# alephllm-mini-beatrix-training — Technical Companion (the life record)

Companion to *Raising Beatrix: A Byte-Level Model's Measured Childhood*
(AbstractPhil with Claude Fable 5 & Claude Opus 5, August 2026). This
repository is the primary artifact: every checkpoint of every era —
including the checkpoints a rewind discarded — every arm (refuted arms
included), every exam and boundary report, and the manifest that makes
the run resumable from any machine. Stack-side detail lives in the
[AlephLLM companion](https://github.com/AbstractEyes/alephllm);
arm-system detail in
[amoe-lora](https://github.com/AbstractEyes/amoe-lora).

## Layout

```
mini-beatrix-1/
  manifest.json                  # phases, checkpoints, token accounting
  checkpoints/step_XXXXXXXX.safetensors   # bf16 weights (72 with val gauge)
  checkpoints/fp8/               # e4m3 ship-only copies
  resume/latest.pt               # optimizer + stream state (resume-first)
  runs/                          # TensorBoard event files
  arms/                          # chat arms (provenance-matched to cores)
  arms/night/  arms/day1/  arms/bdist/    # the arm-program waves
  reports/curriculum/            # baseline, gate_*, boundary_* JSON
  reports/arms/                  # wave ledgers
  reports/lexicon/               # lexicon census editions
```

## The eras (all on one gauge: fineweb-edu holdout bpb)

| era | steps | bytes | val bpb | note |
|---|---|---|---|---|
| wikitext warmup | 0 → ~1k | 0.30B | 1.32 → ~1.2 | |
| fineweb main+extended | → 51,882 | 15.30B | **1.033** | the LOCK; core of record |
| anneal (contaminated) | → 58,664 | 17.30B | 1.045 | chat template + identity leaked into core; law recorded |
| head fold | 58,664 | — | 1.0479 (retro) | γ folded into W_s; semantic no-op 2.4e-07 |
| curriculum S0–S2 | → 66,803 | +2.4B | 1.13–1.14 | |
| S3 first attempt | (discarded) | ~+1.2B | **1.71–2.05** | CONVICTED: template colonization; checkpoints retained |
| REWIND | 66,803 | — | — | resume state restored; corrected mixes |
| S3–S6 corrected | → 81,047 | | 1.14 → 1.17 | logic/arith/causal/learning stages |
| S7 spiral | → 85,795 | | **1.111** | interleaved review repays the drift |
| S8 articulation | → 88,508 | 26.102B | 1.121 | graduation |

**Benchmark-best checkpoint: 83,903** (mid-spiral; 270-item exam mean
~0.43 vs graduate ~0.39). The graduate is the *continuation* core; the
mid-spiral checkpoint is the *evaluation* core. Both public.

## The exam record (`reports/curriculum/`)

- `baseline.json` — the 45-item era batteries on both t0 cores (51,882
  and 58,664): the pre-curriculum report card, including the anneal's
  measured probe cost (clean core beat annealed on 4 of 9 suites).
- `boundary_curriculum_s*.json` — automatic per-stage instruments:
  probe suites, fineweb gauge, hub/bank toggle deltas, head ‖W_s‖.
  Boundaries S0–S6 used the era batteries; S7–S8 the 270-item exams.
- `gate_step*.json` — on-demand exam runs of the newest checkpoint
  (the notebook's instrument gate resolves the newest checkpoint at run
  time; `baseline.json` is reserved for the historical t0 pair).

Headline exam-era facts: in-template logic/arithmetic 1.00 collapsed to
~chance on surface-disjoint items (both cores — template competence,
not transfer); S2 difference/category is the one stage that
generalized (0.50 → 0.77); the spiral produced the only broad rise
(~0.35 → ~0.43 mean) and the final blocked stage receded part of it
(graduation ~0.39).

## The arm record (`arms/`, `reports/arms/`)

Chat arms: `chat-s1200@step51882` (arm A — the space's exhibit pair
with the clean lock) and `chat-s1200@step58664` (arm B — trained on the
annealed core; its near-zero deltas are the anneal-confound
measurement). Night/day waves: stop arms (`stopnl` working, `stop` NUL
control refuted), capability arms (d5 default/wide, sub3 ×4 variants
all refuted, defs, identity, poly generalist), each with ledgers.
`arms/bdist/` receives the distillation lane (pre-registered; teacher
gate T0 = Qwen2.5-3B at 0.96 after 1.5B rejected at 0.86; pushforward
gate A0 = 1.0000).

Every refuted anchor ships. The ledger census equals the anchor census
(weights-completeness law).

## Gauges, defined once

- **val bpb**: mean cross-entropy in bits/byte on the fineweb-edu
  reserved head slice (fixed, unshuffled; measured fact: 100%
  crawl-year 2013), identical across all eras (gauge continuity law).
- **toggle ledger**: bpb with one mechanism disabled minus full bpb —
  the causal contribution of banks / hubs / head aleph.
- **exam accuracy**: byte-NLL argmin over 3 options, 270 items,
  surface-disjoint from every training generator; family sub-scores
  including the reserved holdout families (rule-chain depth-5,
  three-digit subtraction) that no training row has ever contained.
- **strict generative accuracy** (arm program): order-aware full-
  sequence match — substring matching measured as inflating (1.00 vs
  0.70) and retired from verdicts.

## Provenance discipline

The chat space serves only provenance-matched pairs (numeric step
match; anchor `base_model_id` verified at load) after an adversarial
review demonstrated silent foreign-arm attachment. The manifest is
embedded in the resume payload (atomic consistency); crashes never
overwrite `resume/latest.pt`; interrupted sessions upload before exit.
The rewind's discarded checkpoints remain public deliberately: the
conviction, the rewind target, and the corrected path are all
auditable from this repository alone.