Deep research should not rely on one model pass. Use structured disagreement and evidence audit.
Finds papers, repos, specs, negative results, security analyses, and stronger analogues.
Builds source ledger, claim matrix, and pattern matrix.
Proposes CAK-native design candidates.
Attacks assumptions and identifies human-software defaults.
Proposes non-obvious / agent-native alternatives that are not direct copies of human software engineering.
Looks for unsafe execution, supply-chain risks, data exfiltration, prompt injection, approval widening, and self-poisoning.
Defines metrics, minimal experiments, kill criteria, and comparison baselines.
Synthesizes, but cannot introduce unsupported claims.
Reviewers independently inspect the same research packet.
Use when source quality, pattern transfer, or architecture pressure needs independent assessment.
One agent defends a hypothesis, another attacks.
Use when a candidate design is plausible but may hide weak assumptions.
Adversary tries to show why the design will fail.
Use for runtime hooks, skill admission, package trust, policy widening, and anything that may become a security boundary.
Reviewer checks whether claims are actually supported by sources.
Use before an RDR enters decision-ready status.
Reviewer checks whether a borrowed pattern actually transfers to AI agents.
Use when the run borrows from human software engineering, planning systems, papers, or external agent frameworks.
- nitpicker-style multi-review suggests useful patterns: parallel reviewers, actor/critic debate, aggregator/meta-reviewer, repo tools, transcripts, provider diversity.
- CAK should extract these patterns into its own document process rather than depend on a particular tool.
These references are not evidence until inspected and recorded in the source ledger.
debate.md must include:
- participants / model providers if known
- prompts used
- positions
- objections
- concessions
- unresolved disagreements
- judge synthesis
- changes made to the RDR after debate
The judge can:
- compare positions;
- summarize supported claims;
- identify unresolved disagreements;
- request source-ledger updates;
- request new experiments;
- recommend decision status.
The judge cannot:
- introduce unsupported claims;
- hide counterevidence;
- turn seed references into evidence;
- decide architecture before a minimal experiment;
- override the quality gate.