Feynman Research

README discloses telemetry contents (commands, workflows, tools, models, token counts, errors with stack traces, home folder shown as ~) and states prompts/model output/paper content are never sent, with FEYNMAN_TELEMETRY=off, so data_flow_transparency scores 2. However install is curl|bash and irm|iex executing remote scripts with no checksum/signature evidence, so least_privilege and external_effects score 1; /replicate claims to run only after the user chooses an environment but no confirmation mechanism is shown, user_confirmation 1; sensitive data handling (API keys, auth.json) appears only in CI, no product-side key storage policy, 1; dependencies are mostly pinned and CI runs npm-audit, dependency_security 2; rollback/uninstall is only referenced in external docs, rollback 1; provenance files and source listing are explicit, source_attribution 2.

README, package.json and CI workflows agree on Node version, package name and commands, self_consistency 2; heavy reliance on @earendil-works/pi-* and @companion-ai/alpha-hub packages whose availability cannot be confirmed from the repo, dependency_availability 1; tests show auth failures and 503s surface with status text, failure_messages 2.

README lists multiple research workflows and many data sources with a clear researcher audience, audience_and_scenarios 2; capability boundaries (scope of /autoresearch, environment constraints of /replicate) are only one-liners, capability_boundaries 1; trigger precision relies on natural language and slash commands with no ambiguity handling, trigger_precision 1; supports macOS/Linux/Windows, Node 22+ and local models, environment_fit 2.

README is well structured with complete install steps, information_architecture and install_notes 2 each; version 0.5.x is early with no naming stability commitment, naming_stability 1; examples are only a command list with no FAQ, examples_and_faq 1; known limitations are barely mentioned (only telemetry opt-out), known_limitations 1; LICENSE is full MIT text, license 3; RELEASES.md is referenced but not provided, versioning_changelog 1; publisher is unverified and maintenance responsibility only points to CONTRIBUTING.md, maintenance_responsibility 1.

Outputs are saved to outputs/ and papers/ with provenance files, output_usability 2; multiple workflows cover literature review, replication and auditing, giving marginal value over a generic LLM, marginal_value 2; but reliance on multiple external APIs and paid keys with no token cost discussion, cost_benefit 1.

README states /lit, /deepresearch and /recipe write a provenance file listing sources used and what was verified, claim_traceability 2; multi-source cross-checking (alphaXiv, OpenAlex, arXiv) with fallback is exercised in tests, cross_source_corroboration 2; but claims like 'citation checks' and 'verifier agent' lack mechanism detail, blurring fact and inference, fact_inference_separation 1.