# E7 source snapshot

This archive contains customer-owned source and experiment-owned extensions only. It contains no model weights, installed environments, managed library implementation, or credential values.

## Safe offline checks

Use Node22.13.1 and Python3.12+ from the extracted directory:

```
node --test tests/*.test.mjs
PYTHONPATH=src python3 -m unittest discover -s tests -q
```

The64 JavaScript and30 dependency-light Python tests need no subject endpoint and perform no numerical fitting. The original customer code is retained under customer/.

## Recompute saved analysis

Download the report's raw collection resources into COLLECTION, preserving the original collection layout: frozen-manifests.json, journal/run.json, journal/journal.jsonl, metadata.json, status.json, request-accounting.json and worked-records.json. Then:

```
node src/analyze-local.mjs COLLECTION NEW_ANALYSIS
node src/verify-messages.mjs COLLECTION NEW_MESSAGE_AUDIT.json
node src/prepare-audit.mjs COLLECTION NEW_AUDIT
```

No source collection files are modified. The corrected endpoint exporter preserves regime-specific high-price fees separately from the base fee. A complete full-grid bootstrap requires compute and --compute; the full grid was not collected in E7.

The audit merger expects the saved raw packets in scratch/verified-analysis/audit-packets.json, the frozen selection links and twelve annotation shard files under results/. They are separate report resources, not in this code-only archive. Exact quote validation is independent of the labels. Analyst adjudication records, including borderline interpretations, remain available with the report.

## Environment and model

Subject collection used the public lmsysorg/sglang:v0.5.19 image (OCI digest d6e7288627be8b02be88e4bba38e73f6d50e2826869f753c13a4c4385ab3eda9), official Qwen/Qwen3.6-35B-A3B-FP8 revision95a723d08a9490559dae23d0cff1d9466213d989, and eight independent H100 replicas. The image supplies the numerical environment; no base torch/numpy changes were made. Numerical calibration utilities use NumPy2.3.5 and SciPy1.18.1 in that validated image. Plotly figure rebuilding uses the Lab's installed [private-platform]-figures styling API; no styling-library source is included.

Existing collection authorities and deadlines are consumed. This archive grants no permission for new subject requests or new allocations. Any new collection requires fresh researcher approval; do not blindly rerun direct_gpu_entry.py.