0006 — The hanzi nursery, and the child's two memories

2026-08-22. Researcher rulings, verbatim: "i dont know how to teach the child. i almost want a chinese curriculum" → "yes. you see the rotations. and yes hanzi you understand. spec it."

The pun resolves to three design commitments at once: a Chinese-ROOM curriculum (no meanings, ever — structure only), Chinese PEDAGOGY (drill first; understanding condenses as sediment, never stated as rules), and HANZI as the literal corpus (a graded compositional structure system humans curated for three millennia).

A. The child's two memories (ADR-0002 ruling 5, built at last)

1. Trace memory (the field) — a riverbed instance for wink, delta's pattern: machinery imported frozen from ~/Development/riverbed, config + data owned here (field/config.toml, field/data/). History-dependent, taste-accumulating, footprints.
2. Fingerprint memory (the index) — WL colors + spectral signatures of every drilled structure, deterministic, ahistorical, zero tokens.

Discovery queries ask BOTH; their disagreement is recorded as data (what trace finds that fingerprints miss = the child's history; the reverse = structure it hasn't lived).

B. Contamination control — the load-bearing constraint

The models in riverbed's deposit machinery CAN read Chinese. Therefore no hanzi, pinyin, gloss, or meaning ever enters any prompt, deposit, or stored representation. At ingestion, every character and component receives an opaque content-hashed ID (K-ids). Composition operators become spatial relation labels (lr, tb, surround, overlay...). The field's generator rotates over sentences like "K4021 is K102 beside K88" — an LLM as rotation engine over meaningless tokens: the Chinese room, operating on Chinese, on purpose. The ID↔character mapping is stored once, deterministically, for EVALUATION ONLY, and never re-enters the loop.

C. Corpus and grading

• Source: cjkvi-ids ids.txt (verified: 88,939 entries, IDS operator trees). Stored as immutable artifact with provenance.
• Each entry → a structure graph (producer kind=code): component tree with spatial-relation edges, recursively expanded to base components. No rendering, no vision model — the IDS IS the scene graph, human- curated. Purest sense we will ever have.
• Grading: by tree depth/size (structural difficulty; imports no semantics). Cohort 1: ~200 base components + ~800 graded composites. The other 87,900 wait; cohorts are cheap to extend.
• Drill order groups shared-component families so recurrence does the alluvial work — then D3 permutes exactly this (below).

D. Drill loop

For each character in curriculum order: build graph → kernel artifact + representation (lineage) → fingerprint into the index → render opaque structural sentence → deposit to the field (batched by family to control LLM cost). Perturbations from birth (ADR-0003): a recorded fraction of drills are mirrored compositions, component substitutions, and one-component-off near-misses — the control population grows up with the taste.

E. Pre-registered discovery tests (frozen before the drill starts)

• D1 recall: held-out characters presented cold; do shared-component families come back top-k? Both memories scored; disagreement kept.
• D2 the built-in semantic control — the writing system's gift: hanzi share components two ways — SEMANTIC radicals (meaning-related families) and PHONETIC components (sound-related, meaning-unrelated). Raw component-sharing cannot distinguish them; positional composition might (radicals occupy characteristic positions). The test: can the child's structure-space separate semantic from phonetic sharing using position alone — without ever having seen a meaning? The writing system provides matched meaningful/meaningless sharing pairs: a null family built into the corpus by history.
• D3 curriculum permutation (ADR-0003): two field instances, same corpus, families-grouped vs interleaved order; compare tastes on identical queries. Path dependence of the child, measured.

F. Explicitly refused

Meanings, pronunciations, model-mediated hanzi parsing, the full 88.9k in cohort 1, any claim about the Chinese language. This is a structure corpus that happens to be culturally deep — and the cultural depth is quarantined behind the K-ids until an experiment deliberately lifts it.

Sequence

field instantiation → ingestion + K-id mapping → fingerprint index → freeze D1–D3 predictions → drill cohort 1 → run discovery tests → §116 inspection before cohort 2.

What the critic said (added 2026-08-22 — first entry of the standing practice)

1. IDS is the scribes' basis, not "the purest sense" — a curated document ontology adopted as truth because venerable (§8 risk). PARTIALLY DISSOLVED on inspection: hanzi natively carries THREE bases — composition tree, canonical stroke sequence (temporal), glyph geometry — so the corpus is multi-basis after all; ingest at least two before any claim.
2. D2's position→semantics result may recover orthographic convention rather than structure→meaning. Stands; weakens the claim, not the experiment.
3. Riverbed's deposit machinery has never eaten semantically empty K-id sentences; the generator may confabulate gloss over gibberish. Stands as an untested assumption — test on 20 characters before drilling 1000.
4. Single-corpus first-food imprinting (D3 risk at the crib). RESOLVED by researcher ruling, below.

Researcher rulings (2026-08-22, same session)

• "i keep forgetting its not solely about language. jesus that is going to be a constant failure mode for me" — language gravity is a SHARED failure mode (the agent's strongest modality and the researcher's native instrument, same attractor). Added to the standing watch.
• Mixed diet ruling: "we have nothing to lose really, we should do a mixed handful of your original atom, hanzi, siblings you mentioned." Cohort A (deposited first): the existing multi-type nursery + solar text + bakery control, with the atom text held out as EXP7's cold probe. Hanzi is a cohort, not the childhood. Siblings staged: molecules (graph+sequence+geometry, properties as quarantined semantics), Go records, origami crease patterns.
• Success-judgment disorientation ("we've abstracted my own abstractions") is handled by the standing machinery: §55 frozen predictions decide, not in-the-moment feeling.