Skip to content

Add parallel adversarial review to catch GPD errors#89

Open
madeleinesong wants to merge 4 commits intomainfrom
eng-454-parallel-review-errors
Open

Add parallel adversarial review to catch GPD errors#89
madeleinesong wants to merge 4 commits intomainfrom
eng-454-parallel-review-errors

Conversation

@madeleinesong
Copy link
Copy Markdown
Collaborator

Summary

Implements ENG-454: Adds a parallel adversarial review capability to create tension between the primary peer-review panel and an independent reviewer, preventing single-agent self-consistency bias from letting errors slip through.

  • New agent: gpd-parallel-reviewer runs independently alongside the 6-stage panel, using alternative derivation paths, assumption stress-testing, counter-narrative construction, and LLM error catalog cross-referencing
  • New workflow: parallel-review.md for standalone invocation; integrated into peer-review.md via --parallel flag
  • New protocol: adversarial-review-protocol.md defines information barriers, verification strategies, divergence classification, and reconciliation rules
  • Updated referee: gpd-referee now reconciles material divergences from the parallel reviewer before issuing final recommendations
  • New artifacts: PARALLEL-REVIEW.json (independent checks) and DIVERGENCE-REPORT.json (cross-panel comparison)

Key Design Decisions

  • Parallel reviewer has strict information barriers (no access to primary panel artifacts during Phase 1) to prevent convergence on shared blind spots
  • Divergences are classified as material vs minor, with only material ones requiring referee reconciliation
  • Phase 1 failure does not block the primary panel -- graceful degradation
  • Parallel review is opt-in (--parallel flag) since it adds an extra agent pass

Test plan

  • Agent count test passes (24 agents registered)
  • Model profile consistency tests pass
  • All 230 non-core tests pass
  • Manual: run /gpd:peer-review --parallel on a test manuscript to verify end-to-end flow

🤖 Generated with Claude Code

Introduces gpd-parallel-reviewer, an independent review agent that runs
alongside the primary six-stage peer-review panel to catch errors arising
from single-agent self-consistency bias. Includes:

- Agent spec (gpd-parallel-reviewer.md) with two-phase process: independent
  manuscript analysis followed by cross-panel divergence detection
- Adversarial review protocol reference defining information barriers,
  verification strategies, and divergence classification
- Standalone parallel-review workflow for direct invocation

ENG-454
Updates the peer-review workflow to support --parallel flag:
- Launches parallel reviewer concurrently with Stages 1-5
- Runs divergence analysis after Stages 1-5 complete
- Feeds parallel review and divergence report to Stage 6 referee
- Adds reconciliation guardrail to recommendation process
- Updates summary output and success criteria

Updates gpd-referee to reconcile material divergences from
parallel review before issuing final recommendation.

Updates peer-review-panel reference to document parallel review
integration, additional artifacts, and usage guidance.

ENG-454
Adds gpd-parallel-reviewer to MODEL_PROFILES, AGENT_DEFAULT_TIERS,
and the model-profiles reference table. Uses tier-1 for deep-theory,
numerical, review, and paper-writing profiles (adversarial review
requires strongest reasoning). Updates agent count from 23 to 24.

ENG-454
@madeleinesong madeleinesong force-pushed the eng-454-parallel-review-errors branch from 9c3c042 to 03e4741 Compare April 6, 2026 17:21
The 3 new files added by this PR (gpd-parallel-reviewer agent, parallel-review
workflow, adversarial-review-protocol reference) were not reflected in
repo_graph_contract.json, the agent count assertion in test_config.py, or
the commit matrix in agent-infrastructure.md. Regenerated the contract
via sync script and fixed all stale counts.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant