Knowledge Islands model radar
This governance skill maintains a portable, reviewable snapshot of models, executable routes, benchmark applicability, and evidence. Read the model-radar standard before assessing or changing the snapshot. The generated rubric publishes its criteria, the radar snapshot holds current structured state, and the tracked sources govern refresh inputs.
The radar informs decisions; it does not rank models automatically, install agents, change runtime defaults, buy access, provision infrastructure, or transmit private material. ki-acquire pulse discovers public signals, ki-env-tokenomics owns portable model-purpose categories, runtime adapters own effective configuration, ki-work-recurring owns recurring review templates, ki-work next captures approved downstream work, and ki-skills assesses the skill artifact itself.
Operating modes
Invoked as help / -h / ?, emit the generated HELP block and stop. With no recognised mode, emit the same HELP and, only in an interactive session, offer the mode choice.
Mode AUDIT
Run ki repo audit --skill ki-model-radar --repo <repo>. The structured catalogue checks the bundled snapshot without fetching sources or ranking models. Apply its judgment prompts to evidence independence, applicability, uncertainty, and proposed lifecycle movement.
Mode CONFORM
Run AUDIT first. ki repo conform --skill ki-model-radar --repo <repo> --dry-run may regenerate only the readable rubric publication. Snapshot, evidence, and lifecycle findings remain diagnostic or guarded because their correct repair requires authorship or review.
Mode EDUCATE
Run ki repo educate --skill ki-model-radar --repo <repo> to render the catalogue. Explain recommendation, support, retirement, movement, evaluation-unit, and evidence classifications from the standard without treating a benchmark or provider claim as an adoption decision.
Mode REFRESH
REFRESH writes only this skill's canonical files in ki-agent-backpack. When invoked from an installed copy, stop and redirect to that Harness. Read the tracked sources, gather bounded current evidence, reconcile it with the snapshot, and propose reviewed changes. If the last review is still within cadence and no material release or deprecation triggered the run, confirm before forcing an early refresh.
Movement inward requires current identity and lifecycle evidence, materially independent evaluation where available, and local-fit evidence proportional to the recommendation. Preserve counter-evidence and uncertainty. Route any approved consumer change through ki-work next; do not mutate a consumer from this mode.
Notes
- Compare results only within a declared evaluation unit: bare model, provider endpoint, model-agent route, or complete task environment.
- Treat provider performance claims as corroborating evidence, never decisive independent evidence.
- Record exact openness and licence per model variant; do not infer local practicality from weight availability.
- Begin local-fit evaluation with public or synthetic repositories. Private-repository evaluation requires separate authority.