Skip to content
NGL321Public

About

This repository contains my exploration, learning, and research into the idea of cognition as a graph of small inference networks operating in independent metric spaces

Resources

Contributing

Stars

0 stars

Watchers

0 watching

Forks

Repository files navigation

Mosaic

A long-running research programme in computational cognitive science, conducted openly with LLM acceleration.

Mosaic studies cognition as a heterogeneous network of inference engines whose schemas each carry their own metric space, and the transport between them — of which the primary observable is where transport fails. The resolved vocabulary for that sentence lives in CONTEXT.md; every term there names its source, and marks where Mosaic's usage departs from it.

The programme is framed on Lakatos: a Hard Core that is not revised, a Protective Belt of falsifiable claims that is, and a Positive Heuristic governing how belt rungs are added and retired. Work is planned and decided in the issue tracker, indexed from the wayfinder map.


Where the human is

Mosaic is machine-accelerated and says so. The obvious objection — how much of this did the researcher actually do? — is answered structurally rather than by assertion, so that it can be checked rather than believed.

Two rules, in PROTOCOL.md:

Custody. Agents write the record; the human writes the beliefs. Research documents, notebook entries, and Run output are agent-writable. CONTEXT.md and the charter are human-only — agents propose exact text, the human applies it. So:

git log --format='%an' -- CONTEXT.md   →  one name

Warrant. Evidence lands freely, carrying its provenance tier. But a claim entering the Protective Belt requires the researcher to write its falsifier, in their own words, and a revision to the Hard Core requires a public, human-authored work product. Those drafts are grilled on the pull request by an agent that did not produce the work. Failing that gate does not reject the merge — it starts a teach cycle, and the fact that it fired is recorded.

Custody without warrant is transcription. Warrant without custody is a signature on someone else's page. The pair is the claim.


Reading the version

Two independently versioned tracks, tagged research-v* and tooling-v*.

On the research track, the number is Lakatosian rather than sized by effort:

  • MAJOR — the Hard Core changed. This is an admission that the programme was refounded. 0.x is pre-charter.
  • MINOR — a Protective Belt rung was added, retired, or falsified.
  • PATCH — the record accreted: evidence, vocabulary, corrections.

Bump and diff size are deliberately uncorrelated. A 457-line research document is a PATCH; a two-line edit retiring a belt rung is a MINOR. The version tells you the one thing the diff cannot: whether the programme's commitments moved, or its evidence.

Which also means the version is honest about how young this is. Mosaic currently has language and evidence and no belt rungs at all — so it sits at 0.0.x, and will until the first falsifiable claim is named.


Layout

CONTEXT.md Resolved vocabulary. Human-authored. Read before using any of its terms.
PROTOCOL.md Branches, commit types, custody and warrant, review, collisions.
docs/DATA-PROTOCOL.md What is stored, what is regenerated, and where run artifacts live. Drive-first.
docs/research/ Agent-produced research output. Record, not belief.
.claude/skills/ The skills that run the sessions — wayfinder, grilling, research, teach.

Repository structure beyond this is not yet settled.


Licensing

Mosaic is a research record with software attached, not a software project with docs attached, so it carries two licences split by artifact kind — code under MIT, the prose record under CC BY 4.0. Take the code freely; if you carry the ideas, name where they came from.

This table is the single place the mapping lives.

Copyright © 2026 Noah Litov. Attribute the CC BY 4.0 material to Noah Litov, Mosaic, with a link back to this repository and to the licence.

Path Licence
CHARTER.md, CONTEXT.md, PROTOCOL.md, README.md, CONTRIBUTING.md CC BY 4.0 (LICENSE-DOCS)
notebook/, docs/, curriculum/ CC BY 4.0 (LICENSE-DOCS)
.claude/skills/, .agents/skills/ Third-party — neither. Vendored from mattpocock/skills per skills-lock.json, under their upstream licence.
CITATION.cff MIT — the default, and assigned rather than inherited: it is machine-readable metadata about the record, not part of it, and a reference manager should copy it freely.
everything else, inquiries/ included in full MIT (LICENSE) — the default

LICENSE and LICENSE-DOCS are licence texts rather than licensed material: MIT's is reproduced under its own terms, and Creative Commons dedicates its licence texts to the public domain under CC0.

The directory wins, and MIT is the default: a file nobody has classified fails safe toward permissive rather than into unlicensed limbo. inquiries/ is MIT in full, prose write-ups included.

Three riders:

  • The licence follows the whole file. A code block inside PROTOCOL.md is CC BY; a prose docstring in tools/ is MIT. There is no sub-file licensing.
  • Generated files inherit their directory. curriculum/open.md is CC BY whatever emitted it.
  • A new top-level file or directory is assigned a licence in the pull request that creates it — the one rule here with teeth. It reaches files as well as directories because a root-level file has no directory to inherit from, so the default is all that catches it, and the default is wrong for prose: CONTRIBUTING.md is in the table above for exactly that reason.

Contributions are inbound = outbound: you licence each file you touch under whatever that file's path already carries. See CONTRIBUTING.md for the checklist and PROTOCOL.md §8 for the reasoning.

Citing Mosaic

CITATION.cff at the root. A citation must pin a version — a research-vX.Y.Z tag, or a commit sha for anything predating one. main is the programme's current state of belief, not its published state, so an unversioned citation names a target designed to move.

About

This repository contains my exploration, learning, and research into the idea of cognition as a graph of small inference networks operating in independent metric spaces

Resources

Contributing

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages