Anesthesia Pipeline — Network-State Validation

Thesis

General anesthesia is driven by coordinated partial perturbation across a multi-class target ensemble (SNARE release, Complex I, K2P channels, nAChR, NCA leak, GABA-A potentiation), not a single-target binding event. Mutants in any one target give partial phenotypes; non-immobilizers like hexafluoroethane lack engagement at the high-magnitude classes the architecture requires.

What this project earned, stated once: within-organism volatile EC50 prediction ~3× tighter than Meyer-Overton on held-out isoflurane across worm, fly and mouse, with one free parameter per organism; mutant magnitude predictions inside published literature bands; and a 4–6 mechanism-class minimum sufficient subset in which SNARE-or-Complex-I is universal (25/25 passing subsets).

What it does not establish. A 2026-06-12 preregistered, accept-either-way re-examination (see Adversarial close-out) proved the V1 operator is rank-2 — every anesthetic and genotype profile collapses to a uniform current scalar plus one global synaptic gain, so the network cannot by construction distinguish which targets are engaged from the aggregate magnitude of two coordinates. Cell-type targeting, tested on a rank-lifted substrate built specifically to make it testable, returns a null on worm (46th percentile). The immobilization readout is demoted to a network statistic, not validated behavioral quiescence. Whether the specific target identities are load-bearing, rather than one magnitude-matched ensemble among many, is not established here.

Retraction (2026-09-21). Earlier versions of this page headlined a worm-anchor-overfit / fly-uniquely-special / mouse-magnitude-driven trichotomy. That reading came from a null-model bug — it added a phantom 50 pA SNARE current the operator never applies, and never matched the actual synaptic-gain coordinate. Corrected percentiles are worm 28% / fly 26% / mouse 38%: worm ≈ fly, both modestly significant, mouse marginal. Sections asserting the old reading have been removed; the historical record is preserved in a collapsed block near the foot of this page.

The validator operationalises this directly. Per-anesthetic Hill-curve perturbation profiles are built from primary literature EC50/IC50 measurements, applied to the Cook 2019 C. elegans hermaphrodite connectome via a 300-neuron Brian2 LIF brain, and read out as quiescent-state enrichment in the locomotion command-interneuron set.

Adversarial close-out (2026-06-12) — what the gated re-examination corrected

A preregistered, accept-either-way gated re-examination — every fix capable of deflating a claim, with leak floors and pre-written fail branches — certified the foundation and corrected several headline claims to exactly what the evidence licenses. Full record: SESSION_CLOSEOUT_2026-06-12.md.

Foundation — the operator is rank-2 (P18). A static + dynamic certificate proves the V1 perturbation operator collapses every anesthetic/genotype profile to two scalars: a uniform hyperpolarizing current total_pa and a single global synaptic gain snare_factor — QF = G(total_pa, snare_factor), max per-neuron non-uniformity 2.7e-16. This is the lens for everything else: the network cannot, by construction, distinguish which specific targets are engaged from the aggregate magnitude of two coordinates.

The Sub-Q1 trichotomy was a control artifact (P8). The original Match #2 null matched only an aggregate-pA scalar that (a) added a phantom 50 pA SNARE current the operator never applies and (b) never matched the actual snare_factor (≈0.75, a synaptic cut). A corrected two-coordinate null (verified to reproduce the operator to 7e-15) overturns the dramatic pattern:

Worm is not anchor-overfit and fly is not uniquely special; both show a modest, significant conserved-profile advantage (~26–28%), mouse marginal. A mean-field theorem (P3: per-neuron input CV² ∝ 1/K, slope −1.03, R² 0.999) shows mouse-at-median is a derived consequence of the structureless graph, not a failed test, and predicts the compression of all three toward the median.

Cell-type targeting — tested on a rank-lifted V2 substrate → NULL (P1_P2). The minimal-delta-V2 keystone replaces the all-ones broadcast with CeNGEN per-class expression vectors, raising operator rank so Match #3 becomes testable for the first time (built non-destructively; bit-identical to V1 when expression = ones). On worm, the conserved profile’s cell-type spread sits at the 46th percentile of magnitude-matched surrogates — lifting the rank did not reveal conserved-target cell-type specificity. (Fly cannot be tested — no Drosophila cell-type-expression atlas exists.)

The behavioral readout is not held-out-validated (P17). Scored against held-out neural + behavioral data (Kato; Atanas / NeuroPAL), the command-interneuron quiescent-fraction / 3 Hz immobilization readout is demoted to a network statistic, not validated behavioral quiescence — Kato labeled states support it but Atanas pose-correlation straddles zero and the threshold-valley test fails. The Paper-2 (anesthetic vs natural quiescence) bridge is blocked pending a re-grounded readout.

Molecular-layer occupancy — engagement robust, saturation inflated (P11). Vina Kd and the clinical EC50 are both aqueous-referenced; the shipped occupancy multiplied concentration by the partition coefficient (Kp = 250) without scaling Kd — a frame double-count, since Kp cancels when applied consistently. Corrected: multi-target engagement is robust (29/30 targets > 10% at 1× EC50) but saturation collapses (0/30 at > 90% vs 26/30 before) → moderate partial occupancy ~0.25–0.5, not saturating.

Calibration honesty narrowed (P7). Git archaeology shows the sat_pa magnitude ladder’s a-priori provenance is undecidable and the 3 Hz cutoff value was hardcoded after the preregistration. “One free parameter (α)” narrows to “one fitted scalar conditional on a ladder of undecidable provenance and a post-prereg cutoff value.” See PROVENANCE.md.

One constructive positive (P20). Three mutant phenotypes (gas-1, unc-79, unc-80) have literature magnitudes the rank-2 operator structurally cannot express — routed to the metabolic (Tier4) or cell-type-expression (V2) substrate. The model names epistasis it cannot capture and points to the substrate feature each needs. (Separately, P13-SOL28 confirmed the 4–6 class quorum is robust to the biophysically-corrected NCA current magnitude.)

Net. The core thesis survives — a multi-class quorum (4–6 classes; SNARE-or-Complex-I universal) reproduces MAC, and anesthetics discriminatively engage a biologically-motivated target panel. But the project’s own substrate, even after the rank-lift built to test it, does not establish that the specific cell-type-resolved target identities are load-bearing rather than one magnitude-matched ensemble.

The successor result (2026-09): a cross-state mechanism map, and an asymmetry nobody has asked about

The “Paper 2” question this project posed — does anesthetic-induced quiescence act on the machinery of natural reduced-activity states? — was blocked here on the demoted readout. It was answered by a different method: a frozen mechanism-repertoire × state map built directly from primary literature, with the search instrument itself calibrated against a measured recall standard.

Scope: 12 scored (state, species) rows × 12 mechanism classes defined by physical action = 144 cells, spanning sleep, anesthesia, dauer, L1 starvation arrest, cryptobiosis, and both seasonal hibernation and mouse daily torpor. Three-valued evidence logic (TESTED_POSITIVE / TESTED_NEGATIVE / NOT_TESTED, never collapsed to a binary), an orthogonal causal-role axis, and three phenotype axes that are never merged into one “depth” score. 162 verified PMIDs, every title fetched rather than recalled. Measured internal retrieval recall 1.00 on every row.

No mechanism has ever been shown to produce the anesthetized state. Exactly zero, across four rows and 25 populated cells, against 17 in the endogenous states. Not a coverage artifact: the most completely scored row in the table — mouse volatile anesthesia at 11/12 cells — is one of the four.

The matching claim about exit was falsified, deliberately. An earlier version of this table had zero ENABLES_EXIT across every endogenous state. Hibernation was then scored specifically because it was the state most likely to break that, and it did: intracerebroventricular prostaglandin E₂ provokes arousal from an established hibernation bout (PMID 11893609), and central adenosine antagonism reverses established torpor (PMID 22785425). The second is scored the same way as caffeine-accelerated emergence from anesthesia because it is the same pharmacology delivered the same way — scoring them differently would have preserved the zero and been indefensible.

What replaces it is sharper. All seven endogenous rows share a production mode; only hibernation has an exit literature. The discriminating variable is not biology but cost: interbout arousals consume 60–80% of a ground squirrel’s entire winter energy budget, and their function is, in the source’s own words, “unknown and disputed.”

Exit is studied where exit is costly. Production is studied where production is not already explained by the intervention.

Anesthesia has an exit literature because emergence is a clinical problem. Hibernation has one because arousal is an energetic problem. Sleep has none because waking up is taken for granted. Nothing produces the anesthetized state because the drug already answers the question. The shape of the literature follows the shape of the question — and hibernation supplies the natural experiment: two groups matched on production mode, differing in exit cost, differing in exactly the predicted direction.

A convergence nobody looked for. Two states, two sufficiency tests, same structure. Central A1-agonist activation in mouse produces profound hypothermia that is not torpor — it separates from the real state on temperature kinetics, heart rate, shivering, arrhythmia, recovery rate and transcriptional signature (PMID 28378088). Chemogenetic reactivation of anesthesia-activated ensembles produces immobility, discoordination and anti-nociception but not loss of righting (PMID 41573848). Independent literatures, independent methods: the components of a reduced-activity state can be produced separately, and producing some of them does not assemble the state. Neither paper cites the other.

The Paper-2 question, answered. Caspase ablation of parafacial-zone GABAergic neurons in mouse (PMID 40240218) confers durable behavioral resistance to both isoflurane and sevoflurane — ~50% higher likelihood of intact righting, dose-graded against surviving neuron count — and in the same animals, in the same paper, leaves NREM sleep amount, wakefulness and homeostatic drive unchanged with 70–80% of the population destroyed. A node of the sleep-promoting circuitry is required for the drug-induced state and dispensable for the endogenous one. Anesthesia co-opting sleep circuitry survives as a hypothesis, with a constraint it did not have before: the co-option is not symmetric.

The same paper carries a second dissociation its authors state explicitly — the behavioral endpoint moved and the EEG signature did not. A single collapsed “anesthetic depth” score would have erased both findings.

The closest anyone has come to the missing experiment, and it came apart. PMID 41573848 (preprint) captured isoflurane-activated neurons brain-wide with a Fos-driven DREADD system and reactivated them without anesthetic. Immobility, discoordination, anti-nociception and increased anesthetic sensitivity all appeared. Loss of consciousness, measured by righting reflex, did not. The gap above is therefore not “unattempted” — it is “attempted at ensemble level, and the components of the anesthetic state dissociated under the attempt.”

Agent specificity, established four independent times. Anesthetics are routinely discussed as a class; the manipulation evidence says they are not one. Ndufs4 knockout mice are 2.5–3× hypersensitive to isoflurane and halothane, 2× to propofol, and resistant to ketamine. Stomatin knockout shifts diethyl ether sensitivity by 12% and leaves halothane and isoflurane unchanged (a measured null with a positive internal control). EAAT3 knockout lowers isoflurane hypnosis threshold while leaving isoflurane immobility untouched. N-type Ca²⁺ channel knockout is more halothane-sensitive and less propofol-sensitive — oppositely affected, in one genotype. “Anesthetic sensitivity” is not one quantity; it is at least agent-specific and, within one agent, endpoint-specific.

The most confident claim in the dormancy literature is unexamined. A prediction frozen before any scoring said that for the deep states, the membrane-excitability classes would come back NOT_TESTED rather than TESTED_NEGATIVE — and called that the single most valuable possible output. It holds, now on four deep states: dauer, cryptobiosis, seasonal hibernation and mouse daily torpor are all empty across the entire merged excitability block, at recall 1.00. “Dormancy reduces neural excitability” is asserted constantly and has not been tested at the mechanism level in any of them.

Method result, portable beyond this project. The literature-search protocol used in every earlier run had a measured recall of 0.05 — NCBI eutils ANDs bare terms, so concatenated multi-concept queries demanded every term appear, and the founding worm anesthesia-genetics paper returned zero hits. Rebuilding it as boolean OR-groups with calibrated retrieval depth took measured internal recall to 1.00 (112/112), on a hold-out target that grew 170% during calibration. Two results generalize: broadening query vocabulary at fixed retrieval depth costs recall, and every residual miss must be attributed to vocabulary-versus-depth by direct diagnostic before either is changed. A null is only as interpretable as its instrument’s measured recall — which is why 53 of 120 cells can be reported as “searched and empty” at all.

Interactive: V3 ensemble explorer

Five tabs over the V3 worm ensemble + V4 fly cross-species ensemble. Pick a volatile to see its Hill-shaped dose-response against the published worm EC50; toggle a mutant to see the Hill curve shift left (hypersensitive) or right (resistant); compare the three Eger compounds side-by-side to see why the conserved-substrate model distinguishes anesthetics from non-immobilizers; open the perturbation profile to see which mechanism classes each compound engages with primary literature anchors; or open the Cross-Species tab to compare the same architecture’s predictions on the C. elegans (Cook 2019, 300 neurons) and Drosophila larva (Winding 2023, 2,952 neurons) connectomes side-by-side.

Architecture

  literature-grounded perturbation table  ──▶   300-neuron Brian2 LIF brain   ──▶   network-state metrics
  (per-(anesthetic, target_class)               (Cook 2019 connectome,                   • quiescent fraction
   EC50, max effect, Hill n,                     Loer & Rand 2022 NT identity,            • command-interneuron mean rate
   primary-PMID-anchored)                        per-edge CeNGEN Glu sign overlay)        • population state autocorrelation

• Perturbation profiles for 9 compounds × 8 mechanism classes built from Mihic 1997, Patel & Honoré 1999, Forman 1996, Stewart 2000, Hanley 2002, Lu 2007, Belelli 1997, plus Eger 2001 for non-immobilizers. Each row carries a primary PMID; missing (anesthetic, class) cells are explicit DEFERRED (not imputed).
• Mutant baseline shifts (Complex I rate factor for gas-1, gas-2, nduf-6, ndus-8, nuo-1 per Kayser 2001; NCA leak factor for unc-79, unc-80 per Sedensky 1992; synaptic weight scaling for Gαo loss-of-function in goa-1, dgk-1 per Lackner 1999, Nurrish 1999).
• One free calibration parameter (α = 0.13), tuned only against halothane WT (Crowder 1996, EC50 = 340 µM aqueous). All downstream gates use the same locked α. (P7 close-out caveat: this narrows to “one fitted scalar conditional on a sat_pa magnitude ladder of undecidable a-priori provenance and a 3 Hz quiescence cutoff whose value was hardcoded post-preregistration” — see Adversarial close-out.)
• Ensemble protocol: 60 s simulation × 5 seeds per (anesthetic, dose, genotype). 545 sims total; 8-core parallel; 23 min wall on RTX 4060 Ti host.

Validation gates

Calibration uses one anchor and one parameter. Held-out tests then run with no further tuning.

The four fold-errors above are point predictions from 5-seed ensembles. Worm isoflurane is the only one whose 95% bootstrap CI [276, 310] contains the published value; worm halothane CI [297, 334], fly halothane [343, 373], and fly iso [315, 328] all sit just outside published at current seed count (see Adversarial controls below). Mouse bootstrap CIs now computed (halothane WT 296.9 µM [289.9, 307.4], isoflurane WT 273.2 µM [268.8, 277.6]) — both published anchors sit just outside the tight CIs, extending the precise-but-just-outside pattern to all three organisms (published-inside-CI tally 1/6 across WT volatile anchors).

† Sign-only V5 baseline (no network integration) reproduces 32/35 mutant directions across worm + fly + mouse, indistinguishable from the validator. The 9 / 9 directional score is real but not network-distinctive — direction is carried by the perturbation table’s sign convention. Magnitude (the in-band column of the mutant detail table) is where the validator does work the sign baseline cannot.

‡ V7 Sub-Q2 stage-attrition analysis shows zero attrition from Stage 2 (isoflurane) to Stage 3 (Eger) across worm, fly, and mouse — every subset that produces correct halothane AND isoflurane EC50s at frozen α also correctly classifies the Eger non-immobilizers. Gate 4 is entailed by Gates 1 + 2 via the perturbation table’s structural sparseness on non-immobilizers; it is not an independent positive result.

Mutant directional accuracy — Gate 3 detail

WBPhenotype anchors: WBPhenotype:0001611 (halothane hypersensitive), WBPhenotype:0001618 (halothane resistant). Citation paths preserved per gene. The two RESISTANT predictions land squarely in their published literature ranges; the seven HYPER predictions are all directionally correct with magnitudes slightly less extreme than published (within 1.5× of upper bound).

Eger non-immobilizer specificity — Gate 4 detail

The Eger 2001 panel was designed to falsify lipophilic-pocket-fit theories of anesthesia: hexafluoroethane and trans-1,2-dichloroethylene have appropriate lipid solubility for Meyer-Overton anesthetics but produce no immobilization in mammals. Cis-1,2-dichloroethylene, the stereoisomer, is an anesthetic. Single-pose docking pipelines cannot distinguish them. The network-state validator can:

Same architecture, same α, no compound-specific tuning. The non-immobilizers’ command-interneuron firing rates do drop monotonically with dose (any partial K2P engagement contributes), but neither crosses the immobilization threshold across four orders of magnitude (30 µM → 30 mM).

V7 framing update — Gate 4 is entailed by Gates 1 + 2. V7 Sub-Q2 stage-attrition analysis shows zero attrition from Stage 2 (isoflurane held-out) to Stage 3 (Eger) across worm, fly, and mouse — every subset that produces correct halothane AND isoflurane EC50s at frozen α also correctly classifies the Eger non-immobilizers. The discrimination is entailed by the perturbation table’s structural sparseness on non-immobilizers (cis-DCE engages SNARE / Complex I / NCA; trans-DCE and hexafluoroethane do not), not by an additional capability of the integration substrate. The result is real; the credit belongs to the perturbation table’s sparseness rather than to network-state dynamics.

What this replaces

An earlier project iteration built a binding-occupancy → kinetic-shift → network → behavior chain driven by AutoDock Vina docking against AlphaFold structures of 30 C. elegans anesthetic targets. That pipeline failed three diagnostic tests:

• Vina ΔG was dominated by ligand chemistry: per-target regression of log(predicted Kd) on (clogP, MW) across 14 ligands gave median R² = 0.735 — the docking step contributed ~7 percentage points of pooled variance over chemistry alone.
• Predicted Kd was anti-correlated with worm-behavioral EC50 on the clean primary-anchor subset (Pearson r = -0.84, n=4, wrong sign).
• Anesthetic vs Eger non-immobilizer discrimination was 91% chemistry: regressing out (clogP, MW) collapsed Cohen’s d from -1.18 to -0.11 (Mann-Whitney p went from < 0.0001 to 0.50). Meyer-Overton beat the pipeline as a worm-potency predictor.

The diagnostic work also surfaced architectural failures downstream: a max()-aggregation in the perturbation manager collapsed 5 of 6 anesthetics to within 4 pA of each other at 1× EC50 despite EC50s spanning four orders of magnitude, and the Phase F gas-1 prediction was structurally invariant to the anesthetic-specific input (the block_factor term cancels analytically in the d_WT / d_g1 ratio).

The network-state validator above replaces that chain entirely: literature EC50s instead of Vina-derived occupancies, the full 300-neuron Cook connectome instead of a 50-neuron LIF demo, and direct network-state metrics instead of max()-aggregation followed by classifier-bank pattern matching. Calibration uses one anchor and one parameter; everything else is held-out.

Three-organism validation (V4 + V6) on a substrate-agnostic LIF integrator

The architecture transfers to Drosophila larva (Winding 2023 connectome, 2,952 neurons) and Mus musculus (generic LIF random graph, 3,000 neurons — no mammalian connectome required per V5 M2). Three independent organisms, three single-anchor calibrations.

What survived adversarial controls — V5 + V5+ + V7 findings

V5 M2 (connectome permutation) and V5+ (no-integration sign-only baseline + Meyer-Overton baseline), now joined by V7’s four pre-registered controls (Sub-Q2 mechanism-subset search, Sub-Q1 random-ensemble nulls, V5 M3 parameter sensitivity, V5 M4 anchor-swap cross-validation), jointly narrow what’s actually load-bearing in the architecture:

The substrate is largely interchangeable. V5 M2: fly result transfers to fully randomized graphs (P1, P2, P3 all pass Gate 1 at frozen α). Mouse V6 uses a generic random graph by design. Worm V3 needs at minimum cell-type-aggregate connectivity (P3 passes; P1, P2 fail). The “cross-phylum” framing was over-claiming organism-specific structural recovery.

Gate 3 directional accuracy is largely sign-propagation. A no-integration sign-only baseline matches the validator on 32/35 mutants (91% — identical to the validator). Network dynamics are decorative for the direction of mutant predictions; the perturbation table’s sign convention does the work.

Cross-organism MAC similarity at ~340-350 µM is largely Meyer-Overton. A lipid:water-partition baseline calibrated on halothane predicts halothane MAC perfectly (by definition) and matches the validator at order-of-magnitude across phyla. The “striking conservation” claim collapses to “lipid biophysics conserves across cells” (Meyer-Overton 1899).

V7 Sub-Q2 + Sub-Q1 sharpen the conserved-target claim and narrow it further. The architecture requires a 4–6 class minimum sufficient subset per organism (worm 5, mouse 4, fly 6); no 1-/2-/3-class subset passes anywhere. SNARE OR Complex I appears in 100% of all-stages-passing subsets across all three organisms (25/25, exceeding the pre-registered 75% threshold). Pre-registered P2 and P5 were falsified in the brittle-to-low-subset-counts direction. On the corrected random-ensemble null (Match #2b), all three organisms sit in a narrow band — worm 28% / fly 26% / mouse 38% — a modest, significant conserved-profile advantage in worm and fly and a marginal one in mouse. A mean-field theorem (per-neuron input CV² ∝ 1/K, slope −1.03, R² 0.999) shows the compression toward the median is a derived consequence of graph structure rather than a failed test.

V7 M3 + M4 — sensitivity and cross-cal hold. OAT: 9 load-bearing parameters across organisms, max halothane fold-error under any ±50% single-parameter perturbation = 1.44× (well inside 2× pass band). LHS (worm only): 95% CI [178, 448] µM with median 311.6 µM well-anchored — lower tail wider than pre-registered [200, 600] (M3c deviation) but falsification range [100, 1000] not crossed. M4 anchor swap: re-calibrating α on isoflurane MAC gives the same α value at grid resolution in all three organisms; halothane prediction with iso-anchored α is within 1.05–1.21× of published. Calibration generalizes anchor-to-anchor cleanly.

Where the architecture genuinely beats Meyer-Overton:

Honest single-line claim, post-controls

A literature-anchored multi-class Hill perturbation table, integrated through a substrate-agnostic LIF integrator, recovers held-out volatile EC50s ~3× tighter than Meyer-Overton across worm, fly, and mouse with one free parameter calibrated per organism. The architecture requires a 4–6 class minimum sufficient subset per organism (V7 Sub-Q2); SNARE or Complex I is universal across all 25 all-stages-passing subsets. Against a corrected two-coordinate magnitude-matched null, the conserved profile shows a modest advantage in worm (28%) and fly (26%) and a marginal one in mouse (38%). Cell-type-resolved target specificity, tested on a substrate rank-lifted expressly to make it testable, is null on worm (46th percentile) and untestable on fly. Calibration generalizes anchor-to-anchor at grid resolution (V5 M4). Eger non-immobilizer specificity is entailed by correct volatile EC50 prediction (V7 Sub-Q2 stage attrition), not an independent capability.

The architecture is largely substrate-agnostic. The mechanism map is what does the work, at the 4–6 class quorum scale. Cross-organism MAC similarity is consistent with Meyer-Overton (lipid biophysics) and not strong evidence of conserved targets on its own. The architecture’s distinctive contribution is within-organism volatile EC50 precision and mutant magnitude prediction. Whether the specific target identities are load-bearing rather than one magnitude-matched ensemble among many is not established on this substrate, in any of the three organisms.

V6 mouse scope: LRR / immobilization phenotype only. Higher-order mammalian features (cortical EEG burst suppression, NREM-like slow oscillations, gamma suppression, consciousness disruption) are NOT in the architecture and not claimed.

Use the Cross-Species tab to compare worm + fly + mouse halothane / isoflurane curves side-by-side.

V5 adversarial controls — bootstrap CIs and connectome permutation

Honest follow-on testing reveals where the V3/V4 result is genuinely structural and where it is over-determined:

Bootstrap 95% CIs (1000 resamples on 5-seed ensembles): predictions are precise but tight. The worm-isoflurane held-out test (predicted 291 µM, 95% CI [276, 310] µM) sits inside its CI containing the published 290 µM — the strongest of the four EC50 predictions. The other three EC50s sit just outside their 95% CIs (worm halothane CI [297, 334] vs published 340; fly halothane CI [343, 373] vs 340; fly iso CI [315, 328] vs 290). Fold-errors of 1.06–1.11× are real but not statistically consistent with published values to within seed noise.

Connectome permutation tests (Erdős-Rényi rewiring, configuration-model degree-preservation, cell-type block shuffle):

The “connectome-constrained” framing was over-claiming. The worm result requires at minimum cell-type-level structural connectivity (random rewiring breaks it; cell-type-block-preserving rewiring doesn’t). The fly result is NOT connectome-dependent — random graphs of comparable density also pass. The Eger specificity result is NOT a connectome claim in either organism — it’s a perturbation-table sparseness story.

The conserved-substrate hypothesis itself survives: the perturbation-table + network-integration architecture recovers behavioral phenotypes from molecular pharmacology in both organisms. What’s narrower than the original claim: the specific Cook 2019 / Winding 2023 wiring is NOT the load-bearing structural feature in either organism.

Where this went next

Paper 2 was answered, by a different method. The planned bridge experiment — anesthetic-induced quiescence versus natural reduced-activity states, tested inside this simulator — stayed blocked on the demoted readout and was never run here. The question was instead answered at literature scale by the cross-state mechanism map described near the top of this page, which is independent of this substrate and of the readout that blocked it. Its answer: the two states share circuit nodes, but the dependency is asymmetric — a node can be required for the drug-induced state and dispensable for the endogenous one.

Framework claims about the thermodynamic necessity of reversible reduced-activity states remain out of scope, and a separate audit has since sharpened why. The estimator the field uses for this — coarse-grained forward/reverse KL on neural recordings, reported as “entropy production” or “arrow of time” — requires Markovianity, stationarity, observable reversibility, structure-preserving lumping, and no hidden drive. All five fail for calcium imaging, and the failure is measurable: on 38 whole-brain C. elegans recordings the real irreversibility (0.0133 nats/frame) sits below the indicator-kinetics floor (0.0552) and barely above phase-randomized surrogates (0.0120), with 0/38 animals above the floor. The quantity is a statistic of the observed dynamics, not entropy production and not a bound on it.

A controlled follow-up (2026-09-21) found two confounds, with opposite signs. Latent processes with analytically known ground truth were measured with the identical estimator. The prediction I went in with — that a causal asymmetric filter manufactures irreversibility — was falsified: on a smooth reversible source, causal filtering lowers the estimate.

What actually drives it is finite-sample bias that scales with the latent timescale. On a provably reversible process with true entropy production of exactly zero, the estimator reads 0.1286 for fast dynamics and 0.0075 for slow — 17× across the range, with zero real dissipation at either end — and the effect decays with recording length, confirming it as finite-sample. Anaesthesia and deep sleep slow neural dynamics, so this confound carries the same sign as every published wake > anaesthesia result. A second confound, causal-filter asymmetry in the event-like regime, runs the opposite way: it inflates the anaesthetised state. The two compete rather than add.

The estimator’s positive control passes — a rotational Ornstein–Uhlenbeck process with genuinely broken detailed balance reads ~8× the reversible floor — so this is not an estimator that returns noise on everything.

Stated narrowly, and this is the whole claim: the estimator has a documented, sign-known, uncontrolled dependence on exactly the signal properties that define the states being compared. No study in that literature reports either control that would separate confound from effect — a timescale-matched reversible surrogate, or a symmetric-filter comparison. Both are cheap and run on data already published. (A 1996 Biological Cybernetics paper made the same argument form against chaos metrics in EEG; the objection was never answered for irreversibility metrics.)

Status

• V3 worm + V4 fly + V6 mouse three-organism validator shipped. All four gates pass in all three organisms with V5 + V5+ + V7 adversarial-control caveats per sections above.
• V7 closeout complete (2026-05-12). Pre-registration locked 2026-05-05 (hash 533b624a…, commit 4061f4f). Sub-Q2 + Sub-Q1 + V5 M3 + V5 M4 ran 2026-05-05 to 2026-05-08. Closeout docs: v7_final_summary.md, v7_redundancy_analysis.md, v7_preregistration.md.
• Adversarial close-out complete (2026-06-12). Preregistered, accept-either-way gated re-examination: foundation certified rank-2 (P18); deflations P7 (α honesty) / P8 (Sub-Q1 trichotomy) / P11 (occupancy saturation) / P1_P2 (cell-type-targeting worm-null) / P17 (behavioral readout); mean-field theorem (P3); robustness pass (P13-SOL28 nca quorum); constructive positive (P20 → V2/Tier4 routing); one fabrication caught before shipping. See SESSION_CLOSEOUT_2026-06-12.md. Audit harnesses + verdicts under AnestheticSimulator/audits/.
• Cross-state mechanism map complete (2026-09-17, extended 2026-09-21). 12 rows × 12 mechanism classes = 144 cells, all terminally disposed; 162 verified PMIDs, 0 unresolved; measured internal retrieval recall 1.00 (120/120) on every row; schema and completion validators both exit 0. This is the successor line and it is independent of the simulator substrate.
• Hibernation scored and the mandated class-coarsening run (2026-09-21). Hibernation was scored expressly to attack the headline and broke half of it; the coarsening merged classes 1–4 and reversed nothing. Both were the project’s top two self-declared debts.
• Page corrected (2026-09-21). The superseded worm/fly/mouse trichotomy was removed from the live claims and preserved in a collapsed historical block.
• Repo: AnestheticSimulator/ (public). Built at $0 external spend on an RTX 4060 Ti.
• Open work, simulator line: re-ground the behavioral readout (P17 demote → re-validate against a held-out target); build a Drosophila cell-type-expression atlas to make fly Match #3 testable — Fly Cell Atlas and FlyWire make this buildable rather than data-gated; write the P16 held-out structure→activity scoring loop (strict shuffle + fast gates done, heavy path stubbed); regenerate Wave-P occupancy magnitudes under the corrected reference frame (P11); per-mutant Complex I factor refinement (nduf-6 / ndus-8 / nuo-1 share an estimate).
• Open work, mechanism-map line: measure inter-rater reliability on causal-role assignment — now the top methodological debt; add a thermal phenotype axis; commit a stress-induced-sleep row; score the remaining 7 of 19 committed rows.