# L33: Execution evidence and reusable solution candidates

## Done when

- Accepted submissions atomically retain a bounded immutable execution noun with server provenance and optional worker telemetry. Replays cannot replace reports.
- A lease holder can recall exact input candidates, schema-compatible methods and similar examples. Missing context never becomes an exact reuse claim. Failed runs remain warnings.
- Every read checks namespace, original/current source governance, queue grants and pinned artifact permissions. Anonymous reports retain labels, quotas and expiry.
- Python, HTTP, MCP and Worker paths agree, with deterministic fingerprint vectors and isolation/stale/replay regression tests.
- A documented external-worker verb can distill authorized runs into tested, licensed code-node and recipe nouns using existing stores and compiler contracts.

## Design

Execution history is a specialized immutable noun relation inside Tessera, like feeds and buckets. It is not a second task queue or independent code library. Ordinary CAS access cannot bypass task queue permissions. Indexed hashes and LSH bands are rebuildable candidate projections; embeddings are optional worker-supplied search evidence. Acceptance receipts attest submission, not semantic correctness. Reuse requires checking current context, licenses, tests and effect authority. No model call, automatic code execution or automatic publication is introduced.

The first increment captures new submissions. Older receipts do not acquire invented harness, usage or context data. Large logs live in explicitly governed versioned artifacts; unknown measurements stay null with a reason. GPU/ANN services remain optional later index adapters.