Conversation
A 0.3 usage edge is the resolver saying "I could not type the receiver, this bind is a guess", and extract's law stamps it below blast's traversal floor precisely so impact analysis ignores it. But MinConfidence is also the depth control, so the MCP server's 0.3 (a chain of real 0.7 edges has to survive three hops) re-admitted every guess: discourse's Admin::BadgesController, whose whole claim on the word `new` is a two-line empty Rails action, served 60 direct callers of which 59 were spans calling some other class's .new, risk high and total_affected 357 against a true 6. The floor on a single edge's own confidence is now separate from the cumulative-path one, and applies to calls/references only: a 0.3 composes edge is a weak association and a temporal confidence is a co-change ratio, neither of them a name guess. That call now returns the six route entries, "6 direct callers" and verdict complete - the answer min_confidence 0.6 already gave. sense_graph is untouched and still shows the stratum. Collateral, measured on the same indexes: 20% (discourse) to 27% (rails) of symbols with any inbound usage edge now report zero direct callers. That stratum is already invisible at every other default floor - CLI blast, sense_graph and the summary's hub counts all cut at 0.5 - and 82-86% of it points at a name several symbols share, i.e. a coin flip among N. Where it hid a real row it hid it among fakes: maket's Account::Cart::DashboardController served 102 route helpers from unrelated routes, all bound to one `#show`.
Codecov Report✅ All modified and coverable lines are covered by tests. 📢 Thoughts on this report? Let us know! |
|
Rebench by sense-lab at this PR's build,
Closed by sense-lab: a WORSE rebench does not merge. |
|
Correction from sense-lab: the WORSE result above was not a measurement of this change. The Claude arm ran claude-opus-5-5, a newer model than the claude-opus-5 the bench was banked on, and it never called Sense in either draw, so both draws measured grep. sense-lab should not have scored them or closed the PR. It now leaves such draws out, runs a rebench on the model the bench was banked on, and never closes a PR. A rebench on claude-opus-5 follows. |
|
Rebench by sense-lab at this PR's build,
Nothing was measured that can judge this PR. It stays open. |
|
Correction from sense-lab: the "claude opus VOID" line above is left over from the earlier claude-opus-5-5 run and is not part of this rebench. This rebench, on claude-opus-5 (the model the bench was banked on): SAME. Sense arm 1.00 before and 1.00 after, plain arm 0.42 before and 0.33 after, over 2 draws, and no row the finding says Sense already carried was lost. The Sense arm was already at 1.00 on this question, so this rebench shows no regression; it cannot show a gain. Ready for a person's review. |
|
Rebench by sense-lab at this PR's build,
sense-lab measured this PR WORSE. A person decides whether to close it; sense-lab never does. |
|
Rebench by sense-lab: this PR's build
Ready for a person's review. sense-lab never merges. |
This branch has not been deployed
1 participant
Opened by sense-lab for #323.
Closes #323
sense-lab rebenches this PR at its own build and posts the result below. A person merges; sense-lab never does.