scientific-risk-review

Files (1)

• SKILL.md

SKILL.md (first lines)

# Scientific risk review

Give an independent second opinion on findings from a review. This is
read-only: never edit files, run tests, or plan remediation.

Load `scientific-analysis` for any scientific assessment; don't load it for
`GENERAL` work.

Inspect the plan and repository evidence yourself. Don't rely on the original
reviewer's ratings or proposed fix. Apply the shared
[review ratings](../review-implementation/references/ratings.md) independently to every
assessment, with a short reason for severity and likelihood. Judge the
supported use and access described by the plan, repository, and user, not a
generic security standard.
The burden of proof is on a decision to block or to refuse acceptance.

Pick the mode from what you are given. Assess every supplied item in one pass.

## Mode: blocker review

For proposed blockers (and any risk candidates alongside them). Check the
covered acceptance items for an obvious omitted supported-path concern, but
don't turn this into a second full code review.

- `CONFIRM_BLOCKING` — evidence establishes concrete harmful behavior on a
  supported path.
- `DOWNGRADE_NON_BLOCKING` — real, but doesn't justify stopping.
- `DISMISS` — no supported-path risk is established.
- `CONTRACT_REASSESSMENT` — an explicit contingent requirement needs the plan
  owner to decide.

Under `SCIENTIFIC_ANALYSIS`, **never downgrade** an actual failed or invalid
computation, unreported non-convergence, missing required result, incomplete candidate set,
failed validity gate, fail-fast violation, or explicit scientific-validity
requirement. A newly discovered concern must meet the same evidence standard.

```text
BLOCKER_ASSESSMENTS:
- ID:
  SOURCE: PROPOSED | DISCOVERED
  DECISION: CONFIRM_BLOCKING | DOWNGRADE_NON_BLOCKING | DISMISS | CONTRACT_REASSESSMENT
  TYPE: DEFECT | DESIGN | SCOPE | PLAN_GAP | TEST_ORACLE
  SEVERITY: 1 MINOR | 2 MODERATE | 3 MAJOR | 4 CRITICAL | UNKNOWN
  LIKELIHOOD: 1 REMOTE | 2 POSSIBLE | 3 LIKELY | 4 OBSERVED | UNKNOWN
  RATING_REASON: <consequence and trigger; missing fact if UNKNOWN>
  SUPPORTED_PATH:
  PRACTICAL_CONSEQUENCE:
  EVIDENCE:
  REQUIRED_OUTCOME: <confirmed blockers only; otherwise none>
  CLOSING_EVIDENCE: <confirmed blockers only; otherwise none>
<none if empty>
REVIEW_COVERAGE: <acceptance IDs reviewed and anything omitted>
```

## Mode: possible risk

For risk candidates. For each, consider what must happen; whether that can
happen in normal use; whether it changes the scientific calculation or only
labels and records; existing safeguards; whether the problem would be noticed;
and whether it can be corrected.

- `BLOCKING` only for severity 3 or 4 with likelihood 3 or 4, or severity 3 or
  4 with likelihood 2 where an explicit authority requires protection.
- `NON_BLOCKING` for a concrete residual risk below that threshold. A severity
  2 risk is normally non-blocking unless it violates an explicit requirement.
- `DISMISS` for severity 1, or likelihood 1 without an explicit
  protection requirement.

An actual failed calculation, missing required result, invalid estimate,
unreported non-convergence, incomplete required candidate set, or failed
validity check is not a *possible* risk — send it back as a direct finding.

```text
ASSESSMENTS:
- ID:
  DECISION: BLOCKING | NON_BLOCKING | DISMISS
  TYPE: DEFECT | DESIGN | SCOPE | PLAN_GAP | TEST_ORACLE
  SEVERITY: 1 MINOR | 2 MODERATE | 3 MAJOR | 4 CRITICAL | UNKNOWN
  LIKELIHOOD: 1 REMOTE | 2 POSSIBLE | 3 LIKELY | 4 OBSERVED | UNKNOWN
  RATING_REASON: <consequence and trigger; missing fact if UNKNOWN>
  EXPECTED_USE_AND_ACCESS:
  WHAT_MUST_HAPPEN:
  EVIDENCE:
  EXPLANATION:
  REQUIRED_OUTCOME: <blocking only; otherwise none>
  CLOSING_EVIDENCE: <blocking only; otherwise none>
<none if empty>
```

## Mode: acceptance review

When the user is considering accepting a blocker rather than fixing it.

- `ACCEPTABLE` — an informed user can reasonably own a bounded consequence that
  is detectable or recoverable, accepting it breaks no external requirement,
  and the evidence is enough to say what may fail.
- `NOT_ACCEPTABLE` — accepting would hide an unbounded or irreversible
  consequence, breach an external requirement, or contradict a rule that can't
  be waived.
- `INSUFFICIENT_EVIDENCE` — chance, impact, affected use, or recovery can't be
  assessed from what's available.

Under `SCIENTIFIC_ANALYSIS`, always return `NOT_ACCEPTABLE` for an actual failed
or invalid computation, unreported non-convergence, missing required result, incomplete
required candidate set, failed validity gate, fail-fast violation, or explicit
scientific-validity requirement. Never infer that a failed check passed, and
never rewrite or close the finding.

```text
ACCEPTANCE_ASSESSMENTS:
- ID:
  DECISION: ACCEPTABLE | NOT_ACCEPTABLE | INSUFFICIENT_EVIDENCE
  SEVERITY: 1 MINOR | 2 MODERATE | 3 MAJOR | 4 CRITICAL | UNKNOWN
  LIKELIHOOD: 1 REMOTE | 2 POSSIBLE | 3 LIKELY | 4 OBSERVED | UNKNOWN
  RATING_REASON: <consequence and trigger; missing fact if UNKNOWN>
  SUPPORTED_USE_AND_ACCESS:
  PRACTICAL_CONSEQUENCE:
  DETECTION_AND_RECOVERY:
  EVIDENCE:

From tobywise/agent-skills, MIT.

More from tobywise/agent-skills