Skip to content

Make render review genuinely independent without leaking internal agent details #9

Description

@tomvaillant

Problem

The same agent currently proposes the visual and performs the “independent” render review. The documentation acknowledges the residual rubber-stamp risk, but the review still depends on the reasoning that authored the spec.

This is especially risky for semantic defects: a misleading title, inappropriate chart type, omitted takeaway clause, or invented source can look correct to the agent that created it.

@rmdms

Suggested solution

Add a separate, hidden critique pass with a fresh context and structured inputs. It may use a separate review model/agent or another isolation mechanism, but internal identifiers and orchestration plumbing must never appear in journalist-facing output.

The reviewer should receive only the evidence needed to review:

  • Confirmed takeaway and accepted proposal.
  • Article/data or bounded evidence extracts with provenance.
  • Actual render(s) and target viewport metadata.
  • Interaction-test results.
  • Shared review rubric and source policy.

It should return validated structured findings with criterion, severity, evidence, and confidence—not free-form internal reasoning. The host orchestration records the findings and presents only newsroom-facing summaries.

Independence must be observable in the run manifest: reviewer mode/version, input hashes, output hash, and whether a fallback self-review was used. If the independent reviewer is unavailable, do not silently claim independence; record self-review and apply the configured policy.

Define privacy/data-retention requirements before sending unpublished reporting to any additional model or service. Support a local/same-provider isolation mode for sensitive newsrooms.

Acceptance criteria

  • The critique pass runs in a fresh context that did not author the proposal.
  • Inputs and outputs are structured, hashed, and recorded without exposing private chain-of-thought.
  • Journalist-facing messages contain findings only, never agent IDs or internal task plumbing.
  • Unavailable independent review is explicitly recorded and handled by policy.
  • Re-production triggers a fresh critique of the new artifact.
  • Tests include seeded semantic defects that the critique must catch, privacy/redaction cases, and no-internal-ID output checks.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions