AI Evaluation

FixNow uses Google Gemini 1.5 Flash for two jobs: value scoring at registration time, and resolution auditing when judging whether a closed issue was actually fixed.

Value score (0–100)

On registration, the issue body is sent to Gemini with instructions to return JSON:

{ "score": 0, "reasoning": "one sentence" }

Guidance given to the model:

  • 100 — extremely high value (critical bug, large performance win, essential feature)

  • 0 — trivial or low impact

If the API key is missing, the app returns score 50 with reasoning that the key is not configured. On model or parse failure, the score falls back to 0.

Resolution audit

evaluateResolution classifies comment threads into one of:

  • implemented — the requested change landed

  • ignored — closed without a real fix

  • efficacy_alternate — a different but effective solution was accepted

  • unknown — key missing or evaluation failed

The classifier prompt asks for the category name only. Treat outputs as assistive signals, not authoritative project management truth.

Model choice

Gemini 1.5 Flash is used for low latency on first-pass assessments suitable for registration UX. Upgrade paths (larger models, multi-signal scoring with reactions and linked PRs) can be layered later without changing the public registration flow.