File size: 10,081 Bytes
2b86c60 27e707f 64b0986 7ff807e | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 | # HyperFlow status and claim register
Snapshot: 22 October 2026. This register distinguishes completed engineering checks from release gates. No production readiness or live-provider certification is claimed.
## Current phase and gates
| Phase | Completed evidence | Remaining gate |
| --- | --- | --- |
| 0 — truth | Original BRAIN archived; canonical documentation created; mounted routes and synthetic metrics corrected | Independent fresh-checkout setup remains unverified |
| 1 — identity | Explicit live authentication failure; persistent credential encryption; account isolation; strict OAuth state/redirect/rotation; read-only verification endpoint; negative-path tests | Authenticated provider discovery and actual OAuth integration require local credentials/configuration |
| 2 — BasketOps | [500-case solver oracle](solver-substitution-results.json), version-bound approval and reconciliation, [60 process interruptions](crash-matrix-v2-results.json) | Live cart-read evidence and supervised cart integration remain unverified; no upstream exactly-once guarantee |
| 3 — harness | Durable turn records, physical-request budgets, read-only tool allowlist, retry conflict handling; [full-intent baseline](intent-substitution-results.json) | Rules achieve 65/80; reserved 20 cases and model comparison are NOT_RUN; quality gate remains open |
| 4 — ML | [Retail reproduction](retail-v1-results.json), [controlled censoring](censoring-v1-results.json), explicit promotion gates and rollback tests | Exposed evaluation cannot establish unseen generalization; production quality thresholds remain unconfigured |
| 5 — FDE | Real dependency diagnostics, persisted incident traces, [PostgreSQL backup/restore](backup-restore-results.json) | Human pilot, unaided deployment, off-host restore and application rollback drill remain open |
| 6 — release | Mounted workflows restored; frontend build and isolated browser checks recorded below | Independent accessibility/mobile review, Docker deployment and final release review remain open (Compose configuration validated; daemon unavailable) |
## Current evidence and exact scope
- Solver: 500/500 seeded bounded baskets agree with exhaustive feasibility, minimum cost and normalized excess. Synthetic catalogue inputs; not provider catalogue certification.
- Crash matrix: 60/60 abrupt subprocess exits across six persisted boundaries, ten trials each, using real PostgreSQL and an independent simulated receipt ledger. Zero duplicate simulated receipts. UNKNOWN is retained when execution cannot be established; no blind re-placement. The added sixth boundary exits after receipt validation but before local success persistence; the original 50-trial report is retained separately.
- Intent parsing: 65/80 exact completions (50/60 development, 15/20 validation), zero accepted constraint violations in those cases. Initial failures remain in [initial report](intent-development-initial-results.json). Labels are synthetic author labels; validation was inspected during repair. Reserved test and LLM comparison were not run. Brand requests and ambiguous substitution scope still require clarification. Standalone substitution directives work for a single explicit product.
- Load: [complete socket report](socket-load-complete-results.json). Three five-minute runs at five scheduled HTTP requests/second completed 4,500 requests (3,600 reads, 900 simulated action executions), zero failures. Read scheduled-completion p95 by run: [26.56, 34.92, 41.18] ms; action p95: [66.73, 65.34, 68.81] ms, including preparation. All three runs passed the frozen 250 ms read / 1,000 ms action thresholds. Apple M2, 8 GiB, Python 3.11, same-host PostgreSQL/API/client, no TLS/provider/model latency. Other checks ran concurrently during the first trial. Fixed offered load does not establish maximum capacity. The [earlier interrupted campaign](socket-load-partial-results.json) remains retained as historical evidence.
- ML: current UCI study reproduces LightGBM MAE 73.5896818466 versus seasonal-7 88.4160714286 (16.7689% reduction); WMAPE 0.8341981303. Twenty initial-training-selected products, 560 exposed evaluation observations, three chronological validation windows, one-step observed-history forecasting. GBP UK retailer data across countries, not Indian quick-commerce. No promotion without both explicit configured quality gates; the study did not change active serving.
- Restore: same-host pg_dump/pg_restore into a separate disposable PostgreSQL database preserved account/session/event counts, encrypted credential decryption with the same key, and UNKNOWN reconciliation. Post-snapshot writes were absent as expected. Not off-host disaster recovery or a recovery-time SLO.
- Browser: [workspace](workspace-browser-results.json), [mounted workflows](workflow-browser-results.json), and [UI source trace](ui-source-trace-results.json) retained from isolated simulation runs. Source tracing covers three displayed metrics across eleven surfaces; it proves that specific UI/API lineage, not universal data truth. Browser runtime used existing Python 3.14 Playwright; backend tests use Python 3.11.
## Runtime limits and authority
Model use is disabled by default. Both HYPERFLOW_ENABLE_LLM=1 and HYPERFLOW_LLM_FREE_QUOTA_CONFIRMED=1 are required; the latter is an operator assertion, not billing detection. Twenty physical requests per day and 120 per campaign are enforced globally alongside account limits. No actual model calls were made in this campaign.
Completed turn IDs return their stored response; changed payloads conflict. Failed/incomplete turns are inspectable and are not automatically resent. This is conservative durable recovery, not continuation of an interrupted model loop. The 60-second deadline is cooperative between steps; an in-flight provider call can outlast it. Model/provider text cannot approve actions or choose account credentials. Live purchases and bookings are disabled in this campaign.
## Verification policy and next action
The machine-readable [verification summary](verification-summary.json) records the final suite outcome and source hashes. Historical reports remain scoped to their original runtime/source. Default pytest selection excludes historical HTTP specifications; passing selected tests does not certify every legacy module.
Next: configure credentials locally for read-only discovery; conduct the three-task independent pilot. Do not put secrets in chat, documentation or evidence. No paid requests, purchases, bookings or public deployment were performed.
## Latest repair evidence
The substitution repair is bound to [current intent results](intent-substitution-results.json) and [solver results](solver-substitution-results.json); prior reports remain historical. The [current regression log](verification-substitution-tests.txt) records 344 passes with PostgreSQL enabled, zero failures/skips. CI now runs the seeded solver and development/validation intent report and retains offline evidence. This workflow change was checked locally; no remote CI run is claimed. Docker now includes the evaluation corpus required by the offline CLI, but no container build was possible without the daemon.
The [claim-by-claim audit](claim-audit.md) independently reran 344 maintained tests with PostgreSQL and the offline parser/solver evaluations. It recalculated retained load results and corrects “reserved cases untouched” to “not evaluated; public labels are loaded.” External gates remain open.
## Public deployment — 23 October 2026
- Live at https://gaurav711-hyperflow.hf.space: one Docker container on a Hugging Face Space running the API and the worker (`deploy/start.sh`), with Neon PostgreSQL 18.6 (`us-east-2`, direct unpooled endpoint, which session advisory locks require).
- [Live smoke results](deploy-smoke-results.json): 9/9 passed. These cover health and readiness with a recent worker heartbeat, security headers and HSTS, every protected `/api/v2` route in the OpenAPI document refusing anonymous callers, and a reviewer replay journey (an unapproved proposal cannot execute; the approved action executes once and replays by idempotency key; logout revokes the session).
- Live provider mode remains disabled; reviewers use isolated replay workspaces.
- Not yet verified: the number of proxy hops in front of the Space (`HYPERFLOW_TRUSTED_PROXY_HOPS=1` is assumed), the off-host restore drill against Neon, and Cart Guard store parsers, which wait on captured pages.
- Worker idle polling: the Space sets `HYPERFLOW_WORKER_IDLE_SECONDS=1800`. A worker polling every 2 seconds keeps Neon's free 0.25 CU compute awake permanently (about 180 CU-hours a month against the free 100), so idle polls run every 30 minutes and Neon scales to zero between them. Queued work drains every 2 seconds once found; a new training job can wait up to 30 minutes. Readiness treats the heartbeat as stale after three idle intervals.
- LLM vs rules ([NVIDIA results](intent-llm-nvidia-results.json), [partial Gemini results](intent-llm-results.json)): all 100 cases on `nvidia/nemotron-3.5-lightning-30b-a3b` with reasoning switched off, the same instruction, grounding validator and grader, and 0 provider errors. Rules 82/100 strict (reserved 17/20) against the model's 32/100 strict (reserved 9/20). A secondary, post-hoc name-tolerant score, which accepts 'rice' inside '1 kg rice' while every other field stays exact, gives the model 58/100 (reserved 11/20). The model passed no case that the rules fail. Its 26 remaining accepted mismatches are all product-name wording; quantity, unit, brands, substitution and budget were right. The run found one real safety gap: for '-1 L milk' the model returned 1000 ml and the grounding validator accepted it. The validator now rejects negative quantities, and that case is rejected. Conclusion: model-assisted parsing stays disabled; the rules are more accurate here. Gemini's free tier (20 requests per model per day) scored only 20 cases before its quota ran out. Model p50 latency was 1.3 s.
|