IZIFIT team contribution, recorded 27 September 2026
Coaching that stays within the facts.
Improving Nutrition Day feedback without inventing a user's habits, overstating nutrition outcomes or reducing a coach's personality to one repeated joke.
The result, and its limits.
| Campaign | 86 provider calls using synthetic inputs. |
|---|---|
| Accepted review set | 24 cases: 16 variety responses and eight factual controls, in Russian and English, across the Tony and Jane coaching voices. |
| Result | Accepted against the variety and factual criteria below. |
| The final repair | The targeted check passed. A full rerun of the review set wasn't done. |
| In production | Deployed on 27 September 2026. Newly generated coaching text wasn't observed in production during the verification. |
Keep the coaching grounded.
The coaching response needed to recognize the recorded day while preserving each persona's warmth. Repeated jokes about an unnecessary trainer made different responses feel alike. Other failures went further: they inferred habits or discipline, described unsupported body benefits, or discussed metrics when a target was missing.
I kept nutrition facts governed by the application and used the selected persona for tone. Quiet approval could be valid; a joke was not a requirement. The follow-up changed prompts without adding response postprocessing, history lookup or regeneration to the product.
A review designed around specific failure modes.
The harness used the actual product projection and prompt assembler, normal streaming transport and a fresh runtime-configuration snapshot. It blocked application database and media access. There were no transport retries or fallbacks, and the model settings stayed unchanged.
The variety checks used four independent generations from the same facts for each language/persona pairing. Factual controls covered protein shortfall, excess, uncertain causes, substantial shortfall, missing targets, Cheat Day and instruction-like food names. Female and unknown-gender inputs were included.
| Criterion | Recorded result |
|---|---|
| Unnecessary-trainer theme, at most four of sixteen | Three of sixteen accepted variety responses |
| Repeated comic premise within a four-response group | At most twice per group |
| Factual controls | All eight controls passed their defined review |
These are bounded review criteria, not an overall accuracy percentage. English Jane still tended toward similar quiet approval, and verdict-first openings remained common. Passing the thresholds did not establish broad semantic variety or guarantee future non-repetition.
A targeted final repair.
A Russian response described alignment with the target as a rare occurrence. The supplied day contained no history that supported that frequency claim. The final correction reinforced the factual boundary, and a targeted check of the repaired response passed.
Earlier failures also informed tighter factual scope, persona instructions limited to tone, and conditional protein advice that did not promise a numerical result from a food substitution.
What shipped, and what was observed.
The focused nutrition briefing, source and route suites passed 41 tests. The whole-backend typecheck retained ten pre-existing diagnostics in four unrelated test files, unchanged from the recorded baseline.
The API and worker were delivered on 27 September 2026. Health checks passed and deployed source hashes matched the tested nutrition assembler. That release did not include portals or mobile artifacts.
No production generation was triggered, so new customer-visible prose and production saved-result reuse remained unobserved. This verification covered source tests, synthetic provider checks and deployment; it did not include a new native acceptance pass or establish a real-user outcome.
Evidence basis.
This account is based on the internal Nutrition Day message-variety engineering record dated 27 September 2026. Its ledger identifies the accepted cases, every provider call, first failures, source identities and release checks. Private transcripts, credentials and internal runtime records are not published here.
Discuss the evaluation approach