Repository navigation
Validate a clearer Sarvam reporter and document the research history - #58
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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 final reporting stage was losing otherwise usable results on saved Sarvam service histories. This adds a frozen, metered comparison that changes only its first instruction sentence, and publishes all 288 calls plus the complete Sarvam research history.
With fixed model profiles, schema and token caps, factual answers improve from 59/72 to 72/72 for base 105B and 70/72 to 71/72 for Conversations. Base empty outputs fall from ten to zero. This passes the predeclared saved-history screen, not fresh end-to-end validation or a new memory benchmark. One Conversations regression, raw status errors and the previous missed approved action remain explicit; the candidate is not promoted.
Validation: 154 focused tests pass in a clean committed export. All 1,243 artifact files verify; primary replay and an independent raw-response/accounting calculation reproduce exactly without credentials or API calls. All 72 source renderings, 16 incorrect outputs and 55 changed raw pairs were assistant reviewed. Known new usage estimate INR 38.453490; cumulative native retained INR 244.451544 remains within the existing INR 6,000 ceiling. All 6,992 founder files/status and English current state remain unchanged.
Preserve both independent commits when merging. Next phase is a documented proposal for fresh complete-workflow confirmation, not another executed paid run.