Qualify measured JO input before scoring

Status: complete design for review, not an implemented or qualified input, 2026-10-17. Applies to entry protocol 6 and M4 v2. The ADR, accepted 2026-10-18, defers runtime implementation until a measured source establishes the mode. No source response archive, Baker/Pacheco outcome, simulation, seed search or new validation lock was opened or executed for this design.

Measured JO input remains blocked. The audited sources establish useful mechanical and calcium constraints, but none establishes an accessible, licensed, identity-resolved absolute JO-A/B/D spike or calibrated membrane input covering the admitted stimulus domain. This is a bounded source audit, not proof that no suitable recordings exist. A normalized band-pass can support a labelled engineered diagnostic; it cannot satisfy protocol 6 by receiving an arbitrary Hz gain. The next executable step is source acquisition and qualification, not fitting Baker responses. Historical song v1, its stopped outcome and the failed v1 benchmark remain unchanged.

Source observations constrain different links

The required chain is sound particle velocity → arista displacement/rotation → JO native neural observable → declared simulation input. Direct arista actuation bypasses the first link. Air velocity and arista velocity are different observables: v = 2πfx applies to velocity and displacement of the same object, not across this biological transfer. Pressure-to-particle-velocity conversion requires a justified acoustic model and geometry; do not assume a plane wave in this near-field preparation. Every transfer needs its own units, uncertainty and domain. Preserve peak, peak-to-peak and RMS conventions explicitly; their familiar sinusoidal conversions do not apply blindly to pulse trains or noise.

A publication's open licence does not establish the reuse licence of recordings available only from authors. Archive-level and file-level notices must be checked. Replication units must distinguish animals, antennae, cells, ROIs, sessions and trials. Do not pool female mating conditions or invent a total n from figure-specific counts. Calcium exposure integrates away fast timing: neither band-pass shape nor ΔF/F ratios determine baseline, saturation, latency, carrier phase, adaptation, absolute rate or intercell correlations. Preserve calcium in native units for source validation; conversion to Hz requires separate simultaneous electrical calibration and independent validation.

Identity is a separate gate from transfer calibration

The frozen MaleCNS selector ^JO-[ABD] selects 146 bodies: A=50, B=88, D=8 from data/neurons.parquet, SHA256 7ec0020f08495c35398deafa3183292dfb489ce6717d889b80cfcc11cb757c7c. These are local anatomical counts, not measured transfer functions. Side comes from instance, since somaSide is absent. There are 85 left/61 right bodies: A 18/32, B 59/29, D 8/0. Left-only D is reconstruction coverage, not biological unilaterality. Do not invent right D cells or mirror their spike trains.

The 37 -unclear bodies remain subgroup-only. Body 308087 lacks a flywireType; 138093763 is MaleCNS JO-B-unclear with flywireType JO-B, so neither names a subtype and neither is upgraded. All selected male dimorphism fields are absent. Female raw annotations have 412 selected rows, all annotated isomorphic; processed female parquet has 411, including no root 720575940616042508, and includes JO-A5 absent from the male selection. The processing discrepancy needs provenance resolution before any female coverage denominator. Female anatomical isomorphism does not establish male physiological equivalence. Dataset provenance remains authoritative for source versions.

A/B/D projection zones, overlapping genetic drivers, calcium clusters and modern EM subtypes are distinct identities. Do not assign pulse-only A, sine-only B, or a D centre frequency/gain from a qualitative preference. Group-level evidence can admit an explicitly labelled population model at group resolution; it cannot justify subtype-specific transfer functions. Freeze one mapping row per source observation × target body, retaining source animal/ROI linkage. Fan-out is not 146 independent observations. Female-to-male physiology needs direct evidence or a restricted female claim, which does not automatically qualify MaleCNS protocol 6.

Choose the numerical boundary from the observed signal

The current song generator supplies voltage-kick counts, not measured JO emissions. Defaults give approximately 48.08 mV per event against a 7-mV rest-to-threshold gap; the binary envelope has no measured carrier transduction. Dead-time filtering uses 2.2 ms and thinning equalizes per-cell counts across conditions, deliberately discarding integrated-count differences. Stimulated cells bypass ordinary refractory deadlines but retain recurrent conductance and thresholding. Thus 454.5 Hz is not their hard output ceiling. The kernel thresholds before delivering input, so ordinary kicks also acquire next-step emission timing. Actual JO spikes and input kicks are separately recorded and are not guaranteed equal.

The proposed primary boundary is outgoing source-event replay. Observed JO spikes prescribe emissions directly; their source rows cannot also generate LIF spikes. Preserve outgoing chemical weights, signs, fan-out and delays. Record excluded feedback to clamped JO cells and limit the claim to the network driven by observed source output. Audit incident gap junctions before admission: reject unsupported electrical coupling rather than invent source voltage. The decision is pending in the ADR.

Absolute measured rates require a separately justified rate-to-event model: rate is not an observed train. Independent Poisson, shared latent drive and empirical trial sampling make different correlation assumptions. A pooled ROI cannot establish independent cells, and cloning one train fabricates perfect correlation. Require source-only evidence for timing/correlation assumptions and predeclared uncertainty/sensitivity treatment; otherwise refuse the claim. Do not silently impose dead time or count matching on an absolute rate. Calibrated JO current/conductance or membrane recordings offer a separate membrane mode only with recording location, holding conditions, reversal potentials and independently measured JO conversion parameters. Recurrent feedback and the declared biological refractory treatment then remain active; do not simultaneously force emissions. Neither a postsynaptic current nor an arbitrary pA-to-kick gain qualifies this mode.

Freeze typed assets and falsify their transformation

The following is an artifact contract, not a new implemented format. A manifest references immutable native files, mapping and transforms by content hash; no implicit latest. Use explicit null plus missing_reason for unknown metadata, never zero, empty-string eligibility or guessed defaults. Required unknowns block qualification. Distinguish absence of events from missing acquisition intervals; missing intervals cannot become silence. Invalid units, timestamps, identities, negative rates, nonfinite values and corrupt bytes are typed refusals.

Stimulus preprocessing must resolve the carrier with documented acquisition and anti-alias filtering; neural sampling must independently resolve the claimed neural timing. A 0.1-ms simulator tick supplies neither lost carrier phase nor missing calcium bandwidth. Align stimulus and neural clocks explicitly, retaining latency rather than fitting it downstream. Use half-open intervals [t0,t1) and bins [k·dt,(k+1)·dt) relative to frozen t0: an exact boundary enters the next bookkeeping bin. These bins do not specify physical emission time. Preserve the native timestamp t and dispatch at tick q(t) = ceil((t−t0)/dt): an event exactly on a tick stays on that tick, and dispatch never precedes the recorded event. At dt=100 µs, source times 0, 99 and 100 µs dispatch at 0, 100 and 100 µs. Thus the latter two collide at one dispatch tick, although they occupy different acquisition bins. Preserve multiplicity at dispatch or refuse the incompatible timestep/mode; never boolean-clip events. The deterministic quantization error lies in [0,dt) relative to the declared source clock, with clock-alignment uncertainty accounted separately. Do not fit away this shift using downstream responses. Apply the exclusive source end t1 before quantization. A valid source event that rounds to the run horizon must be retained in an explicit continuation or cause horizon refusal, never silently dropped; without a continuation, extend the declared run before admission or refuse that boundary.

For rate models integrate the declared intensity over bins rather than sample an arbitrary endpoint. Materialize complete events offline and hash them, or use a vetted deterministic key schedule containing asset hash, algorithm version, master seed, condition, trial, source/bodyId and absolute bin. Chunk boundaries, worker order and checkpoint restart must not change realizations. Source data, ordered identities, transform, generated events and numerical semantics enter run identity; no historical lock is rewritten.

Independent acceptance cases must prove the transformation, not just internally consistent hashes. Hand-derived two-source events at zero, bin boundaries and exclusive end test timing and multiplicity. Close spikes inside 2.2 ms remain unchanged under event replay or receive an explicit refusal. Strong recurrent feedback must not add source events under clamp; it must affect a calibrated membrane-mode control. A one-source/two-target fixture checks exact signed, weighted, delayed deliveries with no duplicate LIF spike. Constant, step and analytically integrated sinusoidal rates test integral, latency and aliasing against formulas independent of production conversion. Statistical tolerances come prospectively from the chosen process variance and instrument uncertainty, not invented empirical thresholds. Chunk/restart/reordered-worker tests require identical per-body event bytes. Negative cases include ×1000 unit errors, A/B swaps, missing sex/side, unsupported calcium-to-Hz conversion, leaked held-out animals, altered hashes and out-of-domain stimuli; all refuse before run mutation.

Logic and architecture make admission explicit

The selector defines candidate anatomy, not a claim that all 146 bodies share one auditory transfer. D's static-deflection responses and coarse 100–200-Hz vibration preference do not resolve DA/DP functions or a universal auditory encoder. Observed train replay is conditional on its measured stimulus/trial; it is not a reusable sound-to-spike transducer for novel stimuli. A reusable transducer requires independently validated stimulus-domain transfer evidence.

The causal graph is stimulus → measured JO output → chemical delivery → central network. Source clamping cuts central feedback → JO output; excluded recurrent contacts remain in an audit ledger. Central efferent perturbations therefore cannot establish effects on JO transduction in this intervention. Such claims need a new closed-loop sensory-physiology model and protocol.

Admission follows catalogued → acquired → identity_reviewed → calibrated → source_validated → frozen → numerically_qualified → protocol_eligible. Every transition consumes immutable evidence and emits a decision/reason record. Missing evidence yields blocked at that stage; incompatible units/licence, identity contradiction, corruption or unsupported domain yields refused. Changing a consumed asset creates a new identity starting at the affected gate; there is no override from a downstream score. protocol_eligible still does not mean independently validated downstream prediction.

The proposed integration separates read_native_asset(manifest) (units, licence, missingness and checksums), resolve_body_map(source, dataset) (explicit identity rows), qualify_transfer(source, domain, split) (scientific admission), materialize_source_events(asset, mapping, clock, sampler) (immutable event artifact), and deliver_source_events(events, connectivity, tick) (accepted numerical boundary). These are prospective contracts, not existing APIs. A run invokes the first four before mutating runtime state; the last consumes only qualified events and preserves chemical edge semantics. A verifier compares native-to-event transformations and emitted/delivered events using an independent analytic oracle. Calibration provenance and held-out outcomes remain separate inputs, with no target-response dependency in source admission.

The preregistered mathematical controls, standalone script and results supply six finite cases and 19 passing assertions. They expose carrier/envelope ambiguity, sampling aliasing, amplitude conventions, multiplicity/quantization loss, underdetermined correlation and normalized-gain ambiguity. These counterexamples support design refusals; they test no production adapter, biological recording, calcium forward model or downstream prediction.

Experimental decisions are fixed before observations

Before the pilot, freeze acquisition/identity rules, instrument error accounting, missing/exclusion policy, response definitions, experimental units, split method and the algorithm that turns pilot variance into confirmatory sample size. Choose the target precision or minimum relevant effect and power/error rates from the scientific claim before confirmatory acquisition; record pilot-derived n before those trials. No fixed empirical threshold or powered n is defensible from this source audit alone. Calibration animals cannot reappear as validation animals; retain failures and acquisition dropouts. If new choices follow an external validation failure, the next validation needs fresh animals and a new version.

Prelaunch decision record

Before confirmatory acquisition, freeze the unit of inference as an independent animal, the calibration/validation animal allocation, the supported stimulus matrix and endpoint families. Within each animal, estimate absolute firing-rate residuals in fixed baseline/response windows, onset-latency residuals with a fixed detection rule, and adaptation residuals as the prospectively defined late-minus-early response contrast. Native membrane-mode endpoints use physical voltage/current units instead of an invented rate. Pool repeated trials only through the declared animal-level estimator; cells and ROIs within an animal are nested observations, not independent n. Identify every primary condition × endpoint comparison before observing validation trials.

For each endpoint j, define an equivalence margin δ_j from source/instrument uncertainty and the scientifically acceptable prediction error, with units and justification retained. A margin cannot be enlarged after seeing model error. Compute the preregistered animal-level residual confidence interval with the chosen coverage and multiplicity adjustment. An interval wholly inside [−δ_j, +δ_j] passes equivalence; wholly outside that interval fails; any overlap with a boundary is indeterminate. Requiring all primary endpoints to pass defines the overall pass; any primary failure yields failed, otherwise an indeterminate endpoint yields indeterminate. A nonsignificant difference never establishes equivalence. Report each condition even when the overall decision is stopped.

A concrete planning rule is precision-based. For K prespecified primary comparisons and familywise error α, use αj = α/K (or another frozen justified allocation). Given a source-only pilot estimate s_j of the standard deviation of independent animal residuals and target confidence-interval half-width ε_j, start with n_j = ceil((z*(1−α_j/2) · s_j / ε_j)^2) and take the maximum across required endpoints. Before confirmation, recalculate upward with the t critical value at n−1 degrees of freedom until the planned half-width criterion holds; record the resulting n and pilot variance estimate. Choose ε_j prospectively relative to δ_j and a justified planning residual, so the intended interval can fit inside the equivalence region. This rule plans precision, not guaranteed equivalence power*. If a power claim is required, specify an assumed true animal-level residual, target power and a prospective equivalence-power analysis; the precision formula cannot replace it. Small/nonrepresentative pilots, non-normal residuals, clustering, exclusions and uncertainty in s_j require a prespecified conservative variance bound or simulation-based planning under the chosen analysis model. Freeze attrition handling and maximum recruitment; do not stop opportunistically when equivalence first appears.

Missing calibration, bandwidth, licensing, subgroup coverage or sex-transfer evidence yields blocked before confirmatory comparison. Adequate measurements showing a residual interval outside its margin yield failed. Adequate-domain measurements with insufficient precision yield indeterminate. These states cannot be exchanged by relabeling acquisition exclusions. A narrower A/B or female-only source can support an explicitly restricted experiment but cannot close the full MaleCNS A/B/D requirement. Reconstruction's missing right D bodies remain absent; do not invent them or claim full biological bilateral coverage. No Baker/Pacheco target response enters endpoint selection, tolerance, sample-size planning, exclusions, source-transfer fitting or subsequent reconsideration.

Downstream experiments are later: compare the qualified input with constant-rate, spike-silent, envelope-only and correlation-assumption controls under a fresh prospective protocol, with simpler stimulus-following baselines and independent negative controls. These controls identify which input features matter; they cannot be selected by inspecting Baker/Pacheco responses. A successful input qualification permits proposal of those experiments, not their automatic scoring.

Acquisition and review precede a new experiment

First pursue original JO trials from Wilson/Kamikouchi sources and the public Yamakoshi JO-only file inventory, checking allowlists before archive inspection. Request waveform, mechanics, clocks, electrical traces if available, unnormalized fluorescence otherwise, anatomical verification, animal state and reuse terms. If a new experiment is necessary, first establish whether identified JO output is spiking or graded and whether it is technically recordable. Synchronize native neural acquisition, near-fly acoustics and same-preparation arista mechanics; include baseline, randomized carrier/amplitude/pulse conditions spanning the exact protocol domain, recovery and adaptation blocks. Pilot variance determines a later preregistered sample size; historical stimulus frequencies do not supply one. An independently validated paired electrical/calcium calibration is needed if only fluorescence is ultimately available.

This sequence completes the design while leaving source eligibility honestly blocked. Input qualification alone does not make Baker eligible, resolve its sex/readout gates, authorize a new lock, or turn physiological success into behavioral M4 completion.