> ## Documentation Index
> Fetch the complete documentation index at: https://solo-09d10f60.mintlify.site/llms.txt
> Use this file to discover all available pages before exploring further.

# Evidence and artifacts

> How Enoch preserves run evidence, writes generated artifacts, and frames the corpus.

Autonomous AI research runs produce scattered evidence: notes written during execution, intermediate metrics, result summaries, and claim records. Without a system to collect and preserve that evidence before artifact generation begins, generated reports have no stable grounding — they become prose detached from the work that produced them. Enoch treats evidence collection as a required pipeline stage, not an optional export. Evidence sync is available before paper generation and rewrite work, but it is configuration-gated: `paper_evidence_sync_enabled` defaults to `false` in `config.example.json`. When enabled, the control plane attempts to copy high-signal run evidence from the worker before rewriting artifacts. When disabled, or when no local evidence is present, generated drafts must stay bounded by the available run record, evidence bundle, or claim ledger and should not be presented as evidence-complete artifacts.
For current runtime topology and worker boundaries, see
[current runtime snapshot](/current-runtime-snapshot).

Enoch separates generated prose from the evidence that grounded it. That separation is the main reason the generated corpus is reviewable.

## Evidence files

The corpus snapshot contains per-artifact folders with files such as:

* `paper.md`
* `metadata.json`
* `evidence_bundle.json`
* `claim_ledger.json`
* `paper_manifest.json`

The Hugging Face export flattens those artifacts into `data/artifacts.jsonl` and reports fields including `paper_markdown`, `metadata`, `evidence_bundle`, `claim_ledger`, `paper_manifest`, `github_url`, `ai_generated`, `human_authorship_claimed`, and `review_status`.

Before the artifact writer runs, Enoch syncs evidence from the worker project workspace to the control VM. When evidence sync is enabled, the primary sync method is HTTP: the control plane calls the worker-gate API to retrieve evidence files. If the HTTP sync is unavailable or returns an error, an optional SSH fallback uses `paper_evidence_sync_ssh_host` and `paper_evidence_sync_remote_root` from your config to copy files directly.

When `paper_evidence_sync_enabled` is true, the control plane can sync evidence from a worker project root before paper generation. Config fields include `paper_evidence_sync_ssh_host`, `paper_evidence_sync_remote_root`, and `paper_evidence_sync_timeout_sec`.

Docs should not imply evidence exists for a run unless artifact files or metadata show it.

## Artifact writing

The default writer is `deterministic`, which avoids an external model call. The code also supports aliases for a `synthetic.new` OpenAI-compatible provider. Configure provider URL, model, API key, timeout, temperature, max tokens, and fallback behavior through `paper_writer_*` fields.

Provider support does not mean generated papers are validated or peer reviewed.

## Corpus gates

The corpus packaging/provenance policy requires generated artifacts to avoid fake citations, placeholder markers, implied human authorship, and peer-review claims. It also expects provenance metadata plus evidence-bundle and claim-ledger files. The stricter claim/evidence audit separately checks whether ledgers contain evidence-linked claims and whether referenced result files are public or explicitly unavailable.

## Current local corpus snapshot

<Warning>
  Do not publish or finalize a paper until packaging/provenance lint and strict claim/evidence audit status are explicit. A paper with placeholder citations or missing provenance artifacts misrepresents the evidence behind its claims. A paper can pass packaging/provenance lint while still failing the stricter claim/evidence audit if its claim ledger is empty or its result-file references are not public.
</Warning>

## Correct framing

Operator-facing paper state is intentionally simpler than the internal compatibility enums:

| Operator lane      | Meaning                                                                                 |
| ------------------ | --------------------------------------------------------------------------------------- |
| Write              | A completed run passed the paper-positive decision gate and has no live paper row yet.  |
| Finalize           | A generated publication draft is missing its automated finalization package.            |
| Publish/import     | A finalized publication draft is missing a corpus-import ledger row.                    |
| Published/imported | The corpus-import ledger represents the finalized draft in the public corpus.           |
| Done / no paper    | A completed run is negative, ambiguous, already represented, or otherwise not writable. |

The API still exposes raw compatibility values such as `draft_review`, `publication_draft`, `human_review_required`, `in_review`, and `approved_for_finalization` for historical rows and integration clients. Treat those as internal/detail states. They should not be read as the normal operator workflow. The dashboard and state doctor derive operator counts from the paper-positive decision gate, finalization package evidence, and the corpus-import ledger.

## Provenance framing

The reports Enoch produces are AI-generated research artifacts, not human-authored or peer-reviewed papers. They are built from run notes, evidence bundles, claim ledgers, and reproducibility traces produced during autonomous agent runs. The provenance chain is:

```text theme={null}
agent run → run notes + metrics + results
         → evidence_bundle.json + claim_ledger.json (synced to control VM; strict audit is separate)
         → artifact writer (consumes evidence context)
         → Markdown report + LaTeX + paper_manifest.json
         → packaging/provenance scan
         → corpus release
```

The authorship statement for all artifacts produced by Enoch runs is:

> The reports produced by this system are AI-generated research artifacts created from automated run notes, evidence bundles, claim ledgers, and reproducibility traces. The maintainer releases the corpus for inspection and critique but does not claim personal authorship of the generated papers, arguments, or prose.

When citing a paper produced by an Enoch run, credit the system operator as the release maintainer, not as the human author of the generated content. See the [authorship and provenance reference](/reference/authorship-provenance) for recommended citation language.