OPEN_QUESTIONS

Live unresolved questions and competing hypotheses. Removing an entry requires recording what resolved it (constitution §45).

Representation

• What is a "relation" concretely at Layer 2 without smuggling in an ontology? Candidate: freeform typed tuples whose types are data, not schema. Competing: don't reify relations at all yet; store prose parses and extract relations only inside experiments that need them.
• Unit of artifact segmentation (§16): who decides passage boundaries, and is segmentation itself a recorded transformation from day one? (Design note says yes; the cost is unknown.)
• How does a parse's content get encoded so that two parses of the same passage are comparable without forcing them into one schema?

Evaluation

• Null-model construction (§33): what makes a control "relevant" for a myth vs a manual vs a diagram? Currently no answer. This is flagged as research, not plumbing — it is probably where this project succeeds or dies.
• Label destruction (§13): which perturbations are label-destruction vs structure-destruction? The boundary is itself a hypothesis to test.

Provenance

• §40 mentions "the original Winks work on inducing states in models." Does that material exist as an ingestable corpus/notes? If so it is both prior art and a candidate artifact set. Ask the researcher.

Instrumentation

• Which model families serve as generators vs critics at bootstrap (§31), given local qwen + frontier access on this box, and how do we record enough instrument metadata (§50) without drowning in it?

Senses

• The harp/guzheng probe (researcher, 2026-09-01): the researcher repeatedly hears/feels similarity between 17th-century Irish harp music and guzheng music. Wink-form of the question: does that felt similarity survive translation into the nonlinguistic audio basis — onset order, duration, raw pitch, returns structure — with timbre (and therefore instrument identity) destroyed? Candidate experiment: event graphs for both traditions, WL/spectral/GED within-basis, against nulls. What is a relevant null (§33) here? Candidates: other plucked-string traditions (lute, kora, koto — nearest neighbors), non-plucked melodic traditions (matched for scale systems?), the same pieces onset-shuffled (destroys returns topology, keeps pitch inventory), synthetic melodies matched for pitch-class distribution. If harp↔guzheng scores no better than harp↔shuffle, the felt similarity lives in something the basis destroyed (timbre? ornament? idiom carried by the listener?) — which would locate the similarity in the observer (§40), itself a result.
• Audio sense v0's known shear points fall exactly on these two instruments: polyphony (harp chords collapse to dominant f0) and glissando/ornament (guzheng bends forced into discrete events). Does the correspondence survive its own sense's loss vector, or is v0's basis too coarse to see what the researcher's ear sees? A negative result cannot distinguish "no shared structure" from "wrong basis" until a polyphonic/continuous-pitch sense v1 exists.
• The caption confound, staged for later: a model-mediated hearing channel ("describe this music") calibrated against the deterministic sense, the EXP3 move replayed for audio. Prediction worth freezing when the time comes: named-channel descriptions of harp and guzheng will import cultural structures not present in the event graphs (EXP3's beach-ball confabulation, but with "Celtic" and "Chinese" doing the importing).

The protocivilizational question (stage-setting, 2026-09-01)

Raw artifact (researcher, verbatim, lossy by own declaration — "everything i say is lossy so everything in quotations right"):

"i dont know how to compare: chinese/japanese top to bottom writing, divine to profane structure, totem poles, recursive metaphorical languages, exlusion of the I in structure, cultural behaviours vs. linear literal languages, english, linguistic evolution, phonetic structures from across the world (much accumulation on many of these things exist) and the assumptions are that human sounds are wired to have meaning. i dont not assume this, but i do believe there might be a protocivilizational fingerprint of a shared language. [...] i dont' 'believe' i want to know if wink sees a wink."

Parse α (agent, 2026-09-01 — one parse, not the truth): three candidate invariants + one meta-hypothesis.

1. Verticality as meaning-bearing axis — vertical writing, divine/profane cosmological axis, totem poles. Cross-domain claim: spatial axis carries ordered significance. Scene-sense family applies (writing IS layout). Trap with teeth: "lowest = least" on totem poles is largely an English-idiom projection — in several NW Coast traditions the bottom figure is the master carver's and most important. Axis-with-inverted-polarity vs. shared-polarity is a real structural distinction; if the felt wink runs on the idiom, the wink is observer-resident (§40) — a result either way.
2. Ego-in-grammar — pro-drop typology (WALS codes it) vs. mandatory-subject languages; contested cultural correlations (Kashima & Kashima lineage). Contested = testable.
3. Sound-meaning wiring — measured, small, robust: Blasi et al. 2016, sound–meaning biases across ~2/3 of world languages, controlled for family and geography. PHOIBLE / ASJP hold the data. Not an assumption to import; an effect to re-measure in our basis.

The fork that makes it falsifiable — fingerprint vs. attractor: inheritance (shared origin) and convergence (shared embodiment: same gravity, same mouth) BOTH predict cross-cultural invariance. They differ in similarity STRUCTURE: inheritance → lineage-correlated; attractor → lineage-independent. Relevant nulls (§33) must therefore preserve embodiment while destroying lineage, and vice versa. "Does the shared shape carry lineage signal above what shared bodies predict?" is the machine-form of the researcher's question.

Standing hazard, doubled here (§30): every model instrument was trained on the cultural associations under test. This corpus needs deterministic/typological bases more than any prior one; model-channel opinions about totem poles are not evidence.

Status: stage-setting. No corpus touches the store before null design and frozen predictions (§55).

The engineer's real questions (2026-08-21, pre-implementation)

STATUS: all five received researcher rulings same day — see ARCHITECTURE_DECISIONS/0002. None fully closed; each redirected: (1) no single Layer-2 scheme exists — typed multi-representation is the rule; still open: comparator design per scheme-pair. (2) three-term variance (artifact / parser / dents); still open: measuring the interaction term. (3) weighting = exposure-grown taste with provenance; still open: day-zero priors and the feeding protocol. (4) nulls = observed near-misses first; still open: when is the near-miss field too thin to judge exceptionality. (5) retrieval = trace-following + fingerprints, disagreement is data; still open: footprint deposit rules for the trace field.

1. Layer-2 encoding tension: fixed relation vocabulary = smuggled ontology (§8); freeform relations = parses incomparable without a further mapping step. Candidate resolution: freeform labels, comparison at structure level (WL/spectral/GED), label-agreement measured separately. Unknown: whether arity/direction choices already smuggle too much. Experiments 0–1 exist to answer this.
2. Parser variance vs artifact variance: if parse variance is dominated by model habits rather than artifact structure, the instrument measures LLMs, not texts. Must be measured before anything else is believed. (Sharpest form of Experiment 0.)
3. GED cost functions are §27's weighting: hand-set, versioned, sensitivity-tested — but this is where subjectivity enters wearing math's clothes. Watch it.
4. Null relevance: run three cheap null families from the start; treat which null is informative as itself a result.
5. Reverse-retrieval index is a basis (§64): name it in an ADR before building it.

The folder question (researcher, 2026-08-22 — entity-policy fix HELD on this)

Verbatim: "eventually the shape of when the clerk passes the folder to the auditor will by its own nature want to be noticed and audited. is this destructive repetition or is there information hidden in the action. i dont have an answer"

The reification the fix would suppress (models refusing to leave the passed object unnoticed) may be information-tracking, not noise. Our terse ground truth (clerk->auditor, no folder) may be the lossy basis. Held: do not force a granularity policy by fiat while this is open. Candidate discriminator: vary whether the object is causally active (altered/swapped en route) vs pure relay — if models reify MORE when the object matters, reification tracks information; if identically, it is habit. Second axis: "step 1 happens again" vs "a new folder starts the same route" — cycle vs helix; missed back-edges may be type/token discrimination rather than loss.

Loss-vector naming (from EXP10's correction, 2026-08-22)

"invented" means "unaccounted," never "false" — undecidable among confabulation / GT-omission / basis-shift from inside a comparison, and free-rename alignment undercounts it mechanically. Candidate schema_v1: rename to "unaccounted" + record edit DIRECTION explicitly on every stored loss vector. Until then: never gloss the dimension as confabulation (the seat did, in the headline of EXP10, and was caught by the researcher in one line).