Repository navigation
Study Sarvam source interpretation with all 240 histories - #62
Merged
Merged
Conversation
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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 previous agent's prose can affect final answers even when original source evidence is present. This adds and executes a frozen three-arm comparison on all 240 saved Sarvam histories: ordinary chronology, lossless source grouping, and grouping with prior agent prose withheld. Models stay fixed; no new actor or backend action is executed.
All 720 reporter calls completed once. Prose omission moves date-plus-clarification accuracy from 104 to 105/120 on base105B and 96 to 103/120 on Conversations. It also increases false abstentions and worsens status reliability, so the predeclared gate fails and the candidate is not promoted. Paired scenario-cluster intervals include zero. The report preserves all gains, losses, incomplete outputs and identical-input variability; it makes no new memory or English baseline claim.
Validation: 246 focused local preflight tests, two matching raw-response audits, complete assistant source/output reviews for all720 slots, and a clean Git export verifying3459 archived files. Clean export also passes five portable tests; the sixth fixture-dependent test was locally exercised and explicitly deselected there. Human/native-language review remains absent. Cost: ₹93.896460 known / ₹117.370843 retained. Existing founder files and closed studies remain unchanged.
Includes reproducible evidence, current README/research roadmap, prior-art limits and the next offline evidence-contract proposal. This PR preserves the individual implementation, results and verification commits.