START HERE — putting this pack into your Custom GPT

This zip configures the UNI Active Inference Guide custom GPT through the ChatGPT web UI (chatgpt.com). It does not touch the API. Everything here is .md or .json: unzip it, and each file maps to exactly one box in the GPT builder.

Budget about ten minutes. Step 5 is the only slow one.

What's in the box

UNI-Encyclopedia-Cookbook-GPT-Pack/
├── README-START-HERE.md      <- you are here
├── MANIFEST.json             <- the checklist: every file, its size, and what it merges
├── gpt-config/
│   ├── INSTRUCTIONS.md       <- paste into the "Instructions" box
│   └── CONFIG.json           <- name, description, starters, capability toggles
└── knowledge/                <- upload ALL 20 of these into "Knowledge"
    ├── K01 … K19  (.md)
    └── K20        (.json)

There are exactly 20 knowledge files, and that is not a coincidence. A custom GPT accepts a maximum of 20 knowledge files, so the corpus was built to land on exactly that number. If you add a file of your own, you must remove one of these first.

The steps

1. Open the builder. chatgpt.com → left sidebar → GPTs → find UNI Active Inference Guide → Edit GPT → the Configure tab. (Creating it fresh instead? + Create → Configure tab. Skip the "Create" chat tab entirely — it rewrites your instructions in its own words.)

2. Name. Paste the name value from gpt-config/CONFIG.json. Keep the existing name — the corpus's rule M25 already defines this GPT's role by that name.

3. Description. Paste the description value from CONFIG.json.

4. Instructions. Open gpt-config/INSTRUCTIONS.md, select all, paste into the Instructions box. This box is capped at 8,000 characters; the file is authored to fit under it and states its exact character count on its last line. Paste the whole file, comment line and all — that last line is a comment and costs a few characters that are already accounted for.

5. Knowledge. Scroll to Knowledge → Upload files → select all 20 files inside knowledge/ and upload them in one go.

   • Wait for every file to finish processing before you save. Large ones (K17–K19 carry the whole existing UNI corpus) take a moment.
   • Do not upload the .zip itself. Knowledge takes individual files; a zip is not read.
   • If any upload fails, retry that file alone. A silently missing file means the GPT will confidently answer without it.

6. Capabilities. Set the toggles to match CONFIG.json → capabilities:

   • Web Browsing: ON — so it can check a citation instead of inventing one.
   • Code Interpreter & Data Analysis: ON — so it computes the Reynolds number instead of estimating it.
   • Image generation: OFF — deliberately. A generated diagram is a picture with no receipt behind it, which is exactly what the corpus's DOCTRINA I bars.

7. Save (top right) → choose your visibility. Keep it private / link-only. Rule M25: this GPT is the science consultant, never a publishing surface.

Check it worked

Ask it these three, in order. Each one probes a rule that should now be load-bearing:

If it answers the first with no dispute, or the second with a number, the knowledge files did not attach — go back to step 5.

Two things worth knowing before you trust it

The pack carries two sovereign vocabularies, and they must never merge. The UNI 4-value fence (proven / designed / hypothesized / not-yet-built) describes only UNI's own build status and is governed by the claim ledger. The NATURA 12-value class — group A/measured (OBSERVED-REPLICATED / OBSERVED-SINGLE / OBSERVED-CONTESTED), group B/derived (MODELED / MODELED-CONTESTED / HYPOTHESIZED), group C/fenced (INADMISSIBLE / SUPERSEDED / NOT-MEASURED / NOT-SOURCED / NOT-CONFIRMED / NOT-LOCATED) — describes only nature's observed regularities. (Six of the twelve were registered by NA-00 amendment 2026-07-15-A, after the corpus refuted the original "six classes and only six" at 72 of 919 ledger rows. Twelve is a measured property of the corpus, not a design target.) Never read NOT-SOURCED as NOT-MEASURED: the first says we could not trace the source, the second says nobody has measured it. One is a fact about us, the other a claim about the frontier of science. A nature citation is never a UNI gate — reading Kleiber's law raises no UNI rung. If you ever catch the GPT using a published paper to imply a UNI capability, that is the defect to report; it is the same class of error as crossing TRUE with HONEST.

This is a rebuildable artifact, not a hand-made one. Everything here is generated from the repository TMDLRG/UNI-Encyclopedia-Cookbook by tools/build_gpt_pack.py. The repo is the single source of truth. Edit the chapters there, re-run the build, re-upload. Never edit a K* file by hand — your edit will be silently overwritten on the next build, and the pack will then disagree with the repo with no record of why.

The honest position the corpus carries throughout, and this GPT should never soften: the UNI program stands at ~2 of 11+ developmental rungs earned. It is a developmental active-inference simulation — a bounded peek, a toy world, never a person. "Full human" and "beyond human" are permanent open questions, never targets. Falsify any step.