[Sahith59]

The completed ordinary-source comparison separates answer truth, safe backend outcomes and reporting-contract errors across 288 episodes. It finds four unsafe writes in the unenforced Conversations arm, none in the guarded arm, and 18 empty base105B final outputs after useful native answers or successful writes. The frozen runtime, models, raw outputs and original grades are preserved.

Add offline aggregation with explicit read/action denominators and a same-call backend replay that flips only approval enforcement. Publish all 723 request receipts, SQLite states, frozen contracts, usage accounting and the assistant review of all 89 flagged episodes. No memory implementation or general multilingual accuracy claim is made.

Validation: 143 focused tests, Ruff/diff checks, full request/state/meter replay, prospective attribution and independent aggregation. A clean committed Git export verifies all 3,083 artifact hashes and exactly reproduces all four audit outputs. New comparison estimate INR65.255115; cumulative native research retained INR196.384594 within the existing INR6,000 ceiling. No unknown new usage or retries.

Five independent commits are preserved. English baseline and founder files are outside this change. Detailed results, scope limits and the next single-variable reporting experiment are in the evidence report.