# STRICT chronological reference, version 1

This supplement releases the numeric inputs and executable replay for the separate chronological reference evaluation in *Where the Numbers Come From: Auditing Evaluation in Provenance-Based Intrusion Detection*. It covers corrected Velox and non-snooped Orthrus on CADETS E3, CLEARSCOPE E3 and THEIA E3. It adds to the historical score record without replacing its files or results.

## What this reproduces

The replay checks all 64 tuning settings with three members each, reproduces the six validation-only choices, and evaluates ten fresh members per selected configuration. It recomputes tie-safe ADP, its mean and member-bootstrap interval, fixed-budget precision and campaign coverage, exact full-coverage tie bounds, maximum-validation thresholds, saved threshold decisions, and every campaign's detection count and conditional first-alert delay. The secondary tuning selection-gap diagnostic and separate four-of-ten vote with a rolling ten-alert cap are also reproduced.

The underlying selection record predates the reference trainings. Its content hash is retained as provenance, but this release is not a public preregistration and does not establish what any person previously observed. The grid uses eight settings per Velox host and sixteen, eight and sixteen per Orthrus host. It is not an equal-budget architecture comparison. Tuning seeds are 1–3 and reference seeds are 101–110. The selected checkpoint is epoch 11 after twelve training passes.

## Reproduce

Use Python 3.9 or later with the NumPy version in `requirements.txt`. No GPU, database, PyTorch, private repository or service credential is needed.

Download this version with Hugging Face's optional download client, without a token:

```bash
python -m pip install huggingface_hub
python -c 'from huggingface_hub import snapshot_download; snapshot_download("STRICT-APT/strict-score-record", repo_type="dataset", revision="strict_reference_v1", allow_patterns="reference/strict_reference_v1/*", local_dir="strict-reference-download", token=False)'
cd strict-reference-download/reference/strict_reference_v1
```

The download client is not used by the replay itself:

```bash
python -m pip install -r requirements.txt
python -B test_replay.py
python -B replay.py . --out /tmp/strict-reference-recomputed.json
```

The output destination must not already exist and should be outside this release tree. The verifier rejects missing files, extra files, altered hashes and symbolic links before loading inputs. Numeric arrays use NPZ without object arrays or pickle loading. The replay compares every computed member result and summary with the frozen expected record. Missing detections remain missing, not zero-delay successes. The executable checks provide numerical replay from the released evaluation inputs, not independent retraining from raw telemetry.

## Contents and numerical conventions

- Each host directory contains `record.json`, campaign onset arrays, each tuning member's final test scores and validation loss vector, and each reference member's scores, validation loss vector and full node/block score vector.
- Score arrays contain numeric node identifiers, scores, saved predictions, binary benchmark labels and node-to-campaign membership. Validation arrays retain loss order. Tuning means use the original serial floating-point summation. Reference thresholds preserve the original parser's maximum-loss convention.
- Block scores are maximum incident-edge losses per node in two-hour UTC blocks. The released observation calendar selects fully observed blocks. Threshold exceedance is strict and node cooldown is 24 hours. Decisions occur at block end. Campaign credit requires an overlapping attack block ending after the node's canonical attack-period onset. This does not establish that the score-maximizing event occurred after onset, and it excludes processing latency.
- The separate vote is four of ten before cooldown, followed by a rolling ten-alert daily cap. Same-block priority is vote count descending, then node identifier ascending. Rejected candidates still start cooldown. It is not the primary member result or the historical budget-calibrated policy.
- CADETS excludes April 6 and retains two campaigns. CLEARSCOPE has one campaign and THEIA two. Dates, windows, populations and every missed campaign are retained. These are benchmark annotations, not independently adjudicated malicious events.
- `metrics.py` projects the existing numerical functions without changing their arithmetic. `code_sources.json` records the source hashes and function names. `MANIFEST.sha256` binds every released file other than itself.

## Scope and license

This release includes the selection and threshold inputs needed for the reference's reported decisions. It excludes raw command/path/address strings, raw edge-event tables, training features, checkpoints, private execution logs, credentials, internal discussions and reviewer reports. It does not claim independent raw-data retraining, semantic label validation or end-to-end deployment latency. The historical score record and its separate code archive remain available under their existing versions.

Numeric data and documentation are licensed under CC BY 4.0. The replay code is licensed under Apache 2.0, supplied in `LICENSE-CODE`. Cite the paper and this versioned supplement when reusing it.