Open questions

Things this design has not settled. They are written down here rather than left out, because a reference that only describes what works is not a description of the system.

• Tempered SMC for generation. The Boltzmann target is written down but not sampled from. Whether the crossover population kernel is worth the complexity over local climbing is untested; the measured non-concentration of the pool — it widens slightly over a session — weakens the diversity argument for it.

• Cross-island discovery. Closed by measurement, and it was wrong. This entry read: "local refinement from island A will not find island B; a tempering schedule would cross the valley." Measured against a bimodal synthetic user (make islands), 20.9 % of refinement events cross islands — 13.5 % of all events decisively, with both ends more than 1.0 onto their island rather than hovering at the boundary — and 0 of 8 seeds ended with a pool on one island only.

  The reasoning was wrong about the geometry. The walk is not local in feature space: it is a reversible-jump walk over a tree grammar, and a single accepted structural move swaps a subtree, which is a large jump in φ. There is no valley to cross, because the search does not have to travel through the space between the islands. See Refinement.

• Fan-out and feedback in the grammar. Two separate ceilings, deferred together because they are the two things the term algebra cannot say.

  This entry used to read "the grammar is DAG-only today", which understated it in the direction that matters. The genome is a tree — term.rs says so in its first line — and a tree forbids more than cycles. It forbids sharing: one output cannot feed two places, so there is no shared sub-patch. A DAG would already allow that. The distinction is the whole of the first half of this entry, and the docs were describing the looser of the two ceilings.

  Fan-out is the more valuable of the two and the more invasive. One oscillator into both a filter and a delay line, summed — an idiom so ordinary that the app already has to explain its absence. The connect offer does that well, volunteering the constraint as a fact ("A copy: one output cannot feed two places"), and the panel called it the best copy in the product. But it is still a ceiling being narrated rather than lifted.

  The cost is not the grammar rule; it is everything keyed to the tree. Child indices (node/0, node/1) are the trace addressing, so a shared node has no single path and the address scheme stops being a naming of the term. That scheme is load-bearing for panel knobs, locks, live parameter handles, MH proposals and the persisted genome — CONTRIBUTING.md lists it as a sharp edge for exactly this reason. children(), size(), depth() and site_count() all assume each node is visited once, and size is φ's parsimony term. Every structural op in mutate assumes a unique parent. It is a genome-format change with a migration, not a production.

  Feedback needs a mandatory attenuator and limiter in the loop path, and a delay of at least one sample to be computable at all. quiver's graph is evaluated per sample in dependency order, so a cycle needs an explicit unit-delay node to break it — which is a real design, not a relaxation of the acyclicity check.

  Deferred, deliberately, and this is the record of it. Neither is blocked on evidence — no measurement would change the answer — so neither belongs in the "measure it" pile with the rest of this page. They are blocked on being worth a genome migration, and nothing in the loop currently says they are: the search is not starved for expressiveness (refinement crosses islands, and the pool widens rather than concentrates). Re-open this when a listener wants something the tree cannot say, rather than when someone notices it cannot say it.

• Per-style audition phrases — a discovered bass style picks a bassline, a pad style a chord swell. Still open, and worth stating precisely what stands in the way, because the migration mechanism is not it.

  The :p2 stimulus tag does solve the history problem: a phrase change renames the audio coordinates, old votes keep their stimulus-independent structural coordinates, and their old-stimulus audio coordinates are imputed as "no evidence" — which the likelihood now handles honestly rather than as a measurement. Two prerequisites are therefore already met, and one is not:

  • Comparability is fine, contrary to the obvious worry. The phrase is a property of the session (SessionConfig::phrase), not of a candidate, so everything in a pool is auditioned under one stimulus and duels stay apples-to-apples. Per-style phrases only make sense as per-context phrases for the same reason.
  • The tag would have to be derived rather than declared. :p2 is a hard-coded literal in AudioFeatures::NAMES, and Features::phi_names is global. A phrase that varies needs the tag to be a function of the PhraseSpec actually rendered, or the names silently stop describing the numbers.
  • The real obstacle is circular, and it is a design problem rather than an engineering one. A style is discovered — it is an inference from φ. φ is measured under a phrase. If the phrase is chosen by the style, then the stimulus depends on an inference that depends on the stimulus. That loop can be broken (bootstrap from the standard phrase, switch only once a style's share is confidently high, never re-audition history), but every way of breaking it is a decision about how much the instrument is allowed to change what it is measuring while it measures it.

  Deferred until that loop has an answer worth defending, rather than until someone has time — the mechanism is ready and the question is not.

• Where acquisition would earn its keep. Measured, and the tie does not break. BALD ties uniform pairing at session horizon, and this entry named two regimes where that should stop being true: a much larger pool, or a much longer session. Both were run at 20 CRN-paired seeds, bald − random:

  regime	cos θ*	rank r	excess nats
  baseline (pool 48, 6 rounds)	+0.059 ± 0.046 (static)	+0.045 ± 0.068	−0.013 ± 0.012 (static)
  pool 192	−0.002 ± 0.044	+0.015 ± 0.061	−0.001 ± 0.012
  24 rounds (288 duels)	+0.031 ± 0.042	+0.013 ± 0.013	−0.003 ± 0.006

  At the baseline BALD has two marginal wins in the static regime (t = 2.6 and −2.2). Widening the pool fourfold removes them rather than growing them, and lengthening the session fourfold leaves everything inside noise with several signs flipped. The reasoning behind the entry — that a bigger pair space gives an information-seeking rule more redundancy to prune — does not survive being tried.

  What is stable across all three regimes is that BALD beats dueling Thompson (t = 2.9 to 6.9), which was already known and is unchanged.

  So uniform random pairing stands as the default on the same grounds it always had: it ties the information-seeking rule everywhere anyone has looked, has no tuning constants, and makes every duel an unbiased calibration sample. The session-length knob this needed (--rounds) is now in learn_synthetic beside --pool, so the next person can ask a third regime without patching a constant.

• Fit cost at the K cap. Single-site MH re-executes the whole program per step, so a mature fit is both slower and statistically thinner than an early one (205 + S sites over a fixed 10 000 steps ≈ 48 sweeps per site). The address table is hoisted out of the step loop and the chain no longer holds itself in memory, so what is left is purely the statistical shape of the problem — the budget can now be chosen on the recovery tables rather than against a memory ceiling. The written-down option (cap $K$ at 3) is gated on style_share evidence from real sessions.

  That evidence is now collected. Every posterior fit records what fraction of the pool each lens claimed, and the register persists across reloads (Engine::style_shares). The question is still open — it wants sessions, which take time to accumulate, and synthetic runs cannot answer it — but it is now open for want of data rather than for want of an instrument. Rows where k == k_styles are the ones that bear on it: k grows with the log, so an early row with two lenses is not evidence that lenses 3–5 are idle.

• Which state of a refinement walk to inject. Run, and it tied. A walk renders ~40 candidates and keeps one; RefineKeep::Best takes the highest-log π_β state the walk occupied, seed included, and ships switched off. Sixteen paired seeds:

  	Last	Best
  mean gain	+1.927 ± 0.452	+1.774 ± 0.302
  median gain	+2.058	+1.819
  10% trimmed	+1.840 ± 0.383	+1.925 ± 0.190
  climbed on	14/16	15/16

  Paired difference (Best − Last): mean −0.153 ± 0.384, median −0.185, trimmed −0.113 ± 0.318, sign test 8 better / 8 worse (p = 1.000). As exact a tie as sixteen seeds can produce, and it does not clear zero at 2 se on any statistic — so the default stays Last, kept re-checkable rather than deleted, as Acquisition::Thompson is.

  Neither the feared failure nor the hoped-for win appeared. The worry was that argmax over a surrogate would deepen the catastrophic tail; across the pair the tails are a wash. What did show is that Best is the lower-variance rule rather than the better one — half the trimmed standard error (0.190 against 0.383). Injecting the walk's argmax is more consistent than injecting where it stopped; it just does not aim anywhere better on average. That is the argument to re-run this on if the surrogate ever gets sharper.

• Interior signal taps in quiver. Closed, and it was wrong. This entry read: "a compiled patch exposes exactly one output … a quiver-side probe API would turn both into measurements. Not filed — it needs scoping first." There was nothing to scope and nothing to file. quiver's StateObserver has taken Level, Scope and Spectrum subscriptions on any node port for some time, in the release the lockfile already pinned. The gap was here, not upstream.

  The compiler now records where each term node's audio leaves it, and the rack's flow animation multiplies a measured RMS into its reach factor while notes sound. Two expectations about the work turned out not to hold either: sync_output_keepalive is unnecessary, because the genome is a tree and every module's output already feeds a parent, so quiver is already computing every metered value; and the open design question — which of N voices to meter — has an answer, the most recently pressed one, since a sum across the bank averages notes at different envelope phases and is not the level on any wire. The port trace stays an offline render, now by choice: it wants the same phrase every time so that two looks at it are comparable.

• fugue-evo's parallel feature on wasm32. Closed, and it was wrong three times over. The entry read: "it does not compile there, so the workspace takes fugue-evo with default features off and refinement is single-threaded natively too — in the one place the engine is embarrassingly parallel."

  Wrong about the blocker: fugue-evo#22 established that checkpoint, not parallel, was the only thing that did not build on wasm32. Wrong about the remedy: enabling parallel would change nothing here, because every rayon use in fugue-evo sits under #[cfg(feature = "classic")], and Auracle takes ["std", "ppl"] and drives refinement itself through inference::mh::EvolutionChain. And wrong about the prize: the harness is not waiting on single-threaded refinement. search_health and the refinement_improves_pool floor already spawn one thread per seed and saturate the machine, so parallelising inside a refinement cannot make a 16-seed measurement faster — the cores are already busy.

  What is left is real but smaller than the entry implies, and it is a UX number rather than a harness one: latency on a single refinement, which is the app's ⚡ button. Filed as that, not as a build-configuration change.

• Remaining quiver hardening (non-blocking, tracked upstream): voct_to_hz is unclamped — overflow is now recovered by Q198 rather than prevented, and a pitch clamp would also tame aliasing garbage at absurd-but-finite pitches.

• The brightness cluster in φ_audio. rolloff_mean, zcr_mean and centroid_mean are three genuine measurements of one perceptual thing. A fused prior over the cluster is now implemented and switched off, which is a more useful state than either "not done" or "done".

  The VIFs quoted when this was written were 18.4 / 10.4 / 5.9; after the ZCR DC removal they measure 16.9 / 9.7 / 5.9 and zcr_mean no longer trips the collinearity flag at all. A third of the original argument was a coordinate bug rather than a modelling problem.

  Two gates were run at ρ = 0.25 and they disagreed. The closed-loop gate, which scores θ recovery, improved (0.657 → 0.702). The 48-seed paired climb, which scores what the pool is worth to the listener, regressed — −0.579 ± 0.188 trimmed (−3.09 se), sign test 16 better / 32 worse (p = 0.029). So it ships at ρ = 0.

  Both results are real because they measure different things: pooling an ill-conditioned ridge regularizes estimating θ, and biases the search that consumes θ. The general point outlives the feature — a VIF says these coordinates move together across patches, which is a fact about φ; fusing their coefficients asserts a listener's preferences move together, which is a fact about people and does not follow. Re-open if the listener model ever gains a reason to believe it does; the sweep and both gates are there to re-run.