Physical-Learning Commons

The meta-project of the physical-learning-substrates portfolio: it records, adapts, and synergizes the common lessons of 01–05 and the self-assembly parent line. It is simultaneously a knowledge commons and a code commons, joined by a human-triggered synergy loop.

This is project 00. The portfolio's throughline question — asked of every project — is who computes the update? (the substrate's own physics, or a CPU?).

Three pillars

Pillar	Path	What it is
Code commons	commons/	The substrate-agnostic harness (solver, coupled-learning rule, audit hook). Consumed by port, not import; populated by promotion from siblings.
Knowledge commons	synthesis/	Living, hand-curated lessons ledger, verdict registry, pattern catalog, glossary, hypotheses.
Harvest tooling	harvest/	A lean scanner that surfaces promotion candidates from the co-located sibling repos.

BULLETIN.md is the advertise-down digest each sibling reads at session start.

The synergy loop (human-triggered sweeps)

1. Harvest (up) — either python harvest/harvest.py scans the co-located siblings for findings docs and candidate modules, or a sibling stages a promotion-branch handoff (promote/<NN>-<topic> + PROMOTION-<NN>.md, local/unpushed) that the resident 00 agent reviews → merges (--no-ff) → advertises → deletes.
2. Curate — a human + agent decide what is a real lesson or reusable asset; lessons go to synthesis/, code is generalized into commons/ with provenance.
3. Advertise (down) — regenerate BULLETIN.md.
4. Adopt — each sibling reads the bulletin and ports the module or applies the lesson.

Write boundary. Advertising-down is a pull model — 00 publishes BULLETIN.md and each sibling ports what's recommended; the one thing 00 writes into a sibling repo is its CLAUDE.md commons pointer. Contributions flow up via the promotion-branch handoff above (a sibling stages a promote/* branch in this repo; the 00 agent reviews and merges). 00 never edits a sibling's code or findings.

Status

Active. Six promotion sweeps have landed — most commons/ modules are now Tier 1 canonical (filled from siblings): constitutive / network / solver / coupled_learning / tasks / audit (two-axis), the elastic trio (spring_network / mechanical_solver / mechanical_learning), and energy.py — the first gradient-flow substrate (from 05). Remaining: baselines is Tier 2 (one open reconcile — merge 02's two-ceiling ceilings.py), and harness is still a stub. Every sibling now has a result (self-assembly BOUNDARY + Stage-2b interim; 01 A-PASS/B-FORECLOSE; 02/03/04/05 PASS). See BULLETIN.md for the per-sibling scoreboard and what's newly available.

Running

pip install -r requirements.txt
python -m pytest -q        # 84 passed, 1 skipped (only `harness` skips — it's the last stub)
python harvest/harvest.py  # run a harvest sweep (prints a table + writes JSON)

Method discipline

Compute first; foreclosure is a valid outcome; observation kept separate from interpretation; pre-registered verdicts. See ../README.md for portfolio conventions, SEED.md for this project's brief, and docs/superpowers/specs/ for the design.