Council of High Intelligence

The security policy describes the product as prompt files and shell scripts and documents local installation, provider detection, key handling, and private vulnerability reporting. This supports good, though not exhaustive, least privilege and data-flow transparency. The checkout action is pinned and CI tools are listed, but the supplied material shows no complete dependency inventory or vulnerability monitoring, so dependency security is not full marks. README documents dry-run, but the materials do not adequately explain uninstall/recovery or user confirmation before each external call, limiting rollback and confirmation. Persona grounding is summarized without its sources, and publisher identity is unverified, so source attribution is weak.

README describes the round protocol and routing, while CI lists lint, roster validation, and a simulation checklist; the security policy also defines latest-release support. This provides some evidence of internal consistency. Detection covers local CLIs and some API paths, but the materials do not establish service availability or complete fallback behavior, so dependency availability scores low. README says provider failure is reported before fallback to the native host, supporting failure-message credit; no complete error catalog is shown.

The documentation distinguishes full, quick, duo, named triads, member selection, and configuration, and explains suitable decision types and smaller alternatives. This gives thorough audience and scenario guidance. README and the security policy identify supported hosts and provider paths, but coverage of boundaries, compatibility limits, and error cases is incomplete. Examples and dry-route demonstrate invocation, but the supplied materials do not establish ambiguity handling or fully consistent behavior across environments, so those criteria are not full marks.

The repository map and separation into protocol mirrors, agents, configs, demos, scripts, and assets provide strong information architecture. Installation flags, target directories, restart guidance, dry-run, direct examples, and advice on when not to use a council are clearly documented. The MIT license text is present. README points to release history and a maintainer, and the security policy defines supported versions, but the supplied materials do not show specific changelog entries or a clear maintenance update commitment. Cross-host naming stability is also mainly asserted in the documentation, so those criteria retain deductions.

The protocol calls for independent analysis, cross-examination, final positions, and synthesis, with verdicts that preserve unresolved questions, dissent, kill criteria, and a next step. This supports usable outputs. Multi-perspective analysis and provider routing may add value for high-stakes decisions with incomplete evidence, but the marginal benefit is asserted in documentation rather than demonstrated by outcome evidence in the supplied material. The docs recommend direct answers or experiments for cheap, reversible choices; cost, latency, and benefit measurements are absent, so cost-benefit is not full marks.

README links claims to repository protocols, configurations, examples, and validation scripts; CI lists corresponding static checks, providing some traceability and cross-file support. No execution results or check outputs are supplied, so configured checks cannot be treated as having run. The evidence labels explicitly distinguish FACT, INFERENCE, ASSUMPTION, and UNKNOWN and call for preserving unknowns; this design is directly supported and merits full marks.