jpmschweitzer/odysseus
1
0
Fork 0
mirror of https://github.com/pewdiepie-archdaemon/odysseus.git synced 2026-10-06 15:02:20 +02:00
Code Issues Packages Projects Releases Wiki Activity

Record controlled missing-referent and schema-omission diagnostics

Browse Source
This commit is contained in:
pewdiepie-archdaemon
2026-09-17 21:44:46 +00:00
parent e066b5d4dc
commit f8fc4ec4f7
2 changed files with 11 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/search-quality-audit-20260917.md
+6
View File
@@ -75,6 +75,12 @@ Further provider inspection found that the news-to-general fallback dropped the
### Additional informal/multi-part live checks
Ambiguity controls with exact compact search schema/canonical base prompt: `reports/search-tool-choice-probe-1789681406232.json` reproduces invented game-release query even with auto choice. Adding an explicit missing-subject clarification instruction (`1789681407057`) merely changes invented query to France's capital. No network tool was executed by these probes.
`reports/search-tool-choice-probe-1789681431330.json`: tool_choice none while retaining schemas emits raw tool-call markup on two ambiguous prompts. Omitting schemas entirely with the same clarification instruction (`1789681457760`) yields correct clarification questions on all three (“can u look it up”, “please find that”, “what about its price?”), 0.9–1.2 seconds. This supports testing a genuinely context-aware missing-referent boundary, not blanket tool removal or phrase-only blocking of valid follow-ups. No production ambiguity guard deployed yet.
Casual citation replay `reports/clean-v3-search-quality-2026-09-17T21-42-14-297Z.json` failed quality: 29.3s/three rounds and final claims it cannot compare without data despite retrieved evidence. Citation recognition alone did not solve synthesis and must not be described as a successful answer-quality fix.
Extraction replay `reports/clean-v3-search-quality-2026-09-17T21-40-44-367Z.json` completed. Misspelled battery request remained too shallow/uncited (17.1s/two rounds), although it now framed cost reduction conditionally rather than saying batteries are categorically cheaper/safer. `8cd14e3f` recognizes standalone trailing citation requests such as “sources pls” and tests negative topic/negation cases. 926 relevant regressions passed. Deployed after the replay ended; live validation still pending. Concurrent agent committed the unrelated artifact-path regression as `b0a1f7fd`; that edit was not included in our commits.
`reports/clean-v3-search-quality-2026-09-17T21-37-47-091Z.json` remains weak: weekly news took 44.5s and ended by asking the user to open/scroll the page; misspelled battery comparison took 17.6s and gave shallow uncited claims. Its evidence had substantial tag/related-post/reference noise. A fresh inspection of the actual battery page found one article nested within main. Extraction now prefers a single substantive article over its surrounding main wrapper, while multiple article listings preserve main context. Real fetch: 2,413 characters, comparison retained, related posts/comment form absent. 54 extraction/observation tests pass. This does not validate the article's claims: its cost discussion is internally inconsistent, so the model must still qualify/corroborate it. Another agent's unrelated workspace-path test in `tests/test_clean_agent_preview.py` was left untouched and uncommitted by this work.