Repository navigation
Record corrected Sarvam reporting gains and add capped-actor recovery - #60
Merged
Merged
Conversation
Sahith59
changed the title
Sahith59
marked this pull request as ready for review
October 13, 2026 17:04
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
The corrected source questions distinguish current preferences from past requests. Across 240 native episodes with fixed Sarvam models, the clearer reporting sentence improved 105B date-plus-clarification accuracy from 90/120 to 105/120. Conversations moved 95/120 to 96/120 but wrong non-null dates increased 18→22. A capped native response left two reporter slots unreached. The frozen acceptance gate failed; no candidate is promoted.
This PR preserves all 844 calls, 478 issued reporters and two missing slots, complete assistant review, costs and negative results. It adds lossless source/receipt/actor evidence views and a metered three-arm comparison harness; that proposed 108-call comparison remains unexecuted because its parent prerequisites failed. A separate opt-in capped-actor handoff prepares tool-free reporters from verified history without retrying actions, changing frozen results or claiming a live recovery gain.
Validation: 234 focused Sarvam tests pass. A clean export verifies 3,348 artifact hashes, reproduces request/receipt/SQLite replay and separate raw parsing/accounting, regenerates the frozen cases, and reproduces exact capped-input handoff preparation without dispatch. Models, English inference and existing result artifacts were preserved. The synthetic related cases and assistant-only language review do not establish a novel memory advantage, a held-out benchmark or production readiness.