What has been implemented, what has actually been measured, and what is still blocked. Published so the difference between the three is visible rather than implied.
| Check | Result | Status |
|---|---|---|
| Backend tests | 799 passed, 1 skipped | PASS |
| Frontend tests | 90 passed | PASS |
| Lint | clean | PASS |
| Format | clean | PASS |
| Typecheck | clean | PASS |
| Production build | compiled | PASS |
| Branding audit | passed | PASS |
| Database migrations | downgrade and upgrade round trip, 5 checks | PASS |
Recorded when the validation script last ran, at the commit shown above. This is a snapshot, not a live continuous-integration feed.
| Component | Status | Basis |
|---|---|---|
| Database migrations | VALIDATED | Full downgrade-to-base and upgrade-to-head round trip, including native enum types. |
| Paper broker architecture | VALIDATED | Paper-only by construction; a non-paper endpoint is rejected before any request is made. Tested without credentials. |
| Risk engine | VALIDATED | Position sizing, exposure caps, daily loss limits, and trade counts. Every order passes through it. |
| Kill switch | VALIDATED | Persisted, and re-read on start, so an engaged switch survives a restart. |
| Backtest engine | VALIDATED | Documented fill assumptions, no lookahead, and chronological train/validation/test separation. |
| Scheduler | VALIDATED | Completed bars only, driven by the US market calendar including half days, with overlap prevention and backoff. |
| Restart recovery | VALIDATED | State is reconciled against the broker on start, and a blocking discrepancy stops the run rather than trading through it. |
| Duplicate-order protection | VALIDATED | Deterministic client order identifiers, so a retried cycle cannot submit the same order twice. |
| Market-regime engine | PARTIALLY VALIDATED | Classification logic is tested. Its effect on results has only been observed on daily bars, which are not the intended timeframe. |
| Alerting | VALIDATED | Thirteen alert types across modular providers; one failing channel cannot suppress the others. No test contacts a real service. |
| OpenClaw permission boundary | VALIDATED | The agent runtime cannot reach order execution or broker credentials. Tested as a boundary, not as a convention. |
| Strategy experiment approval flow | VALIDATED | Experiments are proposals. Promotion requires an explicit human decision; nothing self-modifies in production. |
| Cost tracking | VALIDATED | Metered and estimated spend are recorded separately and never summed. Budget exhaustion stops research, never risk management. |
| AI ensemble contribution | NOT YET VALIDATED | Whether the AI layer improves decisions has not been measured, because the models it depends on have not run. |
| Out-of-sample trading edge | NOT YET VALIDATED | No result to date supports a claim of an edge. See the backtest and walk-forward sections below. |
No forecast has been produced by the real model, so there are no accuracy, latency, or compute-cost figures to report.
The provider does not fall back to a substitute model. A run configured for real inference either uses the real weights or fails.
The execution path was validated against a locally constructed model with random weights: weight loading, tokenization, the forward pass, batching, device placement, and label ordering all run. Random weights produce no meaningful sentiment, and no accuracy claim is made from that exercise.
It did surface a real defect: the model orders its labels positive, negative, neutral rather than alphabetically, so the adapter reads the ordering from the model configuration instead of assuming it.
No GPU was present in the validation environment, so there are no GPU or VRAM figures. Accelerated execution is implemented; requesting it where no accelerator exists raises rather than silently running on CPU.
Sufficient verified intraday dataset has not yet been acquired in the validation environment.
Blocked for the same reason as SPY: no verified intraday dataset.
AAPL daily is not the intended production strategy timeframe. The strategy under test is designed for 15-minute bars.
A backtest is only as trustworthy as the provenance of its bars. Datasets without a known provider, feed, and adjustment status were not used.
The tested strategy underperformed buy-and-hold on this historical period.
5 non-overlapping windows, 7 trades in total. These results do not demonstrate a repeatable trading edge.
| Validation | Status | Reason |
|---|---|---|
| SPY 15-minute baseline | BLOCKED | No verified SPY 15-minute dataset has been acquired, and daily bars were not substituted for it. |
| SPY regime ablation | BLOCKED | Depends on the SPY 15-minute dataset above. |
| Real Chronos ablation | BLOCKED | The forecasting model has never executed, so there is nothing to ablate against. |
| Historical FinBERT sentiment ablation | BLOCKED | No news archive with genuine publication timestamps covers the test window. Historical articles were not simulated to fill the gap. |
| Full AI ensemble | BLOCKED | An ensemble missing two of its three model inputs is not the full ensemble. |
BirdEye Trading is a validated paper-trading and research platform.
It does not currently have demonstrated evidence of a profitable out-of-sample trading edge.
The intended SPY 15-minute strategy and real AI ensemble have not yet been validated because verified intraday data and real model weights were unavailable in the current validation environment.
Live trading remains disabled.