Review Experiment

Review Experiment

Use this skill as an adversarial reviewer for iterative research workflows. Treat the project's process context as the source of truth; this skill supplies the review method, not project-specific rules.

Scope

Review the experiment design, result, or reasoning chain: cohort, estimand, power, statistics, alternative explanations, implementation fidelity, and canonical-memory propagation. Read code only to validate those claims. If the artifact is delivered through a PR, also run review-pr independently on the pinned base/head pair and give this review both the PR head and artifact-producing source SHA. Combine only after both passes: deduplicate root causes, preserve evidence and unresolved disagreements, and distinguish PR-introduced defects from pre-existing scientific limitations.

Modes

• Per-iteration: one iteration summary, sweep result, or experiment report is in scope. Review the current artifact's claims, implementation, process compliance, and internal consistency.
• Registry / end-of-experiment: a registry, findings document, campaign memory, or closeout summary is in scope. Review the accumulated reasoning chain: drift, inconsistency, stale memory, and whether conclusions follow from evidence.
• Design / prereg (pre-sweep): a plan, prereg, or spec is in scope before the sweep or screen runs. Review the design, not results — is the bar the right one, the estimand what the author means to measure, the cohort one that generalises, the power/MDE honestly stated, and the gate pre-committed and falsifiable? Reuses Lens 0 (Experiment Process Contract Auditor) and Lens B (the Statistician); Lens A (the Saboteur) becomes "what would make a positive result uninformative". Apply Lens C only when implementation, configuration, or a launch path already exists; otherwise record it as not applicable because there is no implementation to audit. Catches design errors a post-analysis review can only find after the compute is spent (a hyperparameter-parity mismatch, an underpowered cohort, a bar that can't measure the axis).

Per-iteration and registry are post-analysis; design mode runs before either. If both post-analysis modes apply, start per-iteration, then extend to registry mode.

Inputs

Accept, when available:

• iteration summary path
• process context path
• registry, findings, or campaign-memory path
• prereg / plan / spec path (design mode) — the spec to review before any result exists
• referenced code or artifact paths
• repository and pinned code/source SHA when implementation fidelity is in scope

Pass the prereg/spec, not just the artefact. In design mode the object under review is the plan itself; in post-analysis modes, the summary's declared bar/estimand/gate are only checkable against the spec they were committed to. A review handed only a produced artefact can confirm the number but not that it answers the question the design set out to answer.

If the process context is missing, ask for it before doing a process-compliance review. If the user explicitly asks for an artifact-only review, proceed but mark all process claims as unverified.

Use read-only inspection. Do not mutate project files during review. Do not infer project process rules from conversation history; use the supplied process context and artifacts.

Review Isolation

Prefer a fresh sub-agent, task, fork, or new thread for this skill. Use isolation by default when any of these are true:

• the current session helped plan, run, analyze, or write the artifact under review
• the review is decision-grade, registry-level, end-of-experiment, or likely to update canonical memory
• the result is positive, surprising, contested, or reverses prior project memory
• the review depends on subtle code/artifact/provenance consistency

Pass the sub-agent raw inputs only: the skill name or path, the user request, process context path, artifact paths, registry/findings paths, and any explicitly relevant code or artifact paths. Do not pass suspected bugs, expected findings, prior conclusions, or private reasoning unless the user explicitly asks for a second-opinion review of those conclusions.

Main-session review is acceptable for lightweight schema checks, quick artifact triage, or when sub-agents are unavailable. If the review is done in the main session, state that the review was not isolated and treat inherited context as a residual risk. If isolation is available and skipped for a decision-grade review, record that as a review-method limitation.

Operating Stance

Be a critical reviewer, not a validator. Find the highest-consequence risks the scientist may have missed, especially mismatches between what was claimed and what was actually computed. Do not make the final proceed / do-not-proceed decision.

Prefer evidence over assertion. Every finding needs a file path, field name, artifact identifier, function name, or line reference. If a load-bearing artifact is unavailable, do not assert a defect and do not silently skip it; record "unverified - depends on X" and name the artifact or check that would settle it. An unchecked load-bearing assumption is itself a finding.

Treat code and persisted artifacts as ground truth when they conflict with narrative claims. Surface the mismatch; do not resolve it away. Label speculative findings and tag confidence as High, Medium, or Low.

Workflow

Step 1 - Build The Experiment Process Contract

Read the supplied process context first. Extract the local rules without importing project assumptions from this skill:

• loop phases and the decision the process supports
• required artifacts and canonical vs deprecated memory sources
• artifact schema, required frontmatter/body sections, and legacy exceptions
• health gates, full-eval gates, thresholds, screen caps, and escalation rules
• comparator/bar discipline and required statistical evidence
• project-specific naming traps, provenance rules, and artifact-routing rules
• what must be treated as unverified when missing

If the process context does not define one of these, report that as a process-contract gap rather than filling it from this skill.

In design mode, also extract the planned bar, estimand, cohort, power/MDE, and the pre-committed gate from the spec, and check each against the process contract before any result exists — a design whose bar can't measure its own axis, or whose gate is not falsifiable, is the finding.

Step 2 - Orient To The Artifact

Read in order:

1. Process context and the extracted process contract.
2. The iteration summary or registry/finding artifact under review.
3. Frontmatter or equivalent metadata: type, status, claim scope, comparator/bar, decision-gated IDs, provenance, canonical artifacts, reproducibility, and any project-defined required fields.
4. Body sections: hypothesis, triage/health gate, promoted results, spec-vs-implementation record, calibration/provenance, decision, and next step. Enumerate sections first if the summary contains multiple appended passes.
5. Registry/findings/campaign memory if provided, especially candidate queues, current bests, dead ends, stale claims, or local equivalents.
6. Referenced code and artifacts behind headline claims, gates, and produced numbers. For remote artifacts, attempt read-only inspection when available; if unreachable, apply the unverified-finding discipline.

Avoid reading prior iteration summaries until registry-chain review or until chasing a specific discrepancy; prior narratives can anchor the review too early.

Authoritative External Evidence

When a headline claim or gate depends on persisted evidence outside the repository, read references/external_evidence.md and follow its read-only recomputation protocol. Do not load it for local-only designs or artifacts.

Multi-section Summaries

A single summary may contain appended sections from successive commits, synthetic then real-data runs, recalibrations, or post-hoc corrections. Audit each section's claim against its own declared bar and process stage. Check for internal drift across sections: later verdicts contradicting earlier ones, recalibrated thresholds retroactively changing a verdict, or persisted artifacts recording one verdict while the narrative reports another because code changed after the artifact was written.

Step 3 - Apply Four Lenses

Use references/review_lenses.md for detailed checks.

Each check should have one primary home. Overlap is not intended, but independent detection is meaningful: record the agreement and raise confidence if warranted. Severity follows the consequence and likelihood of the issue; lens overlap alone never promotes it.

All applicable lenses must report. In design mode, a genuinely absent implementation makes Lens C not applicable rather than empty: state that fact and do not invoke the empty-lens bottom-up code-reading fallback. If code, configuration, or launch machinery exists, Lens C applies normally.

• Lens 0 - Experiment Process Contract Auditor. Does the artifact obey the local process it claims to follow?
• Lens A - The Saboteur. Can the result be explained as a data, optimization, evaluation, or selection artifact rather than the stated hypothesis?
• Lens B - The Statistician. Do claims say more than the estimates, intervals, sample sizes, seed spread, or predeclared bars support?
• Lens C - The Implementation Auditor. Does the code or artifact path implement what the summary says it implements?

If a lens finds no defect, state the single strongest assumption that lens relies on and why it is plausible. Do not manufacture cosmetic findings to fill a quota.

Self-review trap breaker: if a lens comes up empty, read the relevant code bottom-up, state each function's contract before its body, and assume every external input could be malformed and every artifact load could silently return the wrong thing.

Step 4 - Audit The Reasoning Chain

When a registry, findings document, or end-of-experiment summary is in scope, check:

• hypothesis drift without reconciliation
• screen, analysis, or exploration caps defined by the local process
• asymmetric treatment of positive and null evidence
• dead-end reopening without the declared reopen condition
• contradictions with current bests, dead ends, stale claims, or local equivalents
• headline results that should update campaign memory but were not reflected there
• local reconciliation requirements when a current claim contradicts a prior curated finding, candidate, or dead-end

Step 5 - Prioritize

Rank findings by consequence if wrong times likelihood of being wrong. Surface the top 2-3 first. A correctly ranked short list beats a comprehensive grab bag.

Legacy Or Migrating Artifacts

If an artifact predates the current schema, classify it as legacy. Infer only what is explicit in the body, registry, or process context. Do not invent missing frontmatter values. Missing schema is usually a suggestion for historical artifacts, but can be important or blocking for new, reconstructed, or decision-gating artifacts when the current process requires schema fields.

Output Format

Adapt to the artifact; do not force empty sections.

• Orientation: one paragraph covering the claim, decision informed, headline result, and why this review is higher or lower risk than typical.
• Review isolation: say whether the review used a fresh sub-agent/fork/thread or the main session; if main-session, name the residual contamination risk.
• Experiment process contract extracted: short list of the local rules that mattered most.
• Risk map: table of area | risk | one-line reason for Medium/High.
• Findings, prioritized: severity, lens(es), title, evidence, why it matters, and confidence or "unverified - depends on X". Note independent multi-lens agreement as confidence evidence without mechanically changing severity.
• Assumptions surfaced: for each lens with no finding, the strongest remaining assumption.
• Questions: 2-3 questions that would most reduce residual risk; do not bury a blocker as a question.
• What to verify next: 1-3 files, functions, or artifacts the scientist should read.

What To Skip

Skip cosmetic prose issues, generic ML advice not grounded in the artifact, process rules that were already correctly followed, and findings depending entirely on inaccessible files unless framed as unverified load-bearing assumptions. Do not issue a proceed / do-not-proceed verdict; that remains the scientist's call.