ifc-0224

16.6.1 Neural projection expansion in singing mice

The singing-mice package compares five laboratory mice (Mus musculus) with seven Alston’s singing mice (Scotinomys teguina). Rows of the MAPseq matrices represent barcoded motor-cortical neurons; sixteen columns represent target regions. The source study reports selective expansion toward auditory regions and the periaqueductal gray, targets implicated in vocal behavior [ Isko et al. , 2024 ] .

The executable counterfactual replaces singing-mouse AUD and PAG support by laboratory-mouse species means. Their combined mean positive-projection fraction changes from \(0.2102\) to \(0.0595\), a \(71.7\% \) attenuation. This number is grounded in the matrices, but it does not simulate neural dynamics, development, song, or evolution. The bridge from projection expansion to enhanced cortical control of vocalization remains a causal hypothesis.

Experiment: Prometheus substrate: singing mice. Observed layer: 12 animals, 16 target regions, and per-animal MAPseq summaries.
Executable edit: attenuate singing-mouse AUD and PAG support to laboratory-mouse species means.
Result: the combined metric falls from 0.2102 to 0.0595, a 71.7% reduction.
Admitted: dataset-relative arithmetic and provenance.
Not admitted: a corresponding change in song or vocal repertoire.

Controlled SGTE–1 recovery.

The preceding calculation supplies the projection-support identity. A stricter test hides it. The registered support family contains all unordered pairs of target regions with positive species-dependent projection expansion, while the oracle support {AUD, PAG} is unavailable to the estimator. Dataset M194, containing three laboratory and three singing mice, is used for discovery. Dataset M220, containing two laboratory and four singing mice, is sealed for confirmation. The estimator ranks targets by positive pooled standardized species contrast and admits the top pair.

On M194 the recovered support is exactly {AUD, PAG}. Its summed discovery score exceeds those of nineteen balanced relabelings, giving the smallest possible exact one-sided value, \(p=1/20=0.05\), for this split. Without any reselection, both effects remain positive in M220. The transported pair ranks second among all \(\binom {16}{2}=120\) pairs by summed positive standardized contrast, the \(99.2\)nd percentile. Replacing precisely those two M220 coordinates by laboratory-mouse means therefore agrees with the registered oracle counterfactual.

The controls show why typing matters. A random pair has only a \(0.83\% \) chance of exact recovery. Ranking unstandardized mean differences selects AUD and OMCc instead, while attributing the change to all fifteen nonconstant targets yields support \(F_1=0.235\). The result thus admits a sparse, dataset-relative projection-support signature across acquisition batches. It does not turn that signature into a causal account of song.

Experiment: SGTE–1: withheld singing-mouse support. Discovery: M194 only; the two-target identity is hidden.
Construction: positive standardized contrasts recover AUD+PAG.
Counter-witness: exact balanced-label permutation, \(p=0.05\).
Confirmation: rank 2 of 120 in sealed M220; both effects positive.
Baselines: raw differences recover AUD+OMCc; random exact recovery \(0.83\% \); global-support \(F_1=0.235\).
Verdict: admitted as SGTE–1 projection-support recovery, not as a causal theory of vocal behavior.

SGTE–2: constructing the observer.

SGTE–1 supplies positive-neuron prevalence as the measurement. SGTE–2 additionally hides that observer as well as the AUD+PAG support. The learner receives the binary, raw-count, and normalized-count matrices and nine registered summary constructions: targetwise means, totals, and compositional shares for each matrix type. An observer certificate requires a per-animal, targetwise marginal in \([0,1]\) that is invariant to row replication and to changes in positive barcode amplitude. Binary prevalence is the unique candidate with that type. In particular, a compositional share is rejected because one neuron may project to several targets; the target coordinates are not exclusive outcomes.

After constructing this observer, the support estimator again recovers AUD+PAG on M194 and transports it at rank 2 of 120 on M220. The uncertainty is material. Across the nine leave-one-laboratory-mouse/leave-one-singing-mouse jackknifes, the exact pair persists in \(66.7\% \) of replicates; AUD and PAG are included in \(77.8\% \) and \(88.9\% \), respectively. The discovery permutation value remains \(p=0.05\).

The observer ablation is decisive in this study. If the system ignores the typed certificate and simply chooses the summary with the largest discovery score, it selects normalized-count means and proposes HY+OMCi, not AUD+PAG. Across the twenty balanced M194 label assignments, only the true assignment passes the registered permutation and stability gates: attempt coverage is \(5\% \), selective accuracy is \(100\% \), and abstention on the nineteen false label worlds is \(100\% \). These percentages describe this finite counter-witness family only; they are not a general coverage guarantee.

Experiment: SGTE–2: observer and support recovery. Hidden: the projection observer and the two-target support signature.
Observer construction: typed invariance selects binary prevalence.
Support signature: AUD+PAG; M220 confirmation rank 2 of 120.
Uncertainty: exact-pair jackknife stability \(66.7\% \); AUD/PAG inclusion \(77.8\% /88.9\% \).
Selectivity: \(5\% \) attempt coverage, \(100\% \) selective accuracy, and \(100\% \) abstention on nineteen registered false-label worlds.
Ablation: score-only observer selection yields normalized means and the incorrect HY+OMCi support.
Verdict: admitted as a calibrated SGTE–2 result with a narrow counter-witness family.

SGTE–3: active counter-witness acquisition.

SGTE–3 asks not only what should be measured, but which measurement should be acquired next. It uses a controlled world with the same sixteen-target geometry as the singing-mouse example, while deliberately replacing the biological data-generating process by a registered Gaussian assay model. The hypothesis family contains a null theory and all \(\binom {16}{2}=120\) sparse two-target mechanisms. An assay has mean two when its target belongs to the hidden pair and mean zero otherwise, with unit variance. The learner stops when one posterior probability reaches \(0.9\), or after thirty-two assays.

The active policy selects the target on which the live posterior most strongly disagrees about support membership, discounted by its accumulated replicate count. This is a posterior counter-witness rule: it spends the next assay where the surviving theories make different predictions. We compare it with random selection with replacement and a fixed round-robin schedule, first at the full budget and then at a matched budget of twenty-one assays.

Across 6,000 pair worlds, the active policy attains \(86.9\% \) unconditional exact recovery, \(91.6\% \) decision coverage, and \(94.8\% \) selective accuracy while using 20.75 assays on average. At the full budget, random and fixed acquisition reach only \(19.0\% \) and \(33.5\% \) power while using 30.42 and 29.92 assays. At the matched twenty-one-assay budget, their powers fall to \(5.7\% \) and \(6.5\% \); the active gain over the stronger passive baseline is therefore 80.3 percentage points. On 1,000 null worlds, the active policy has \(98.4\% \) null safety and a \(1.6\% \) false-extension rate.

This result isolates the value of active acquisition, but it also marks a firm boundary. The likelihood, sparse mechanism family, and posterior acquisition rule are supplied, and an assay reads a coordinate of a synthetic world. SGTE–3 has therefore selected where to measure under that rule; it has not learned the acquisition language, performed an intervention on an animal, or established a causal effect on vocal behavior.

Experiment: SGTE–3: active counter-witness acquisition. World: null plus 120 two-target mechanisms over sixteen targets.
Acquisition: posterior support disagreement; threshold \(0.9\); maximum budget 32.
Active result: \(86.9\% \) power, \(94.8\% \) selective accuracy, and 20.75 mean assays.
Matched-cost baselines: random \(5.7\% \) power; fixed \(6.5\% \) power.
Null control: \(98.4\% \) safety; \(1.6\% \) false extension.
Verdict: admitted as controlled active measurement selection, not as biological intervention.