ADR 0021 — Computational biology is in scope; exclusions become liftable

Context

karani has excluded computational-biology and bioinformatics roles since the first pre-filter. The reasoning was sound at the time: the candidate’s NeoRx work is biomedical, the titles attract wet-lab and lab-adjacent postings, and a hunt aimed at staff software engineering did not want “Computational Biologist” filling the shortlist.

The exclusion was enforced in four places — built-in excluded_title_terms, [positioning] excluded_domains rendered into every prompt, the qualification rules, and CLAUDE.md. Config could only ever append exclusions, so there was no way to lift one.

The candidate now wants research-engineering roles on both sides of their background. That is not a reversal of judgement so much as a correction: the NeoRx pipeline applied Pearl’s do-calculus with reinforcement learning to drug-resistance prediction. That is systems and modelling work, and the labs hiring for it — Recursion, insitro, Isomorphic, Deep Genomics — are hiring research engineers, not bench scientists.

Decision

1. [profile] allow_titles lifts terms out of the built-in exclusions. The defaults stay as they are: they encode a sensible generic software-engineering hunt, which is the right default for a project anyone can pip install. A candidate whose own background covers an excluded domain declares it, and the loader subtracts those terms before appending exclude_titles.

   One trap this surfaced: allowing "computational biologist" is not enough while "biologist" remains excluded, because the shorter term matches inside the longer title. The allowlist has to name every term that matches, and a test pins the real titles rather than the terms.

2. Wet-lab terms stay excluded. medicinal chemist and chemistry are not lifted: the thesis is building systems and models, not bench work. “Computational Biologist” now passes; “Medicinal Chemist” still drops.

3. excluded_domains may be empty, and the persona then omits the “roles in X are OUT” sentence entirely rather than rendering it against an empty string.

4. A research-engineering target tier (targets.py Tier 3b): AI labs and computational-biology outfits, on the ATSes karani already reads. Deliberately a curated target list and not a web scraper — these employers post to Greenhouse, Lever and Ashby, so scraping would fight bot detection for strictly worse data.

Consequences

• The shortlist will include comp-bio research engineering. If it fills with wet-lab postings the allowlist is too broad — narrow it rather than restoring a blanket exclusion.
• Prompt versions move (qual-v7) because the persona text changes; old and new cohorts stay distinguishable by prompt_version.
• Any pip install karani user keeps the generic defaults, and now has a supported way to hunt in a domain the defaults exclude. The feature is general; this candidate is the first case.
• Rows already judged under the exclusion are not re-qualified (that keys on content_hash), so previously-dropped comp-bio postings only resurface when re-ingested or via karani refilter.