# BTC 6H observation: revised export baseline

Use `results/mlp_observation/btc6h_20260924_revision2` for subsequent observations.
The original `btc6h_20260924` bundle remains unchanged as historical evidence.
This is an explicit new prospective baseline, not permission to accept altered
history into an existing observation series.

## Revision review

The latest exports pass structural validation with historical coverage/gap
warnings. All five BTC timeframes pass score parity against their saved winning
models, with zero threshold-side disagreements. Maximum absolute score difference
is 0.000337, below the 0.01 tolerance. The parity checker previously interpreted
Unix seconds as nanoseconds and misleadingly reported missing scores. Its parsing
is corrected, with Unix/ISO equivalence covered by the eight passing focused tests.

The latest BTC 6H export differs from 53 captured historical input rows, with no
captured timestamps missing: one row on November 14, 2022 and 52 rows during
July 31–August 21, 2026. The original raw August export was overwritten; the
original freeze preserves combined row hashes rather than original feature values.
Consequently, the exact original column/value changes cannot be established from
that freeze. Neither a particular upstream data provider nor layout state is
proven to have caused them. This is not a timestamp-format difference.

Archived signal traces permit a direct behavioral comparison:

| Policy | Available comparison through | Changed action bars | Changed position bars | IS return / drawdown |
|---|---|---:|---:|---|
| Incumbent | August 21, 2026 06:00 UTC | 0 | 0 | Unchanged |
| Original hybrid9090 | February 28, 2026 18:00 UTC | 0 | 0 | Unchanged |
| Higher-low width 3 | August 21, 2026 06:00 UTC | 0 | 0 | Unchanged |

Scores do change: maximum absolute difference is 323.0461 for the incumbent and
higher-low traces, and 5.4436 for the available hybrid trace. Thus these changes
are not dismissed as rounding. Incumbent and higher-low old OOS results also
remain unchanged, with zero trades. No equivalent hybrid OOS trade-trace claim is
made. All three policies pass the twelve-perturbation sensitivity check again on
the revised data. Parameters, model bytes and entry gates are identical across
the two freezes; no candidate was selected or retuned from the revision results.

## Prospective restart

The new freeze was recorded at **2026-09-24 19:29:30.683790 UTC**, after the review.
Its observation boundary remains **2026-09-25 00:00 UTC**, the same next six-hour
boundary as the original freeze. No forward observations existed in either
bundle at restart. All three policies report `awaiting_data`, with null metrics.
The latest retained bar is September 24 at 12:00 UTC; this is expected today.

Run this when another export is available:

```bash
.venv/bin/python tools/validate_chart_data.py
.venv/bin/python tools/mlp_observation.py observe \
  --bundle results/mlp_observation/btc6h_20260924_revision2 \
  --data "data/mlp/COINBASE_BTCUSD, 360.csv"
```

Future history revisions still fail closed and require review. Keep each raw
export; this checkpoint preserves today's files byte-for-byte in gzip form.
Do not repeat baseline restarts after seeing forward outcomes to erase losses.
The 180-day and ten-completed-trade review thresholds remain unchanged.

Score parity is established for the installed winning models, not the two
uninstalled research candidates. Matching current TradingView trade-list parity
remains unverified: the available trade lists are still dated June 25. Nothing
was promoted, and all 941 protected trading artifacts retained their hashes.

## Evidence

- `results/mlp_observation/revision_review_20260924/impact.json`: paired metrics,
  reference paths, score changes and action differences.
- `results/mlp_observation/intake_20260924_1424`: preserved exports and parity logs.
- `results/mlp_observation/btc6h_20260924_revision2`: sealed models, parameters,
  source snapshot, sensitivity results and initial observation receipt.
- `results/research_checkpoint_2026_09_24/revision2_manifest.json`: verified archive
  and member hashes for `btc6h_observation_revision2.tar.xz`.

Both new-bundle integrity and unchanged policies were checked directly. Repeating
the observation returned the same report without adding a receipt. The original
bundle's source/model seal and protected-file hashes still verify.