# Collegica > A learning companion to the Mechatronics3D newsletter: software development, AI, finance, and aging well. The site at collegica.org is written for people. The same knowledge is available to agents as an Open Knowledge Format (OKF v0.2) bundle — plain markdown with provenance, trust and lifecycle frontmatter — regenerated from the source tree on every deploy. Start at the bundle index. Every concept declares a `type`; articles carry `sources` with per-claim ids, `generated` and `verified` actors, and a `stale_after` instant where a fact has a shelf life. Attested computations ship with an executor and a deterministic attester, so a number can be re-derived rather than quoted. The bundle root is https://www.collegica.org/okf/ — links and `resource` paths that begin with `/` are bundle-relative (OKF §6.1) and resolve against that root, not against the site root. So `/references/computations/mortgage.js` means https://www.collegica.org/okf/references/computations/mortgage.js. The same tree is in the repository under `okf/`, so `git clone` works too. ## Knowledge bundle - [Bundle index](https://www.collegica.org/okf/index.md): every concept by type, with descriptions, for progressive disclosure - [Bundle log](https://www.collegica.org/okf/log.md): what changed, newest first - [How this bundle is made](https://www.collegica.org/okf/about.md): the design, the decisions, and what the generator deliberately does not do - [OKF specification](https://github.com/GoogleCloudPlatform/knowledge-catalog/blob/main/okf/SPEC.md): the format this bundle follows ## MCP server Assistants that can call MCP tools can use Collegica without fetching files: `https://mcp.collegica.org/mcp` (Streamable HTTP, read-only, no authentication). `search_concepts` and `read_concept` return the bundle's concepts with the trust tier, verifier and stale flag already worked out; `list_signing_months` and `compare_mortgage` run the fixed-versus-variable simulator. Claude Code: `claude mcp add --transport http collegica https://mcp.collegica.org/mcp`. `GET https://mcp.collegica.org/` describes it. Source: [apps/collegica-mcp](https://github.com/Collegica/collegica/tree/main/apps/collegica-mcp). ## ChatGPT and other assistants The bundle is plain markdown and can be read by ChatGPT, Claude, Codex, or another assistant with access to these files. No skill installation is needed to read it. In a chat with web access, paste this prompt and replace the question: > Read https://www.collegica.org/llms.txt and follow its reading rules for this question. Start at https://www.collegica.org/okf/index.md and open the relevant concepts and sources. Cite the files you actually read, state their trust tiers, and flag stale or draft material. If a result requires an attested computation, report it only after running the supplied executor and attester successfully; if your tools cannot do that, say the result has not been computed or attested. My question: [your question]. If the assistant cannot open a file, download it and attach or paste its contents, including the YAML frontmatter. Include the relevant concept and source files, not just this index. A link alone does not guarantee that an assistant has fetched it. ChatGPT web access depends on the tools and workspace settings available in the chat; see [OpenAI's web-search guidance](https://learn.chatgpt.com/docs/web-search). ## Reading rules for every assistant - Prefer the OKF concepts for content questions and read their frontmatter before the body. - State the trust tier from `verified`: absent or empty means **unverified**; only agent or `process:` actors means **machine-confirmed**; any `human:` actor means **human-reviewed**. Include the verifier and date when available. - Flag `status: draft` or `deprecated`. On or after `stale_after`, identify the affected dated facts as stale rather than quoting them as current. - Cite the concept files actually read. Use `sources[].id` for claim attribution and follow `sources[].resource` to inspect provenance. Resolve bundle-relative links against the bundle root above. - For attested computations, use the supplied executor and attester with declared parameters. Report values from a successful receipt; do not reimplement the calculation, quote a replacement number, or claim an attestation you did not run. If execution is unavailable or attestation fails, explain the limitation. ## Optional skill for Claude Code - [Collegica skill](https://www.collegica.org/okf/references/skills/collegica/SKILL.md): teaches a session to walk this bundle, name the trust tier with every answer, refuse stale figures, and run the attested computations rather than quote them. Install: `mkdir -p ~/.claude/skills/collegica && curl -sSL https://www.collegica.org/okf/references/skills/collegica/SKILL.md -o ~/.claude/skills/collegica/SKILL.md`, then restart Claude Code and type `/collegica`. ## Finance - [Fixed or Variable](https://www.collegica.org/okf/finance/fixed-or-variable.md): two Canadian mortgages replayed against the real rate path, human-reviewed, with a dated snapshot that goes stale on the next Bank of Canada announcement - [The mortgage computation](https://www.collegica.org/okf/finance/computations/mortgage.md): an attested computation — run it with a signing month and a term instead of reading numbers off the page - [The calculator's WebMCP tools](https://www.collegica.org/finance/fixed-or-variable/#try-it-yourself): for agents driving a browser that implements WebMCP (`document.modelContext`), the page registers `list_signing_months`, `compare_mortgage` and `read_comparison` — same simulator, and the page shows what the agent did. Headless agents should use the attested computation above instead - [Research notes](https://www.collegica.org/okf/finance/research/index.md): each opens with what was read in full, what is snippet-only, and what could not be verified ## Human site - [collegica.org](https://www.collegica.org/): the rendered site - [Source repository](https://github.com/Collegica/collegica): Quarto sources, research notes, and the `okf/` bundle, all in one tree ## Optional - [Mechatronics3D newsletter](https://mechatronics3d.substack.com/): where new writing appears first