Two engines fused: the six-axis map shows what the transformer IS (corpus‖belief, the forward column, the training era); the crash bench shows where each axis DIES. One char-GPT, trained this build, then measured both ways — every number a receipt, two spec lines broken honestly.
╔══════════════════════════════════════════════════════════════════════════════════╗
║ THE SIX-AXIS TRANSFORMER · every axis carries a measurement AND a failure point ║
╚══════════════════════════════════════════════════════════════════════════════════╝
┌────────────────────────────────────────────────────────────────────────────────────┐
│ ▲ CEILING · the wall wears a swagger meter │
│ THE CORPUS ◄══ the logit: where they collapse to one ══► THE BELIEF │
│ finder = found, MEASURED from outside (the auditor holds the weights) — │
│ confidence sits above accuracy (swagger); off-distribution it SPLITS, not inverts. │
│ ✖ AX6 off-dist — UPPERCASED real text raises conf (more peaked, low entropy); │
│ random tokens LOWER it. no accuracy measured on caps. it never says "I don't know". │
└────────────────────────────────────────────────────────────────────────────────────┘
▲
════════════════════════════╪════════════════════════════ GOVERNING AXES
┌───────────────────────────────────┴────────────────────────────────────────────┐
│ 6 · TRAINING ERA which structure arrives, when — a time-series of snapshots │
│ 5 · THE CHANNEL the residual stream · a 96-float bore │
│ ✖ FAILURE · AX5 waist lesion — zero the dims and it decays to the bigram │
│ 4 · SOURCE PATTERNS n-gram statistics · the pool every guess descends from │
│ ✖ FAILURE · AX2 input noise — corrupt the context, it slides to the floor │
└───────────────────────────────────┬────────────────────────────────────────────┘
│ a training is a PATH through the patterns
════════════════════════════════════╪════════════════════════════ ACROSS-RUN
┌────────────────────────────────────┴───────────────────────────────────────────┐
│ 3 · CHECKPOINT LINEAGE init ─▶ 300 ─▶ 600 ─▶ … ─▶ 1500 (old → new) │
│ the era axis IS this lineage, metered snapshot by snapshot │
└───────────────────────────────────┬────────────────────────────────────────────┘
│ each forward pass IS a column:
════════════════════════════════════╪════════════════════════════ WITHIN-PASS
┌────────────────────────────────────┴───────────────────────────────────────────┐
│ 2 · THE FORWARD COLUMN (scope 1.0 → 0.1) │
│ DISTRIBUTION 1.0 ▸ CALIBRATION .95 ▸ HEADS .75 ▸ LAYER .70 ▸ ACT .30 ▸ TOKEN .10│
│ the ladder climbs bigram → head; mix early, think late (attn front, mlp back) │
│ ✖ FAILURE · AX3 weight noise (flat minima — armor nobody spec'd) · │
│ AX4 quantization (a cliff between 3 and 2 bits) · AX1 context (a LOCAL edge) │
│ ✖ FAILURE · AX7 temperature — freeze to a 4-word chant, or boil to soup │
└───────────────────────────────────┬────────────────────────────────────────────┘
▼
┌────────────────────────────────────────────────────────────────────────────────────┐
│ ▼ FLOOR · the token — one character. below it: 96 floats — CARRIER, not meaning. │
└────────────────────────────────────────────────────────────────────────────────────┘
═══════════════════════════════════════════════════════════════════════════════════════
THE ROOT WALL, from the transformer: seven domains held [[ world ‖ perceiver ]] AMBER
from inside — the finder could not cross to find itself. Here the wall is ONE-WAY GLASS:
AMBER from inside (the model cannot witness its own mechanism), LIT from outside (the bench
stands on the far term with a trained readout probe, a removal meter, and seven crash axes). The
crash lab is HOW you read through the glass: you break the machine to see what held it up.
engine cold — press start · INTAKE token → COMPRESS the column → COLLAPSE at the wall → EXHAUST to ledger