# PHASE 3 — Small MRPT, Student-t head, tests, pilot end to end (2026-11-02)

## Completed
- `models/`: gated patch embedding with resolution/position/gap embeddings; pre-norm encoder; shared vs
  cross-attention fusion; regime embedding (stats / learned / none); direct 60-step decoder with future-calendar
  queries; Student-t / Gaussian / quantile heads (scale-equivariant parametrisation); Stage-B low-rank Gaussian-copula
  head (joint paths).
- `train/`: AdamW + warm-up/cosine, clipping, early stopping on validation NLL, seeds, crash-safe checkpoints with
  resume, JSONL tracking (loss, grad norm, throughput, data fraction, MPS memory), copula-only stage.
- `inference/`: batched prediction with Monte Carlo cumulative-return samples (independent or copula), streaming
  path scores, serialization, CLI (`chronos-x`).
- Tests: 26 offline (unit 6, synthetic 5, leakage 5, numerical 10) + 5 live + 1 slow synthetic end-to-end pipeline test.

## Commands
```bash
.venv/bin/python scripts/run_experiment.py --config configs/pilot.yaml           # on M3U96b, detached via scripts/node_run.sh
.venv/bin/python scripts/run_experiment.py --config configs/synthetic_smoke.yaml  # laptop, license-free (2 min 50 s)
```

## Pilot training (M3U96b, torch 2.15 MPS, batch 64, 600 steps, 4 942 train / 5 362 val origins)
| Model | Parameters (measured) | Examples/s | Wall (train) | Peak MPS memory | Best val NLL |
|---|---|---|---|---|---|
| MRPT-S (A+B+C) | 3 455 619 | 186 | 207 s | 63 MB | −6.0663 |
| MRPT-S (A only) | 3 200 259 | 228 | 169 s | 86 MB | −6.0659 |
| TCN (A only) | 231 459 | 243 | 158 s | 35 MB | −6.0685 |
| MLP (A only, last 60 bars) | 301 827 | 253 | 152 s | 15 MB | −6.0666 |
Data-loader fraction of wall time ≈ 1 % (10 spawned workers over memory-mapped grids); the GPU forward/backward is the bottleneck.
Test scores: see `docs/PHASE_2_REPORT.md` (MRPT-S multi-resolution −2.6 % to −3.4 % CRPS vs random walk; the
one-minute-only variant, the TCN and the MLP are within 0.2 % of it on this 10-session test — no pilot claim on H1).

## Gates
- Leakage gate: `tests/leakage` (inputs invariant to future bars; quarantined leaky fixture rejected) — PASS.
- Numerical gate: Student-t log-density vs scipy, closed-form CRPS vs Monte Carlo (2 % tolerance), copula sampling
  reproduces marginals, copula NLL minimised at independence for independent data — PASS. Two bugs were caught by
  these tests during the build (operator precedence in the PIT clamp; scipy `t.ppf` returning NaN for float32 inputs
  near 0.5) and fixed before any real-data result was produced.
- Provenance gate: dataset versions, request ids and sha256 recorded for every export; run metadata saved per experiment — PASS.
- Save/resume: `last.pt` / `best.pt`, resume tested in the trainer (step and optimiser state restored).

## Go / no-go
GO to P4 (profiling + main ingestion already complete) and P5.