RESEARCH_STATE

Updated: 2026-09-01 (post-restart; audio sense v0 built and tested; queue item (a) discovered already complete — scene sense + EXP3 run 2 + stored basis transformation predate the checkpoint; checkpoint queue was stale on this point)

Senses on the books

svg-document (v0, deterministic), svg-scene (v1, deterministic; the document↔scene basis change stored as tra_f321edd06bbe89ed), vision (qwen+mmproj), language ×2 — and now audio v0 (kernel/senses.py, deterministic DSP, zero tokens): note-event graphs with precedes + returns edges, raw-Hz pitch (no 12-TET snap), timbre deliberately destroyed (timbre is the noun — label destruction by construction). 9/9 tests (kernel/test_senses.py), including microtonal non-snapping and constructed-ground-truth recovery. Known shear: polyphony, glissando — recorded in SENSE_LOSS_AUDIO and OPEN_QUESTIONS §Senses. Motivating probe (researcher): felt harp↔guzheng similarity — see OPEN_QUESTIONS for the null-design question before any run.

Apparatus

kernel/store.py — implemented and validated (§42 sense; nothing settled): the five bootstrap objects over content-addressed JSON files, with invariants enforced in code, not convention — artifact immutability, competing-parses-as-sets, loss-vector-only (scalars rejected), mandatory producer metadata, lineage walkability, freeze-once predictions, append-only runs. 9/9 software tests pass (kernel/test_store.py). The kernel contains no similarity function, by design (ADR-0000/0001). Design note 0000 is standing (researcher delegated review 2026-08-21).

Lineage

LINEAGE.md cross-references the ancestor systems (jane, anamnesis/ana, riverbed, delta, confluence, llm-apocalypse) in the constitution's §28 correspondence form — inspected directly, not recalled. Researcher's stance confirmed 2026-08-21: those systems tried to build this into an LLM; incompatible; winks builds the system outside and observes. Riverbed/delta are the pivot — several constitutional constraints (§6, §12, §30, §31, §91) were operational there before being written here.

Results so far (first day — small n, toy shapes; nothing settled)

• EXP0: PASS on its frozen criterion. 24 parses, 24 conditional answers, 5 syntheses; 0 broken lineage walks; 23 entity-level disagreements — all reconstructed from the store alone.
• EXP1/2-lite (72 parses + 12 anonymized, 0 JSON failures, both local instruments). Frozen-prediction scorecard 3 supported / 1 failed:
• P2 SUPPORTED, dramatic: the sequence type annihilates branching (tree/diamond via eventseq: top-1 0.00, WL 0.05 vs relgraph 0.83). First clean representation-specific conservation result (§83).
• P3 SUPPORTED, better than predicted: label destruction IMPROVED recovery (0.33 → 0.42). Nouns were hurting structural parsing — models import world-knowledge through them. The §13 worry inverted.
• P4 SUPPORTED: prose is the WORST rendering (0.29 vs 0.46 both other renderings). §74 empirically vindicated on day one.
• P1 FAILED informatively: the designed 3-cycle family did not cohere (0.36 in-family). Mechanism: entity-granularity noise — feedback3's parses grew extra nodes and became diamond-shaped (0/6 recovered). But inversion→cycle3 sibling confusion (4/6) confirms the structural half of the prediction. Parses are noisy at the node-selection level before they are noisy at the shape level.
• Candidate claim (proposed, unregistered — kernel can't yet make claims about experiments, see ADR-0004): parse loss concentrates on re-entrant edges. Cycle-closing and re-merging edges are what vanishes (cycle3→chain, diamond→tree, feedback3 never recovered); purely divergent shapes (tree 6/6, chain 5/6) recover near-perfectly. Models under-parse the act of returning.
• Instrument difference: qwen 0.44 vs mistral 0.36 top-1 — the §50 "changing the model changes what is visible" effect, measured.

EXP3 (senses, run 1) — 2026-08-22

Predictions failed (P1, P3) or floored (P2) — and the failure IS the result: ground truth is basis-relative. The deterministic SVG channel delivers document topology; every model channel delivered scene topology; comparing across bases without a stored transformation floored the numbers. Claim recorded (clm_8dc3ec4c). The senses now on the books: svg-deterministic (zero tokens), vision (qwen+mmproj), language ×2. Kernel gained claims-about-experiments, run_id membership, spectral comparator.

Active research

Queue: (a) ~~svg sense v1 + EXP3 within-basis rerun~~ DONE pre-checkpoint (EXP3 run 2; tra_f321edd06bbe89ed). (b) The re-entrant-edge experiment (claim clm_685599 needs to earn belief) — note: the audio sense's returns edges now give this claim a second, nonlinguistic domain to be tested in; music is made of the act of returning. (c) Near-miss rerun with declared entity policy. (d) eventseq→relgraph conversion stored as transformation (ADR-0004 debt, still open). (e) NEW — harp/guzheng structural correspondence probe: null design first (OPEN_QUESTIONS §Senses), frozen predictions before any corpus touches the store (§55). (f) Parked, low priority: EXP11 hostile control (Ma & Zhang correspondence-degeneracy — see agent memory; inward-facing, runs after outward work).

Settled

Nothing. (§42 — this line should stay short for a long time.)

Standing failure-mode watch (constitution Part III, kept visible)

axis ungroundedness (SHARED — the seat judges performance on axes neither party can evaluate; verdict-language promotes unevaluated axes into settled ones. Discipline: measurement-language vs judgment-language separated in every report; axes versioned, plural, disagreement watched; the bottom axis stays named, not solved. Researcher's reminder 2026-08-22) · certainty gravity (SHARED — researcher supplies intuitions; agent returns them dressed as certainties. Parse-labels ("provisional", "parse α") do not neutralize certainty REGISTER: "the fork that makes it falsifiable" and "music is made of returning" were candidates delivered in the voice of facts, same session they were coined. Discipline: candidate structure in candidate voice; the confidence of a sentence is not evidence; wink exists to notice structure, not to conclude it. Researcher's reminder 2026-09-01: "i have intuitions, you often see certainties. this is our trap") · language gravity (SHARED — researcher and agent both default to language as the medium; the project is nonlinguistic by ruling ADR-0005) · closure bias · feature gravity · schema fossilization · summary drift · demo pressure · semantic gravity · mathematical cosplay · automatic reconciliation · premature generalization. The coding agent implementing this project demonstrated closure bias twice in the week before this repo existed (documented in its own memory). Assume these pressures are active.