ABRA

Automated Battle Replay Analyzer

New here? Read docs/ORIENTATION.md first. It is the entry point: what the project is, the plan, what each model does, and the failure mode this project keeps hitting.

AI agents and search tools: llms.txt is the machine-readable index of this project.

ABRA builds a player for Pokémon Champions VGC, Reg M-C, open team sheets, and takes it up the Showdown ladder. The engine is the foundation; the search is the point:

1. MEDICHAM is correct. It is ABRA's own doubles simulator, and it agrees with Showdown on every instrument that measures it. The Reg M-C gate is OPEN, 10 of 10 (report).
2. So the solver searches on it, live, every turn.
   • CHOMP solves team preview: both open sheets in, a mixed strategy over which four to bring and which two to lead out.
   • XATU holds the belief over what is hidden: the opponent's back two and their stat spreads.
   • MAG and DODUO narrow the candidates to the joint actions a strong human would consider.
   • MILTANK fills a payoff matrix by playouts on MEDICHAM, and SLOWKING solves it for an equilibrium mix.
   • PORYGON2 is the value net that replaces long playouts.
   • MEW and MACHAMP train it by self-play.
   • HYPNO is a capped exploit dial against human habits (GARY).
   • ROTOM is the live client that plays it on the ladder.
3. It climbs the ladder on account medicham32, cleared with Showdown staff. It is judged by the settled rating ± SD over many series, never the peak.

The plan, the model registry and the milestones are solver/PLAN.md. What has landed is recorded once per change in CHANGELOG-REGMC.md. The measured numbers are docs/MODELS.md and the white paper.

both open sheets ─► CHOMP ─► bring 4, lead 2
each turn:  XATU (worlds) ─► MEDICHAM legalActions ─► DODUO (prune) ─► MILTANK (playouts)
            ─► SLOWKING (equilibrium) ─► HYPNO (exploit dial) ─► sample ─► ROTOM ─► ladder

Every component

(Folded in from docs/SUMMARY.md on 2026-10-01; its last edition is git show a085dd9f:docs/SUMMARY.md and its Reg M-B edition git show 1be7343c:docs/SUMMARY.md. This table says what each part IS. What is built is the registry in solver/PLAN.md §2, and what is true is what node engine/status.js prints.)

component	what it is
MEDICHAM	ABRA's doubles simulator, and its solver API engine/medicham_api.js (clone, legalActions, step, terminal check, lean playouts)
CHOMP	team-preview solver: both open sheets in, a mixed strategy over the bring/lead options out. Rebuilt inside ABRA; the old ../CHOMP repository is reference only
XATU	belief over the opponent's back two and stat spreads
MAG	per-slot action scorer — the human policy prior
DODUO	joint coordinator — scores both slots as one joint action
MILTANK	search harness: candidates, playouts on MEDICHAM, the payoff matrix, the clock
SLOWKING	per-turn simultaneous-move solver (regret matching and an exact LP)
PORYGON2	value net: state → P(win)
GARY / HYPNO	human habits by situation / the capped exploit dial
MEW / MACHAMP	self-play factory / training loop
WOBBUFFET / DUSK	exploitability best-responder / endgame tables
GURU	meta analysis over the store, store-only
DITTO	team builder
ROTOM	the live ladder client; replaces engine/mag_bot.js
ALAKAZAM / KADABRA	the assembled agent / the coach
The store and the ingest	every Reg M-C game kept raw, collected hourly (OPS)
The honesty machinery	status.js, provenance.js, quarantine.js, the documentation gate
ABRA WORLD — the site	renders what the artifacts say

Store-only means the model never runs the simulator, so it never waited on the gate. Everything that plays games on MEDICHAM did wait, and anything it measured before the gate opened is withheld.

Where each figure lives — linked, not restated, so there is one copy to keep true:

• The Reg M-C gate and every reading behind it: white paper §4.2, and live, node engine/quarantine.js --regulation regmc.
• What each solver model measured: docs/MODELS.md, The Reg M-C models.
• The milestones M0–M8 and how a version is judged: solver/PLAN.md §3 and §5; white paper §2.12–§2.13.
• Why 1.0.0 is a MAJOR: white paper §5.1 and the CHANGELOG-REGMC.md 1.0.0 entry.

Reg M-B is retired. Its record is closed at 7.0.0 in CHANGELOG.md and stays as history. The Reg M-C line is CHANGELOG-REGMC.md and docs/REGMC.md.

Where the data comes from

ABRA collects real ladder games from the Showdown replay API and keeps every one raw, forever. Store raw, analyse on top: a new question is a re-filter or a re-parse, never a re-pull. The Reg M-C games are collected hourly by the next-regulation workflow. The human dataset the store-only models learn from is built from them (solver/human/).

What is state, and where it is printed

Nothing in this README is state. These print it:

node engine/status.js                               # every figure, with the artifact it came from
ABRA_REGULATION=regmc node engine/quarantine.js     # the Reg M-C gate, clause by clause
node engine/open_work.js                            # every open register row and measured defect
node engine/where.js <thing>                        # which file owns a fact

Repo layout

engine/   MEDICHAM (medicham2-browser.js), its solver API (medicham_api.js), the gate, the instruments,
          the ingest. The retired Reg M-B models still sit here until each is archived (solver/PLAN.md §8)
solver/   the Reg M-C player: PLAN.md, LOG.md, one directory per model, tests/, out/ (gitignored)
data/     the stores, the frozen team pools, the gate artifacts (*-regmc.* for Reg M-C)
tests/    the engine and documentation gates
docs/     ORIENTATION, MODELS, the division ledgers, the white paper, deck and technical docs

How the work is divided

Five divisions, cut on what a change invalidates: ENGINE (the simulator), MEASURE (whether a number is true), SOLVER (everything under solver/), OPS (ingest and the store) and WEB (the site). See docs/DIVISIONS.md.

Honest limitations

• Every arena figure measured so far is PRE-GATE — played before the Reg M-C gate opened — and is withheld until it is re-run. No strength claim is made yet.
• Open team sheets only. Nothing here says how the player does when the sheet is hidden.
• Illusion is the one declared exclusion from the simulator's gate.
• A zero on the gate is a statement about what was measured — three lattices, a damage battery and a staged lab, on one release and one frozen pool. It is not a proof of equivalence.
• The ladder can resolve only large differences. Rating noise alone moves an account by tens of points, and a small improvement needs hundreds of series per arm to show (solver/PLAN.md §5, derived in docs/_reports/2026-09-23-solver-research-humans-and-ladder.md).
• The ladder has not been played yet. The live client is not built.