three of three checked had errors, so assume a similar rate rather than zero.

Per-category qualifying rates. qualifying is a boolean per spend category and can't express New Mexico's real rule (non-resident crew at 15% rather than 25%, capped at a share of the BTL budget). We exclude them, understating NM by $16,875 on a $2M drama rather than overstating by $67,500 — the smaller error, stated on the card.

Deployed and verified 6 Sep 2026. Both halves current; smoke test 6/6 against production. The adversarial pass and the ADK agent have now run live — each found a real bug on its first execution, both fixed. See VERIFY_LIVE.md for what was checked and found.