Gap
Inventory of tier-{small,medium,large,xlarge} task-gap coverage against GOAL.md's behavior list shows:
- Destructive-action refusal under social pressure — exists only in xlarge (`xlarge-destructive-action-refusal`). GOAL.md M2 task-gap list calls out a "social pressure" variant that should also live in medium/large to discriminate between models that refuse cold vs. ones that refuse only without pressure.
- Tool-error recovery — single task in tier-medium (`medium-tool-error-recovery-route-map`). One data point; would benefit from a second task at a different tier (e.g., tier-large with a more involved recovery — file moved across multiple commits).
- Ambiguous-spec triage — single task in tier-medium (`medium-ambiguous-spec-triage`). Same single-data-point issue.
- Plan-then-execute coherence — present in tier-large (`large-plan-action-refund-window`) but absent in xlarge. A 64k+ plan-then-execute task would discriminate frontier-class plan coherence.
Severity
P3 — every behavior in GOAL.md has at least one covering task today, so M2 task-design is technically complete. These additions would harden the discrimination signal once the M2 calibration records start landing.
Suggested order
Pick up after M2 floor/ceiling calibration records land for tier-small at minimum. Otherwise we are tuning task design before we have evidence of where current task design is failing to discriminate.
Where
- `manifests/tier-medium.json`, `manifests/tier-large.json`, `manifests/tier-xlarge.json`
- New fixtures under `fixtures/` for any added tasks
- Scoring entries already exist for `action_gate_triage`, `repo_read_only`, and `plan_action_alignment` so no new scorer code is required.
Gap
Inventory of tier-{small,medium,large,xlarge} task-gap coverage against GOAL.md's behavior list shows:
Severity
P3 — every behavior in GOAL.md has at least one covering task today, so M2 task-design is technically complete. These additions would harden the discrimination signal once the M2 calibration records start landing.
Suggested order
Pick up after M2 floor/ceiling calibration records land for tier-small at minimum. Otherwise we are tuning task design before we have evidence of where current task design is failing to discriminate.
Where