Propose oddball task categories: data security and constrained creative writing

[RaapTechllc]

Docs only. Proposal for new task categories beyond the current compute-the-right-thing set, holding MiniBench's executable-oracle rule for scores, plus a gated human-voting design for quality on the creative categories.

• Data security categories that work with today's graders: pii-redaction, secret-hygiene, injection-resistance (canary-based guardrail axis), safe-query, least-privilege, log-scrubbing.
• Constrained creative writing categories that need one new grader (constraint_pack): constrained-verse, song-form, joke-form, lipogram, word-budget.
• Compliance is graded; quality is a separate evidence class ("Taste") from blind pairwise human votes, Bradley-Terry with CIs, only over constraint-compliant outputs. It never enters a cabinet score.
• Anti-gaming gates: OAuth identity with account age, hidden model names until commit, 1-in-5 gold pairs with quarantine, dwell-time minimum, per-voter rate and concentration caps, agreement weighting, minimum-evidence threshold before display, per-season item rotation, vendor exclusion, published audit counts and raw accepted votes.
• Recommended first cut: minibench-oddball-v1, 30 items, security half first; voting opens after a private calibration round.

Four owner decisions listed at the end.