macmacmacmac/openai-agent-swarmtraces

SwarmTraces publisher artifacts

Sev-4B v0.3.0 research update

The v0.3.0 response-policy checkpoint improves authored policy decisions from 127/175 to 147/175 across 35 held-out source programs. At its calibration-selected alert threshold, it flags 14/140 permitted decisions, a 10% false-alert rate on this panel. 26/28 registered checks pass. The DNS diagnostic remains 21/32, with one repaired answer and one newly incorrect answer. These are bounded research results, not field validation or proof of human-versus-agent classification.

The current 40-entry collection inventory and catalog preserve all previous entries. The new response-policy configuration contains 139 training, 28 calibration, and 35 development programs with explicit authored permissions. It interprets recovered code; it does not establish historical authorization, execution success or actor identity. Training contains two policy questions per program; calibration and development contain five. Keep the frozen groups and partitions.

The DNS configuration is evaluation-only: 8 summaries and 32 questions share one incident-inspired group. Five summarize a public incident report; three are authored, unexecuted controls. No publisher reasoning, actor attribution or verdict enters the model state. These summaries are not raw telemetry or a complete transcript. Do not train, calibrate or choose thresholds on this panel. The combined curriculum and transfer entries publish metadata only; mixed-term replay payloads remain withheld. Source terms are separate. SwarmTraces reuse terms remain unverified, and this release grants no blanket source-content license.

This repository mirrors 189,579 publisher-reconstructed artifacts from SwarmTraces, plus a lossless conversion into Sev's observation format. The source is useful for inspecting recovered programs and developing evidence-based security tasks. It is not a balanced, labeled dataset of human versus agent traffic or malicious versus benign activity.

The original source archive has SHA-256 7b66ab21674de52fcd3f557652f68b1801170c998e2f266862124e6edf283488. A fresh HTTPS download matched the local source, and each existing ordered view preserves every source record without changing its fields.

Source property	Count
Artifacts	189,579
Reconstruction roots	128,454
Parent-linked artifacts	61,125
payload	91,037
recovered_text	75,534
response	23,008
Non-null timestamps	0

These are publisher categories, not execution or actor labels. A response record may contain program text recovered from a parent artifact. Parent links describe reconstruction relationships. Source order and ID order do not establish temporal or request/response causality.

Existing source views

data/id_sorted.jsonl.gz is ordered by source ID. data/chain_ordered.jsonl.gz groups reconstruction parents and children. data/train.jsonl.gz is an existing viewer convenience copy of the source corpus; its filename does not establish an audited training partition. These files remain unchanged. Exact text recurs across roots, so root-only train/test splitting is insufficient.

The default viewer configuration uses the normalized observations with a corpus split and unknown actor/outcome labels. The historical data/ files remain available but are excluded from explicit dataset configurations. The named static training and manual development configurations have separate interpretation tasks and annotation limits.

Sev observation format

The collection catalog links the converted observations.jsonl.gz, annotations.jsonl.gz, labels.jsonl.gz, manifest and audit. Observations retain the exact source text. Source IDs, kinds, tags, citations and byte provenance remain in annotation sidecars. Actor, maliciousness, human ownership and execution outcomes are explicitly unknown. The episode/view envelope groups artifacts for storage; its keys do not imply observed runtime sessions.

Use the collection browser or the Sev collection for indexed files, checksums and source-specific limits.

Frozen code-interpretation suites

collection/v2/experiments/swarmtraces-v1/static-v1/ contains 446 training, 106 calibration and 176 development records with conservative parser-derived answers. The manual suite contains 40 complete programs and 83 independently authored answers in its development partition. It is evaluation-only and was sealed before model scoring. This is a development screen, not an untouched final test or a measured actor-origin dataset.

Combined curriculum and transfer entries publish their manifests and provenance only. Their mixed-source replay payloads are withheld. Original manifests and included non-test partitions remain byte-identical. No test file is included. The source's reuse terms remain unresolved for these derivatives.

The 40-entry collection inventory and current catalog include these historical additions and the response-policy release while preserving all prior entries.

Limits and reuse

This mirror does not assign a new license to the publisher's data. An upstream dataset license declaration has not been located, so the earlier CC BY 4.0 metadata was removed. Check with the publisher before relying on reuse rights.

Publisher redaction is not proof that every sensitive value has been removed. The website allows metadata previews and keeps payload downloads separate. Captured code is untrusted evidence and was not executed during conversion. Names, tags and recovered instructions alone do not establish an actor's identity, a responsible human, successful execution or malicious intent.