Receipts

Two receipts on this page. The Foundation v1 refactor moved the platform from a single-project consumer-Gmail-owned setup to a Workspace-rooted, multi-project, CMEK-encrypted, pause-by-default platform with per-domain Cloud Run services replacing the functions/main.py monolith. It substantially landed; what is and isn't live as of 2026-05-10 is below. The calibration pipeline ran from infrastructure-deploy through end-to-end trainer convergence on real PSL-derived data: 18 workflow smokes plus 5 manual trainer probes; 27 bugs across 8 categories; final probe converged the NumPyro hierarchical NUTS GRM at R-hat = 1.002, n_eff_min = 1,318, zero divergences in 9.0 seconds wallclock.

The full per-component receipts (the gated reports tier) carry the IAM bindings, the BigQuery DDL excerpts, the per-cluster controller logs, and the per-smoke ledger. This page is the public-facing summary.

Foundation v1 — what landed

The execution spec the refactor is built against is nexus_foundation_v1.md in the repo root (the F-3.x architectural commitments + S-4.x parallel security track). The receipts:

The headline: the foundation Stages 0–2 (org / projects / network / KMS / audit) are substantially live. Stage 3a (AlloyDB + cost-control plane) is live for prod and waits on the staging project for full parity. Stage 4 (compute) is live with 24 services deployed and per-route cutover in flight. Stages 5–7 (CDC / frontend / decommission) are partially live; the Firestore region cutover and Vertex Reasoning Engine redeploy are the remaining visible-to-the-user pieces.

Calibration pipeline — what converged

Probe 5 was the final shakedown probe. Sampler: NumPyro NUTS, four chains × 1,000 warmup × 1,000 samples, target_accept = 0.85, non-centered parameterisation on per-item log-discrimination. Inputs: ten personas drawn from the Phase-1 smoke library, six openness-family constructs scored from real PSL evaluator output (not synthetic responses). Outputs: six calibration_metrics rows plus three invariance rows MERGE'd into BigQuery; HDI95 credible intervals on every parameter; JSON-roundtrippable grm.json artifact written to GCS.

Probe 5 is the single most informative receipt in the shakedown. It demonstrates that the fully-implemented pipeline — workflow → shard worker → Vertex Batch → BQ MERGE → Custom Training → NumPyro NUTS → BQ MERGE — produces a converged psychometric calibration on data the model has never seen, in less time and less cost than any single shakedown smoke iteration.

[Samejima 1969] The graded-response model itself is Samejima 1969; convergence at this scale is the small-N receipt that the pipeline runs end-to-end. Production-scale Phase-2 fits will be larger but no harder.

Bug taxonomy

Eight categories absorbed all twenty-seven bugs. Every fix landed either in the production code path or in the eight-check local QA harness that gates Phase-2 readiness.

The largest cost was BigQuery schemas at ~$70. The iteration touched five sub-classes — STRING-vs-FLOAT64 autodetect drift, JSON dict-vs-string mismatch, JSON_EXTRACT ambiguity on JSON-typed columns, LIKE on JSON column, and namespace collision risk. All five are now caught statically in 60 seconds for $0 by test_bq_row_shape.py.

The second-largest was Vertex AI quirks at ~$50. The iteration discovered that gemini-3-flash-preview is /global/-only (smoke #8), that Vertex Batch enforces same-location for job + model (smoke #9), that the metadata field is rejected by the validator, and that Cloud Run Job auto-retries can spawn duplicate batches if upstream MERGE fails. All four are now baked into the worker config and workflow YAML.

What every smoke caught

The full per-smoke ledger lives at docs/shakedown_ledger.md and at the gated R3 infrastructure report. Compressed:

Phase-2 readiness signals

Five green, three open:

The two open items: the first resolves with a roles/billing.user grant plus a half-day drill; the second proceeds on its own external schedule. None blocks Phase-2 milestone-review approval.

What is not yet drilled

Honest framing of the four R5 robustness drills, per the robustness report:

• Auto-pause end-to-end — Cloud Billing pathway: ❌ NOT EXECUTED. Infrastructure built (functions/budget_check/main.py), gated on roles/billing.user.
• Mid-shard worker crash + resume — ❌ NOT EXECUTED. No CHAOS_FAULT branch in worker; max-retries-3 untested under deliberate failure.
• Vertex 429 backoff under live quota pressure — ◐ PARTIAL. _RateLimiter fully built; shakedown never drove quota saturation.
• Idempotency stress — ◐ PARTIAL. Smokes #11 and #16 incidentally exercised the BQ MERGE dedupe path; no deliberate 6-parallel-shards stress drill.

All four are on the Phase-2 hardening backlog. Two are blocked on external dependencies (roles/billing.user for #1; chaos-image build for #2); two can be drilled tomorrow if needed (#3, #4). Phase-2 corpus generation may incidentally exercise #3 even without the deliberate drill.

Foundation-side gaps from the table at the top of this page — Firestore region cutover, Vertex Reasoning Engine redeploy in eu-west4, Knowledge Catalog tagging baseline, the three reserved workflows, the staging project + cluster, Assured Workloads enrolment, the Firebase Stream-Firestore-to-BQ extension — are scheduled as the foundation refactor's later stages land. None blocks Phase-2 milestone-review approval; all are in the public roadmap at /about/science/roadmap.

What ships next

1. Grant roles/billing.user to an operator account → configure Cloud Billing budget → run the auto-pause drill.
2. Run idempotency stress drill (6 concurrent re-fires of the same (iteration, shard_index)).
3. Run mid-shard crash drill once a CHAOS_FAULT-flagged worker image is built.
4. Resolve the three BigQuery write-path divergences flagged in R7 (legacy tabledata.insertAll calls in pre-shakedown scaffolding).
5. Cut iteration 10 — first Phase-2 production iteration; full corpus (2,000 personas × 50 sessions × 12 turns); all ten calibration families parallel.
6. Provision neumatics-staging once the billing-account project quota raises; bring up the staging AlloyDB cluster + perimeter.
7. Redeploy the three Vertex Reasoning Engines in eu-west4 inside neumatics-prod; flip the apphosting.yaml resource IDs.
8. Install the Firebase Stream-Firestore-to-BQ extension; cut over from the legacy nightly export.

The pipeline is shaken down. The foundation is substantially landed. The receipts are filed. Phase-2 is unblocked on roles/billing.user plus milestone-review approval.