Raising Beatrix: A Byte-Level Model's Measured Childhood

AbstractPhil · August 2026 · with Claude Fable 5 and Claude Opus 5 (Anthropic) · continues Agreement, Anchors, Addresses: A Week of Geometric Training

---

This installment is about one model. mini-beatrix-1 — Beatrix — is a 112.5-million-parameter language model that reads raw bytes instead of tokens, built from the aleph mechanisms the previous article introduced: learned anchor directions, signed addresses, expert modules born contributing exactly zero. Over ten days she was pretrained on 15.3 billion bytes of web text, given a measured "childhood" of nine curriculum stages, examined on a 270-item surface-disjoint benchmark, caught cheating on it, rewound once when a stage damaged her, graduated at 26.1 billion bytes — and then fitted with detachable 3.2M-parameter arms that taught her, among other things, when to stop talking.

The point was never the leaderboard. Beatrix exists to be a catalyst: a small, fully-instrumented, fully-public organism on which a series of upcoming experiments can run cheaply and honestly — tokenizer-to-byte distillation, anchored differentiated learning, structured cohesion between detachable skills, accumulation of systems over a fixed core, and prediction formats that respect geometric structure. Every mechanism in her was forced to prove its need through an arbitration — an experiment, an adversarial review, or a refutation — before it stayed. This article is the record of those arbitrations and what they bought.

Everything below is public: the training code, every checkpoint (including the damaged ones a rewind discarded), every arm (including the refuted ones), every exam report, and the research journal itself. Each repository carries a technical companion (TECHNICAL.md) with the full numerics; this article links them throughout.

---

Reading guide

As with the last installment, this article is a narrative. The companions carry the bulk numbers. If you want tables, jump to the artifacts section and open the TECHNICAL.md files; if you want the story of how a small model gets raised, examined, and repaired in public, read on.

Contents

1. The week between articles
2. The craft: a model with no tokenizer
3. Pretraining: organs elect themselves
4. Arms: skills that detach to bit-exactness
5. The head that slept through everything
6. The anneal, the contamination lesson, and the parallel lanes
7. A proper childhood: nine stages, one rewind
8. The exam era: catching our own flattery
9. The arm program: what 3.2M parameters can and cannot learn
10. What Beatrix is for
11. Bytelex: the direction of travel

The findings that matter most

In order of importance, not of date:

• Interleaved review teaches; blocked drilling memorizes — measurable in a 112M model. Every blocked curriculum stage aced its own material and moved the honest benchmark almost nothing; the interleaved review stage (S7) produced the only broad gains on surface-disjoint exams, and the final blocked stage undid part of them. Her benchmark-best checkpoint is mid-review (step 83,903), not graduation (88,508).
• Template competence is scale-invariant. Narrow-format training bought perfect in-format scores and near-chance transfer at every scale tried: billions of tokens into the core, thousands of rows into adapters. This finding forced the program's evaluation laws (surface-disjoint exams, held-out formats, generative tiers) and motivates the distillation lane.
• A 3.2M detachable arm can carry a real abstract capability. The rule-chaining arm generalizes to unseen sentence formats and unseen nonsense-predicate vocabularies (1.00 strict at 32-slot width) — and toggles off to a bit-identical base model.
• Arms steer priors; they cannot mint symbols. The same recipe that taught turn-ending perfectly via an in-distribution byte pair (\n\n, 6/6) taught nothing when the target was an out-of-alphabet byte (NUL, probability unmoved at 4e-8). Control symbols must enter at pretraining or not at all.
• Some capabilities refuse adapter SFT entirely. Three-digit subtraction stayed at zero through four escalating attempts — bare pairs, doubled capacity, a mixed generalist, and the borrow algorithm written out step-by-step. That quadruple wall is the precisely-motivated target of the distillation experiments now gated and running.
• A curriculum can be audited and rewound like code. Stage S3's corrupted mix drove general-text loss from 1.14 to 1.71 bits/byte and leaked its templates into her prose; boundary instruments convicted it, the run rewound to the prior boundary, and the corrected stage taught the same logic at ~1% cost. The damaged checkpoints remain public.
• The organs keep electing themselves. Removing her expert banks costs +2.2 bits/byte; removing the aleph attention hubs costs +3.7 to +4.9 — contributions that grew through every era while conventional diagnostics (rank floors, anchor merges) looked alarming. The one organ that never activated — a third output head — grew structure for 26 billion bytes while contributing exactly 0.0000, the program's cleanest open question.
• The biggest "architecture gap" of the bridge week was an eval bug, and the aleph is ~20× more optimizer-sensitive than softmax attention. A three-hunter audit showed the softmax-vs-aleph plateau was a window-truncation protocol ceiling (true gap ~.10, two readings retracted); the optimizer battery measured hub cross-optimizer spread at ~20× standard attention's (Muon .808 vs Adam .716) — the optimizer is part of the architecture became a number, and Beatrix's Muon/Adam split is that number applied.

---

The body of this article proceeds in date order, 2026-08-08 through 2026-08-17.

1. The week between articles

(2026-08-08, publication day.) The previous article went out in the morning; by nightfall its own artifacts were under retro-analysis. Three batteries ran over the stored checkpoints of the encoder era, and their verdicts became the design citations for everything below. A1 — inhibition-emergence: negative (inhibitory) dispatch is born, its amplitude learned (~14× monotone growth), and the anchor frame is decided in the first ~10% of training (drift 85% complete by step ~5.5k, long before the loss settles). A2 — toggle accrual: routed function arrives late and non-uniformly (one arm tripled its contribution in the back half of training). A3 — basin dynamics: a reversal worth printing — dense trunks reach 90% basin depth 3–4k steps faster on the training gauge, while frozen-random routing is the fastest on capability: a clean gauge dissociation, with learned anchors paying for their reorganization in path speed. The same night, a small-bed screen validated the week's whole screening methodology on its first cell: routing-path effects are visible at small budget (+.031, ~6× the seed band), and the small-budget advantage is routing, not learning.

(2026-08-09 → 08-10: the V-series — the army gets sized, then the bed gets audited.) A registered hypothesis — that expert-army size should follow the data stream's rank — went under an overnight 18-cell ladder and failed as written: knees non-monotone, effects tiny. The census explained the failure and minted the corrected law: the knee tracks the task's addressing demand, not the stream's generative rank — the first task simply made no addressing demand. The follow-up battery built a demand-controlled task (in-context associative recall) and found something sharper along the way: binding is address-capacity-hard for the linear aleph read and easy for softmax attention (.925 vs .056 at equal steps, with the aleph's signature late-bloom curve). The mechanical campaign winner was humbler still: the pad fix (+.053 — the hard-zero law had been live in every cell), with the unigram control cratering to .02–.06 and proving the trigram embedding load-bearing. By the capstone, provisioning and laws alone — zero mechanism changes — had closed 85% of the gap to softmax (.840 vs .936).

Then the program audited itself. A three-hunter corruption hunt over twenty candidates confirmed the worst one: the evaluation protocol scored unanswerable positions, so the .936 softmax "plateau" was protocol, not architecture (true scores .993 vs ~.89; two published readings retracted on the spot). The numerics were exonerated with constants attached — the differentiation structure fits in fp16 (top-1 agreement .9997), so precision was never the limiter — and one dormant landmine (an fp16 clamp flushing to NaN on ~5% of sharp reads) was defused. The optimizer battery then quantified something the program had long suspected: Adam is suboptimal for the aleph under autoregression (Muon .808 vs Adam .716), and the aleph's cross-optimizer spread is **20× wider than standard attention's** — the optimizer is part of the architecture is a measurement, not a slogan. And the clean-bed re-run retracted the campaign's own prettiest constant (K=4R — a pad-artifact) while re-confirming the demand law it decorated. Claim, audit, retraction, survivor: the whole method in four days.

(2026-08-11: the line ships.) The attention research closed publicly: splat_attention.py v2 — the causal, AR-validated form with dtype-aware clamps — plus eleven scrubbed standalone trainers and a README carrying the battery map, the repair flags, and the open problem stated plainly in the file header: high-demand binding saturates below softmax parity and anchor count does not close it — "we clearly need something else." That sentence is why the next ten days happened.

Why these points matter

• Frame early, function late (A1/A2): the two-timescale picture — addresses lock in the first 10%, capability arrives in the back half — later predicted the schedule battery's ordering and shaped the curriculum's boundary instruments.
• The demand law: capacity questions became measurable (consumed address rank) instead of aesthetic — the direct ancestor of the arm program's capacity-data law.
• The protocol retraction: the biggest "architecture gap" of the era was an eval bug; the audit-then-retract cycle became the program's standing shape.
• ~20× optimizer fragility: the Muon/Adam split in Beatrix's recipe is this number, applied.

2. The craft

(2026-08-12.) Beatrix has no tokenizer. Her input is raw UTF-8 bytes; each position composes the current byte with its two predecessors through three learned tables (a trigram embedding, ~591k parameters spanning a ~16.9M-state input space), with a dedicated pad row for sequence starts — a detail promoted to law after measurement showed that faking history with a legal byte starved address consumption. Sixteen pre-norm layers follow, three of them carrying CausalSplatHUB attention (a linear-cost signed-address read over learned anchors instead of softmax attention), every layer carrying an anchored expert bank dispatched by the signed sinh/cosh rule from the prior article — no softmax routing anywhere, negative (inhibitory) weights first-class.

Two design laws govern every new mechanism in her:

• Born null, weight-zero. New structure enters contributing exactly zero — via zero-initialized output weights, never via a zero gate. The distinction has teeth (section 5).
• Instruments before spend. The trainer ships with a census (per-layer effective ranks, anchor drift and merges, dispatch entropy, sign census), a toggle ledger (bits-per-byte cost of switching each mechanism off — the causal contribution gauge), and recall canaries. Nothing trains without its gauges already in place.

Optimization follows the program's measured split: Muon for transport weights, pure Adam (no weight decay) for embeddings and 1-D parameters, flat learning rate, bf16 compute over fp32 masters.

3. Pretraining

(2026-08-12 → 08-14, RTX 6000 Blackwell via Colab, resume-first sessions.) The diet: 0.3B bytes of WikiText warmup, then 15B of fineweb-edu, streamed. The headline of the era is in the toggle ledger, not the loss curve: the born-null organs elected themselves into load-bearing work, and the election was monotone from the first measurement. Across the first day's four ledgers (steps 4k → 10k), removing the banks went from costing +0.30 to +1.06 bits/byte — already ~25–50× the toggle class these mechanisms showed in the encoder era — and the hubs from +0.73 to +1.61. By 8.4B bytes the pair read +2.08 / +2.54; by 13.4–14.0B, the hubs crossed +3.00; at the 15.3B lock (step 51,882, validation 1.033 bpb): banks +2.29, hubs +3.25. Structure that entered at zero ended up carrying the model, with no balance machinery anywhere.

The same era registered the program's core directives (2026-08-13): the ethical-attunement direction for the Beatrix line — moral learning through storytelling early rather than rule-imposition an infant model cannot represent, heavy attunement deferred to later stages, a quiet tamper-evident evaluation planned rather than announced, and the organs' entanglement with the core read as a durability property: behavior woven into load-bearing structure is harder to strip than behavior bolted on. Those directives shaped the curriculum's design constraints two days before the curriculum existed.

The era also set two honesty precedents that shaped everything after:

• Session caps are pauses, not completions. A cap bug printed a mid-phase halt as if training had finished; the fix made every cap session-relative and its message explicit ("a pause, not completion").
• What did she actually eat? An empirical census of fineweb-edu (the exact streamed config) found 84–86% of documents sit at the minimum passing grade of its educational filter, topics diffuse, second-person advice register in ~half the documents, math in 3%, code in 2%, exam-style Q&A near 0%. Her later strengths and failures trace directly to this diet — and the reserved validation slice turned out to be 100% crawl-year 2013, a gauge-composition fact now recorded with the gauge.

4. Arms

(2026-08-12 → 08-14.) The program's thesis is that behavior belongs in detachable adapters while the core stays a raw prior. The arm system is amoe-lora: anchored mixture-of-experts adapters (RelayPatchwork) that attach to a frozen trunk and detach to bit-exactness — the toggle law, asserted with torch.equal, not eyeballed.

Integration was an arbitration gauntlet, and two of its findings became permanent laws:

• The label-shift bug: the adapter trainer passed same-position labels where the core expected pre-shifted targets, so the first arm trained to copy the current byte (loss ~10.7 nats). One semantics fix later, real conditioning started at 0.72 nats.
• Provenance must be structural. An adversarial review (three skeptic agents, fourteen findings) proved the adapter library's identity check was structurally inert for this model class — a foreign-core arm attached silently and produced plausible, wrong output, indistinguishable by inspection. The fix moved identity into the binding layer, made undeclarable identity a hard error, and put anchor-metadata verification into every loader. The public chat space serves only provenance-matched core+arm pairs as a result, with the arm/core switch live as a demonstration.

The first chat arm (3.2M parameters, 16 sites, 1,200 steps) made the locked core conversational — and its memorization guard, firing on a question space smaller than the training draw, planted the seed of the program's central evaluation law (section 8).

5. The head

(2026-08-14 → graduation, and still open.) Beatrix carries a third output head: an aleph read over a 512-anchor codebook, added to the standard readout — born null, intended to earn its way in. It never did, and the reason is the era's best process story. The original design gated the read with a scalar γ initialized at zero: a gate-zero birth, violating the weight-zero law — the gradient to the read is scaled by γ=0 and the mechanism deadlocks. A mid-run migration to weight-zero froze γ "at 1.0" — but freezing applies to the checkpoint's value, and the checkpoint held γ≈0.0015. The head spent the rest of pretraining building genuine structure (‖W_s‖ 17.9, effective rank 20) behind a fossilized ×0.0015 gate.

The repair was surgery, verified as a semantic no-op before it shipped: fold γ into the weights (max logit difference 2.4e-07), leaving a weight-zero-small head with its accumulated direction — exactly the recipe under which the banks elected themselves. Then the honest result: through the entire curriculum — 8.8B more bytes, symbolic stages included — the freed head grew ‖W_s‖ from 0.03 to 24.8 while contributing exactly 0.0000 bits per byte. Structure without function, measured at every boundary.

Late in the arc, a second instrument added a twist: the head's address boundaries went from lexically null to word-aligned (section 8's lexicon census) — it learned to read units it never used. The escalation ladder (forced-utility training screened on the smaller gate craft) is queued; the finding stands as the program's cleanest open mechanism question.

6. The anneal

(2026-08-14 → 08-15.) Before the curriculum, a standard 2B-byte anneal mixed higher-quality sources into the diet. Two of its components — dialogue rendered in her chat template, and identity texture ("I am Beatrix…") — were a design mistake with a precise signature: the bare core began answering "Who are you?" in the first person and bleeding identity into content ("The moon is… a small byte-level language model"). The chat arm trained on the annealed core had nothing left to teach: 3 of 4 probes identical with the arm on or off.

The arbitration that followed is worth recording in full, because both directions were corrected:

1. The first verdict — anneal as contamination, full stop — was half-retracted when live conversation showed the annealed core was the better conversationalist ("substantially more contained and rational").
2. The generalization drawn from that — "conversational competence may live in the core" — was retracted in turn by the program lead: compacting behavior into the core surrenders control; the core's designed role is a substrate that arms differentiate.

What survived is the law: conditioning corpora stay out of core training — anneal mixes are format-neutral prose; identity and persona live only in detachable arms. A retro toggle-ledger on the continuous fineweb gauge priced the anneal honestly (+0.018 bpb general text, with the hubs' contribution concentrating +3.25 → +3.73), and both cores — clean lock and annealed — remain public with matched arms.

The same day, the attention line — quiet since its public ship — returned as a standing improvement program, because Beatrix's instruments had diagnosed her hubs precisely. The mechanism is causal linear attention: additive prefix memory per anchor, a query read, agreement-mass normalization, constant-size decode state. That family's textbook flaw is no forgetting — interference accumulates over the entire prefix — and her gauges show exactly that signature: recall canaries at floor (0.00–0.016) all life long despite a recall thread in every training mix. The program registered six numbered directions against it, each pre-constrained by the house laws (all new machinery born at exact identity — function-preserving from step zero; the 37.6M gate craft as the screen; instruments before variants): per-anchor learned decay gates (each anchor becomes a register with its own persistence), delta-rule error-correcting writes, merge-and-respawn recycling of the 300–650 duplicate anchor pairs every census counts, per-hub toggle attribution (today's ledger only prices the hubs collectively), multi-constellation reads, and a placement/count sweep to decide which depths deserve hubs at the next scale.

The engineering lanes shipped alongside. A seven-explorer speed campaign folded a fused fast path into the hub attention (eager 4.85ms → 1.22ms combined with compilation at ctx 2048 — 4.0×, and wall-clock parity with standard attention by ctx 8192 — exactness-gated at 1.4e-06 against the naive oracle); its training adoption was gated on a gradient-parity check on the actual training GPU, and that gate later caught a compiled kernel whose forward matched to 7e-4 while its backward emitted NaN gradients on Blackwell — loss-only gating would have shipped it. The surgered core also shipped as a one-line AutoModel package (remote-code, parity 0.00e+00 against the native stack), and the chat space survived its own subplot — two gradio bugs root-caused to the framework (one a 6.24.0 regression, both eliminated rather than suppressed) and a platform allocation outage diagnosed with the receipts to prove it wasn't the app.

7. The curriculum

(2026-08-15 → 08-16.) Then Beatrix went to school: nine stages, ~8.8B bytes, designed as "a crash course through childhood" — grounding, self/other (taught as concept, never as identity, per the law above), categories and difference, rule-based logic, arithmetic and early science, cause and effect, learning-from-failure, an interleaved review spiral, and articulation. Every stage's data came from vetted streams or owned generators; every stage carried a fixed holdout — whole task families (depth-5 rule chains, three-digit subtraction, two bAbI families) excluded at the generator level so later causal experiments would have genuinely unseen material. Boundary instruments fired automatically between stages: probe suites, the fineweb gauge for era continuity, the head election trace.

The era's defining event was a conviction. Stage S3's first mix let its rule-chain templates colonize her prose — general-text loss spiked from ~1.14 to 1.71 bits/byte and the boundary probes dropped across unrelated suites. The instruments convicted the stage, the run rewound to the prior boundary (step 66,803), the mix was corrected, and the retrained S3 bought complete in-format logic at roughly 1% general-text cost instead of 48%. The damaged checkpoints stay in the public record; the annotated spike in the life chart above is them.

The second defining pattern was the spiral. Each blocked stage lifted its own topic and receded a stage later; the S7 interleaved review stage repaid most of the accumulated general-text debt (1.178 → 1.111) and — as the next section shows — produced the only broad transfer gains of her whole education. (One infrastructure note that almost cost the final stage: S8's definitions source declares a dataset feature type Colab's preinstalled library predates; the abort cost one minute — crash-safety left the hub state untouched — and the fix landed as a proper dependency floor in the package, not a notebook patch.) She graduated at step 88,508, 26.102B bytes, 70.5 hours of wall clock.

8. The exam era

(2026-08-15 → 08-16, instruments; ratified 08-16.) The first probe batteries were tiny (2–11 items) and, it turned out, flattering twice over. A false-belief suite went 0.00 → 1.00 across the self/other stage — until a surface-disjoint re-probe (novel names, objects, frames) showed the capability had largely pre-existed the curriculum and the jump was template match; the original 0.00 was equally an artifact of probe phrasing. Both directions of that error became the probe-validity law: evaluation items must share no surface with training generators.

The replacement — authored by a nine-agent workflow under that law, validated mechanically — is a 270-item exam battery: nine suites × 30 items, three difficulty tiers, family labels, holdout families preserved as exam sections. Its shakedown rewrote the story of the whole curriculum:

• In-template logic and arithmetic scores of 1.00 collapsed to ~chance (0.27, 0.23) on disjoint surfaces — on both the graduate and the pre-curriculum core. Blocked stages had bought template competence, not transfer.
• The one stage that genuinely generalized was S2 (difference/category): 0.50 → 0.77, with category and property families at 1.00.
• Mid-spiral, the exams caught the only broad rise of the era (0.35 → ~0.43 mean across suites) — and by graduation part of it had receded (0.39), because the final blocked stage displaced consolidation. Benchmark-best checkpoint: 83,903.

Two more gauge laws fell out of adversarial re-scoring during the arm program: multiple-choice-by-likelihood exams are blind to arms that reshape output format (a capability reading 0.00 on MC scored 0.70–1.00 generatively), and substring answer-matching inflates (1.00 → 0.70 under order-aware scoring). Every later verdict uses generative, order-aware, surface-disjoint evaluation.

A companion instrument, the lexicon census, reads her learned vocabulary rather than counting it: her surprisal boundaries define word-like units (with trailing spaces — the opposite convention to BPE's leading spaces), unit-boundary agreement with her internal address switches rises monotonically through the hub stack, and individual anchors acquired legible jobs — one deep-hub anchor fires on whitespace 88% of the time. The full method and tables are in the AlephLLM companion.

9. The arm program

(2026-08-16 → 08-17.) With the graduate in hand, the arm thesis went under formal test: a written process (phases, pre-registered decision gates, standing laws) rather than ad-hoc scripts. The first night produced a working stop arm — the graduate never once ended a turn in her life (0/6; she'd answer and then summon fictional speakers); a 3.2M arm trained on newline-pair turn endings took her to 6/6 with reply length halved and capabilities intact, robust at full sampling temperature. It also produced the program's tidiest controlled pair: the same recipe targeting an out-of-alphabet byte (NUL) moved its probability not at all from 4e-8. Arms steer the prior; they cannot mint symbols. (A first version of that refutation was itself retracted within hours — the gauge had measured a detached arm — and re-established with a valid instrument. The retraction cycle is in the journal; the evaluation law it produced, attach before gauging, is in the process document.)

The day campaign then ran the capability question at proper scale — 16k–24k rows per behavior in varied surface frames with held-out frames, capacity sweeps, guard warnings captured to ledgers. Results, per pre-registered gates:

• Rule-chaining is learnable and abstract: 0.70 strict at default width, 1.00 at 32-slot width and in a mixed generalist — on alien predicate pools never seen in training. The 16-slot arm's failure mode was termination (correct chain, then re-enumeration): pairing it with the stop arm lifted 0.70 → 0.90. Two arms, each repairing the other's failure — the anchored-MoE thesis in one line.
• Three-digit subtraction is not: 0.00–0.02 through bare pairs, doubled capacity, the generalist, and an explicit show-work corpus. Four refutations, one wall.
• Composition has laws: pairs compose cleanly; three same-direction behavioral arms stack into 13-byte degenerate replies. Gating, not addition, is the road forward for deep stacks.

The wall is the point. A distillation lane (bdist-e001) is pre-registered against it with numeric bars: a gated teacher (Qwen2.5-3B-Instruct, accepted at 0.96 ceiling accuracy after the 1.5B variant was rejected at 0.86), a byte-level pushforward of teacher distributions (alignment-gated at 1.0000 gold-byte agreement), and the cleanest possible comparison — same rows, same positions, same arm capacity, hard labels versus teacher distributions. If distillation breaks the wall SFT could not, that is the headline of the next installment; if it does not, the wall is architectural and equally worth knowing.

10. What Beatrix is for

The catalyst list, explicitly. Each of these now has its substrate, instruments, and in most cases its pre-registration:

• Tokenizer distillation — teacher token distributions pushed onto the shared 256-byte simplex (a pinned frame by construction, resolving the frame-ambiguity law of the previous installment without a Procrustes step). Gates passed; running.
• Anchored differentiated learning — the arm collectives: which behaviors and capabilities live in which detachable organs, what composes, what requires gating, what the substrate (curriculum vs control core) contributes causally. Holdout families were reserved from her first curriculum byte for exactly this.
• Structured cohesion — generalist-vs-specialist verdicts (the wide generalist matched every specialist at 1.00 while remaining one anchor), and the routing question underneath the stack-compression law.
• Systems accumulation — a fixed, fully-characterized core accreting instruments, arms, exams, and reports as public layers; the training repo is the ledger of that accumulation.
• Geometric prediction formats — byte-level distributions as the common measure space for models with incompatible vocabularies; her trailing-space lexicon versus BPE's leading-space convention is the first measured consequence.

11. Bytelex

The direction of travel, briefly. On 2026-08-16 the lexicon-census work forked into its own repository under a law-grade redirect from the program lead: never hardcode cross-tokenizer behavior to one Beatrix's internals — model weights drift as training proceeds; corpus statistics do not. geolip-bytelex is the result: a model-free relational system over byte information — corpus statistics (successor-branching entropy as the boundary functional) that don't drift with any model's training, plus token→byte expansion tables for a fleet of tokenizers. Its first verdicts: roughly 27–36% of every generative BPE vocabulary packs multiple corpus-entropy units under one token id (a ByT5 control reads exact zeros), and every generative tokenizer attaches whitespace leading while byte-native models grow it trailing — a convention that must be normalized before any boundary comparison. It supplies the alignment endpoints for the distillation lane above, and it is deliberately not this article's subject: it is the next one's.

---

Artifacts and companions

Method and attribution. This work was run by a team of two humans' worth of hands: AbstractPhil (direction, arbitration, hardware, and every final call) with Anthropic's Claude as research collaborator — Claude Opus 5 through the space productization and early Beatrix era, Claude Fable 5 through the curriculum, exam, and arm programs, with parallel Fable instances handling the splat-attention speedline and bytelex extraction. Mistakes in method described above (the flattering probes, the invalid stop-arm gauge, the over-general competence claim) were the collaborator's; the arbitrations that caught them were the process working as designed.

Prior installments: FT4 — Distill Against a Consensus · FT5 — Agreement, Anchors, Addresses