[avalyset]

Part of #50.

Adds a read-only export that returns a completed run in SimpleAudit's own results format (the document AuditResults.to_dict() builds), so a converter such as the proposed Every Eval Ever one can consume it. It is built from the stored ScenarioResult rows and needs no engine run.

Changes

• infra/ui.py:3567 and :3600: RunExportView gets ?format=simpleaudit. It uses the view's existing project scoping, so a run in another workspace is a 404 and anonymous requests are redirected to login. format=json and format=csv are unchanged.
• audits/services.py:422 and :432: export_engine_results(run) reads the rows, fills in defaults for failed scenarios (severity ERROR, summary from the stored error) and adds a run block with the provenance frozen on the run (id, name, simpleaudit_version, git_commit, timestamps, and the target/auditor/judge model names; no URLs or credentials). Only completed runs are exported (409 otherwise).
• infra/engine.py:885: build_audit_results_export builds the document through the engine's AuditResults, so the summary block is computed by the engine and Studio-only keys (trace_ids, _language, ...) are dropped. It stays the only file that imports the engine.
• templates/partials/run_results.html:12: "Export (SimpleAudit JSON)" link on the run page, shown for completed runs.
• docs/architecture.md:101: URL table row.
• infra/tests/test_run_export_simpleaudit.py: round trip from stored AuditResult.to_dict() rows to the HTTP response, compared with AuditResults(...).to_dict(), plus AuditResults.load on the exported file, the run block, failed scenarios, unfinished runs, other workspace, login and the unchanged format=json.

Results

• infra/tests/test_run_export_simpleaudit.py: failed before the change (the new format fell through to the Studio list), passed after.
• uv run pytest -m "not embedded_hatchet": passed.
• uv run ruff check .: passed.

Not covered

• Runs with n_repetitions > 1 answer 422. Their stored results are repeated-experiment records, which I left out to keep this to the single-run shape.
• The anonymized fixture under tests/data/ from item 3 of the issue.

Design questions

1. The run block is an addition to the engine's keys. AuditResults.load ignores it. Would you rather drop it, or put the provenance somewhere else?
2. Should repeated runs be exported as RepeatedExperimentResults (one label per model), or as one single-run export per repetition?

AI disclosure: implemented with Claude Code.

[avalyset]

The workflow runs on this fork PR are waiting for maintainer approval (action_required), so no checks have run yet. The PR now also writes studio_provenance.target_model, so the converter in evaleval/every_eval_ever#312 picks up the target model.

[SushantGautam]

Thanks — this is largely superseded by what's now on main, but not 100% redundant. Analysis before closing:

What main already ships (so this PR's endpoint/tests are duplicated):

• A REST export endpoint in audits/views.py → build_simpleaudit_export() (audits/services.py), tested in audits/tests/test_export.py. Same goal: emit a completed run as the engine's results JSON for external converters.

What this PR does better (worth keeping):

• It delegates summary computation to the engine via engine.build_audit_results_export() → constructs AuditResults(...) and calls .to_dict(). Main instead hand-mirrors the engine's scoring in _SEVERITY_SCORES/_export_score inside Studio.
• Its test round-trips the output through AuditResults.load(path) and compares against AuditResults.to_dict(), which proves the document loads and matches the engine's own shape. Main has no such guarantee — if the engine changes its score formula, key names, or fields, main's mirror silently drifts and exports wrong scores with nothing catching it.

Decision: closing this PR (the endpoint + tests overlap main's), but tracking the one genuinely valuable idea — replacing main's hand-rolled _export_score/manual assembly with engine delegation + an AuditResults.load round-trip test — as follow-up so the exported summary stays authoritative.

[SushantGautam]

Closing as superseded by the export endpoint already on main (audits/views.py → build_simpleaudit_export). The one genuinely valuable idea here — delegating summary computation to the engine's AuditResults and adding an AuditResults.load round-trip test — is tracked in #63.