I build AI features that would rather be right than impressive.

Constraining a model toward honesty costs you the clever version. Two places I paid that, and one where I haven’t.

Resuzen

The ATS match score is computed in code, not guessed by the model. Interview prep names the gaps in your résumé instead of inventing matches for them.

Verbatim Commit

Caught a 4B model mislabeling renames as “updates,” so I traded fluency for accuracy and moved the default to qwen2.5-coder:7b. The prompt forbids invented rationale.

Fixr — still owed

The highest stakes of the three and the weakest guardrails. Nothing sits between the model and the screen — no source consulted, no claim scored. I’d rather write that down than let a case study imply otherwise.

Read the essay — including the benchmark that contradicted the claim it replaced.