You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Commit 660f50c
Browse filesBrowse the repository at this point in the historyBrowse files
Copy file name to clipboardExpand all lines: eval/knowledge-depth/README.md
+2Lines changed: 2 additions & 0 deletions
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -73,6 +73,8 @@ JiuwenSwarm explicitly selects `semantic_depth.acceptance_mode: lightweight`; ot
73
73
74
74
The campaign runner defaults to 13 isolated workers and accepts `--workers 13`. Each worker owns its checkpoint, model gateway and trace archive. The parent builds one immutable source/test/document index bound to the source pin, policy and index version. Initial extraction packets offer up to 24 KB of source/test lines, 8 KB of relevant document lines and 4 KB of existing knowledge; subsequent corrections retrieve evidence for the specific gap. Format, references, hashes, receipts and counts are checked incrementally; comprehensive audits, retrieval acceptance and full CI run at delivery.
75
75
76
+
For a catalog produced by feature discovery, pass `--discovery-record /absolute/path/to/init/<repo>/feature-discovery.json` to `run_depth_campaign.py`. Use the original completed `published` record after its PR merges, or the original `empty` record when the catalog was already current. The native prerequisite gate checks the real PR state, source pin and exact catalog/report hashes at the campaign baseline. The campaign binds the record's byte hash and copies its exact bytes into every worker before dispatch; native trace paths stay unchanged outside Git. A baseline containing a discovery report requires this explicit handoff, including repositories whose older manifest did not require discovery. Changed records or mismatched worker prerequisites require a fresh campaign; statuses and prerequisite records are never synthesized. Legacy baselines without discovery retain their existing behavior.
77
+
76
78
Each missing facet receives an initial attempt and at most three corrections. The round is saved before dispatch, including review-only retries. Restarting or `--retry-unfinished` does not reset this lifetime limit. Passed facets are frozen; failure in another facet does not discard them. After exhausted repairs, the facet remains unknown with its reason. Creating the campaign's STOP file stops new dispatch and drains active calls; the supervisor records progress in `drain.json`.
77
79
78
80
Validation knowledge labels runtime assertions, source-text assertions, helper-unit assertions or existing documented manual checks. Manual entries include their documented action and expected observation and explicitly say this batch did not execute them. Missing tests never become verified absence, and knowledge about assertions does not mean tests passed.
0 commit comments