Method

Version 0.4.10 — draft. Status: the reproduction gate passed on 2026-09-18 at attempt 3 (gate/RESULT-3.md). Seventeen readings are issued — six computed here (mouse and human, Types A/B), two paired contrasts (Type F), and nine values cited from the published literature (Type D), spanning human TMS-EEG and intracranial stimulation, rat, and ferret cortical slices. Three distinct scales are represented and may never be mixed; §3 governs which. None of the seventeen is conformant under §12; the count is published at zero because that is the only way its rise means anything.

To be maintained by the Nooscope Foundation, which is not yet formed; until it is, this document is maintained by Mind Matter Analytics Inc. Licensed CC BY 4.0. Canonical copy: this document as published at nooscope.io/method. Cite by version, not by date.

---

0. Status and posture

This document is published before any result. That order is deliberate. A measurement standard written after its first numbers exist is a standard fitted to those numbers, and nobody outside the room can tell the difference. Fixing the method first, in public, with a version history, is the only way the numbers that follow can be checked rather than merely believed.

As of version 0.3.7:

• Sixteen readings have been issued. The first two, NS-0006 and NS-0011, came from the deposit the gate was run against; issuing them exposed two defects in the record format, corrected in the same version: a reading carries one state and one value, so the planned "wake versus isoflurane" entry was not a reading but two; and a rejection threshold derived per session is a range across an aggregate reading, not a constant. NS-0012 and NS-0013 are the first human computed readings; NS-0015 and NS-0016 are paired contrasts; NS-0017 through NS-0025 are cited from the published literature. NS-0026 is the first reading issued to test another reading rather than to add to the register. The header above governs; this list is the history.

• The reproduction gate (§9.1) was attempted on 2026-09-17 and did not pass. It reproduced the direction of the published result but not its magnitude, and it found three errors in §2.1 of this document, which 0.2.0 corrected. The attempt is published in full at gate/RESULT.md, including what failed.

• It was attempted again on 2026-09-18 over the whole deposit and did not pass. The primary wake-versus-anesthesia contrast reproduced within error; the superficial-stimulation stratum did not. That attempt found three further errors, which this version corrects: stimulation intensity was not a required field, channel and trial rejection were unspecified, and §9.1 itself named no pass criterion. It is published in full at gate/RESULT-2.md.

• §9.1 now states a pass criterion. It was written after attempt 2's numbers were known, which is recorded there as a defect in its provenance; attempt 2 does not pass under it, and it binds from attempt 3 forward.

• Attempt 3 passed it, 5 of 5, on 2026-09-18 (gate/RESULT-3.md). It was the first attempt run after the criterion existed. The pass came from honoring the deposit's own channel validity flags, which attempts 1 and 2 ignored; the fix introduced no tunable parameter. Four of five measured quantities moved toward their published values and one moved away, which is recorded there.

• The method was applied to human data for the first time on 2026-09-18 (gate/RESULT-PRIOS.md): seven epilepsy patients stimulated on the same implanted grid awake and again under propofol. It separated the two states at 168 of 254 stimulation sites, in all seven patients — statistically solid and far weaker than the same pipeline's mouse contrast, which separated at 21 of 24 sessions at roughly 1.9 times the ratio. A predicted ordering among the anesthetic states did not appear, and the first explanation offered for that was tested and failed. The run found two defects, which 0.3.1 corrects: a between-state comparison had no rule holding its channel set constant, and none holding its trial count constant.

• The register could not express its own strongest evidence, which is what issuing the human pair revealed. Both state contrasts are held as two readings whose intervals overlap; the paired comparison that carries the result lived only in a result document. 0.3.2 adds Type F, the paired contrast (§3.2), and two were issued under it — NS-0015 for the human pair and NS-0016 for the mouse pair, which had the same defect from the beginning.

• The leading explanation for that weak human result was tested and abandoned. Local recording coverage was asserted in three documents here as the reason. It failed in the mouse deposit over the range that deposit can test, and then failed in the human deposit itself across 248 stimulation sites at a permutation p of 0.926. What replaced it was already published: the intracranial paradigm compresses the contrast relative to TMS, 1.9 against 2.8, in one framework on the same measure (Comolatti et al. 2025, entered as NS-0017 to NS-0022). See gate/RESULT-COVERAGE.md and gate/RESULT-COVERAGE-HUMAN.md.

• §5 (in vitro) is no longer entirely unexercised: NS-0025 enters a published ferret-slice value, on a third scale that may never be mixed with this one. §6 (artificial systems) remains specified and unexercised.

• Recording geometry was not a recorded field, which 0.3.3 corrects with §3.3. It surfaced twice: a second human deposit records from depth electrodes rather than a subdural grid, and the pipeline's channel filter had the grid type hard-coded, so the depth deposit computed to nothing without an error. Every Type A or Type B reading now records its geometry, absolute values are not compared across geometries without a stated warrant, and Type F is exempt because geometry is held constant inside a pair. A third deposit then turned out to mix both geometries across its own thirteen patients (gate/RESULT-STREEF.md), which the field now makes visible rather than silent.

• Stimulation waveform and repetition rate were not required fields, which 0.3.4 corrects in §2.1. Forty-five cells matched on patient, contact pair, current and state measured a 2.07-fold median difference between two stimulation protocols — larger than either state contrast in the register (gate/RESULT-CHOCS.md). The standard fixed intensity and left the shape and rate of the pulse unstated, which admitted a source of variation larger than the signal being measured.

• The register's first external check on an absolute value was run, and it held. NS-0012 issues human awake intracranial complexity at 32.26 from seven patients at one center; three patients at a different center, computed by the identical pipeline at matched geometry, protocol, current and trial count, returned 35.83 — inside the issued interval (gate/RESULT-STREEF.md). The same run found that between-patient spread inside one cohort is wider than the gap between cohorts, and that an analysis had read the wrong one of two similarly named result columns and nearly published a failed replication that had not occurred.

• A geometry survey was run across every verified deposit, and eight of the nine readable ones are heterogeneous, which 0.3.5 addresses by extending §3.3 to the mixed implant. Four subjects in ds004774 and the only subject in ds003708 carry a subdural grid and depth electrodes at once; five of eight in ds004696 carry scalp EEG beside depth electrodes. In that case both channel filters return real channels and compute without error, so the earlier guard — which catches a geometry that is absent — cannot see it. The only homogeneous deposit is ds004370, the one already computed, which is why nothing caught this earlier.

Nothing here should be read as a claim that the method works. It is a statement of what will be done, in enough detail that someone else can do it and show that it does not.

---

1. Scope

The Nooscope measures one thing:

the complexity of a system's causal response to perturbation.

It does not measure consciousness. It measures a quantity which, in human cortex under controlled conditions, separates states we independently know to be conscious from states we independently know to be unconscious, with published sensitivity and specificity. Every application beyond that human cortical case is extrapolation. This document requires each extrapolation to be labeled as one.

1.1 In scope

• Human cortex perturbed directly — transcranial magnetic stimulation with EEG (TMS-EEG), or single-pulse electrical stimulation with intracranial or high-density scalp recording (SPES).
• Human cortex assessed from spontaneous recordings, as an explicitly weaker proxy (§3, Type C).
• Non-human nervous systems where an equivalent perturb-and-record protocol exists or can be run.
• Neural tissue in vitro on multi-electrode arrays.
• Artificial systems, by a separate instrument that does not share this scale (§6).

1.2 Out of scope

The following are outside the method and may not appear in any reading, commentary or chart issued under it:

• Any claim about the presence, absence, character or moral weight of subjective experience.
• Any diagnosis, prognosis, treatment recommendation or clinical decision support.
• Any statement about a system that was not actually measured, including inference from a similar system.
• Any ranking of subjects by moral status, sentience, or worth.

---

2. The quantity

Three steps.

Perturb. Deliver a brief, well-localized, harmless input to the system and record the time of delivery precisely.

Record. Capture the system's response across as much of it as the instrumentation allows, at a sampling rate sufficient to resolve the propagation of that response.

Compress. Measure how much the recorded response resists lossless compression, relative to the pre-stimulus baseline.

The third step is the substance. A system that is not integrated returns a response that stays local. A system that is integrated but not differentiated returns a globally synchronized, stereotyped response. Both compress to almost nothing. A system that is both integrated and differentiated returns a response that spreads widely, arrives at different places at different times, and does not repeat itself — and that resists compression. The degree of resistance is the number.

2.1 Reference computation

For version 0.1.0 the reference computation is PCIst — the state-transition Perturbational Complexity Index of Comolatti et al. (2019; PMID 31133480) — as implemented in the reference library cited in §11.

Defect of record, 0.4.0 (gate/RESULT-DUAL.md). The sentence above defines the reference computation by pointer rather than by algorithm, and the published paper's openly reachable text does not supply the operative details either: not the signal-to-noise definition or the role of k, not the range the distance threshold is swept over, not how transitions are counted from a thresholded matrix, not the normalization applied before the baseline is subtracted. A reading under this standard is therefore not reproducible from published sources — only by running one specific repository, which is not what §12 invites a third party to do.

A second implementation, written from everything the specification actually says and resolving the five open points explicitly, produced values for the same evoked responses differing from the reference by 3.76× to 35.67×, with ICC(2,1) = 0.043 — no agreement — while ordering every one of six animals identically. The disagreement is not a scale factor: removing the median 8.32× leaves a 9.49× residual spread.

Closed in 0.4.1. spec/PCIST-1.0.md is a normative annex to this section and specifies the computation in full. An implementation written from that annex alone reproduces the reference on all seventeen register cells at a worst relative difference of 0.00e+00, ICC(2,1) = 1.0000. spec/vectors/VECTORS-1.0.json supplies eight conformance vectors at a 1e-6 tolerance which also name the clause a failing implementation most likely got wrong. A reading declares computation.spec_version alongside the library, version and commit.

PCIst is used rather than the original Lempel-Ziv formulation (Casali et al., 2013) because it requires no source reconstruction, no statistical thresholding against a surrogate distribution, and no binarization, which removes three degrees of analyst freedom from the pipeline.

Every reading must declare its parameters, and the parameter set is per-paradigm. Version 0.1.0 gave a single set, taken from human TMS-EEG, with a note that other paradigms differ. The first reproduction attempt showed that is not good enough: applied to mouse SPES the human set understated both the values and the separation. Each paradigm gets its own named, versioned set.

human_tms_eeg_v1:
  baseline_window: [-400, -50]   # ms relative to stimulus
  response_window: [0, 300]
  k: 1.2                         # signal-to-noise scaling factor
  min_snr: 1.1                   # minimum SNR for a retained component
  max_var: 99                    # % variance retained in dimensionality reduction
  embed: false
  n_steps: 100
  resample: 725                  # Hz - REQUIRED, see below

mouse_spes_v1:                   # Claar, Rembado et al. 2023
  baseline_window: [-800, -2]
  response_window: [2, 800]
  k: 1.2
  min_snr: 1.1
  max_var: 99
  embed: false
  n_steps: 100
  resample: 500                  # Hz - provisional, see below

resample is required and may not be omitted. The reference computation builds T×T recurrence matrices, so cost goes as the square of the samples in the response window: an 800 ms window at 2500 Hz demands roughly 9 GB and does not run at all. More importantly the rate changes the answer — 250 Hz against 500 Hz moved group means by about 25% in the gate attempt. A reading that does not state its resample rate is not reproducible, and the validator rejects it.

The 500 Hz figure in mouse_spes_v1 is provisional: the authors' own rate was not stated in the methods available, and it was chosen here to hold the response window near the sample count the human set produces. It is flagged as an open question rather than presented as settled.

Stimulation intensity is a required field of the measurement, not of the analysis. The second reproduction attempt measured PCIst against stimulation current across eighteen sessions and found an inverted U spanning a 5.4-fold range — 59.62 at 20 µA against 11.11 at 100 µA, under deep stimulation in the awake state. Two readings taken on the same animal in the same state at different currents differ by more than wake differs from anesthesia.

A perturbational reading must therefore record the stimulation modality, the intensity with its units, the site, and the depth, and readings taken at different intensities are not comparable and may not be pooled. Any aggregate across intensities must say so and must report the range pooled. The validator rejects a Type A or Type B reading that does not carry a stimulation record.

This is a worse defect than the unstated resample rate that the first attempt found, because resample is a property of the analysis and can be recomputed from stored data, while intensity is a property of the measurement and cannot be recovered after the fact.

Stimulation waveform and repetition rate are required fields on the same footing as intensity, and readings taken under different stimulation protocols may not be pooled or compared as absolute values. Version 0.3.3 fixed intensity and said nothing about the shape or rate of the pulse. A deposit carrying two protocols on the same patients settled the question: in ds003848, 45 cells matched on subject, contact pair, current and state — every one of them awake, at 2 mA, on the same implanted electrodes — measured a median paired ratio of 2.07 between monophasic stimulation at 0.2 Hz and biphasic stimulation at 1 Hz, higher under the biphasic protocol at 35 of 45 cells, sign test p = 2.5e-4 (gate/RESULT-CHOCS.md).

That ratio is larger than either state contrast the register carries: 1.21 for human wakefulness against propofol (NS-0015) and 1.89 for mouse wakefulness against isoflurane (NS-0016). Changing the stimulation protocol moved this quantity further than losing consciousness does. A standard that fixes intensity while leaving waveform and rate unstated admits a source of variation larger than the signal the instrument exists to detect.

A perturbational reading therefore records, in addition to intensity:

stimulation:
  waveform: monophasic | biphasic     # REQUIRED
  pulse_width_ms: null                # REQUIRED, with units
  repetition_rate_hz: null            # REQUIRED, with units
  protocol_name: null                 # the deposit's own label, where it has one

Two qualifications are owed. The evidence is two patients, at one current, on depth electrodes, and waveform and rate differ together in this deposit, so it cannot say which of the two carries the effect or whether they interact. And the requirement does not depend on the size of the effect being settled: a field that is recorded costs nothing and can be checked later, while a field that was never recorded cannot be recovered, which is the same argument that made intensity required.

Readings issued before this version that do not carry these fields are not invalidated, but they may not be compared as absolute values against readings taken under a different protocol, and the register records that limitation against them.

Channel validity and trial rejection are part of the computation and must be stated. Version 0.2.0 was silent on both, and the reproduction pipeline consequently used every channel in the deposit — including the channels the depositors had marked invalid. A method that ignores the data's own quality flags is not reproducing the published analysis; it is running a different one.

rejection:
  channels: source_flags          # honor the deposit's validity flags where present
  trials: amplitude               # reject trials exceeding the threshold below
  trial_threshold_uv: null        # REQUIRED when trials: amplitude
  report: required                # counts excluded, per category, in the record

Every reading records how many channels and how many trials were excluded and under which rule. A reading that excludes nothing states that explicitly. Silent rejection and silent non-rejection are equally unreproducible.

A comparison between two states holds the channel set and the trial count constant across them. Both of these were found by applying v0.3.0 to human data (ds004370, seven patients recorded awake and under propofol on the same implanted grid), and neither is optional.

Channels. Deposits flag validity per recording, not per subject, so the same patient can have 73 of 80 channels marked good in one state and 70 of 80 in the other. Honoring each recording's own flags then compares two states measured on two different arrays — the same class of error as running the array with the invalid channels left in. The comparison uses the intersection of the sets marked good across every recording entering it, and records what the intersection cost.

Trials. PCIst is computed on a trial average, so a condition with fewer trials has a noisier average. Noise lowers the number. Where the two states were not recorded with equal trial counts — and they usually were not, because the awake recording is an unhurried clinical session and the anesthetized one is inside somebody's operating-room schedule — the larger condition is subsampled to the smaller, per stimulation site, by a rule that is a function of the deposit rather than of a random seed. The unmatched computation is run beside it and both are reported.

This matters more than it sounds. Unequal trial counts depress the complexity of whichever condition has fewer trials, and in this literature that is almost always the unconscious one. An uncorrected comparison therefore produces the expected result for a reason that has nothing to do with consciousness, and the number of published comparisons that state their trial counts is smaller than the number that do not.

matching:
  channels: intersect_across_states   # required for a between-state comparison
  trials: match_per_site              # subsample the larger condition
  trial_selection: evenly_spaced      # deterministic; no seed to report
  report: required

On average referencing. Version 0.1.0 prescribed avgref: true for SPES with referential recording. The gate tested that prescription against real data and it made the result worse — separation fell from 1.75 to 1.28, and subjects separating in the published direction fell from four of six to three of six. The prescription is withdrawn. avgref defaults to false and is set per paradigm only where evidence supports it. A parameter written into a standard before any data has been run through it is a guess wearing the clothes of a specification.

A reading that departs from a documented set must state the departure and the reason in its provenance record, and the departure must be visible in the diff of the repository at the version the reading was issued.

2.2 What the number is not

PCIst is a scalar with no natural unit and no upper bound. It is meaningful only by comparison: against the same subject in another state, against a published benchmark distribution, or against the empirical cutoff established in human cortex.

That sentence was an argument until 0.4.0 and is now a measurement. Across two implementations of this same computation, the magnitude of a reading is not reproducible — ICC(2,1) = 0.043, limits of agreement from 3.27× to 31.49× — while the ordering of states within a subject is reproduced 6 of 6, and the magnitude of a within-subject ratio disagrees by up to 2.5× (gate/RESULT-DUAL.md). Comparison is therefore not the preferred use of this value. It is the only use that survives a change of implementation, and a paired contrast is portable in its direction rather than in its size.

The empirical cutoff is PCI* = 0.31 (Casarotto et al., 2016), established in a benchmark population of subjects whose conscious state was independently known. That cutoff is a property of that population and that recording setup. It is not transferable to other species, other tissue, other instrumentation, or other analysis parameters, and a reading may not apply it outside the human cortical case without an explicit calibration of its own.

---

2.3 The matched null

Added in 0.4.1, in response to gate/RESULT-FLOOR.md.

The reference computation does not return zero when there is nothing to measure, and what it does return is a property of the recording rather than of the method. Measured on six sessions of one deposit — same species, same rig, same parameter set — the value returned on randomly timed epochs ranges 2.80 to 34.05, a factor of 12.2. It cannot be looked up, and it cannot be inferred from a benchmark population.

Every perturbational reading issued under 0.4.1 or later carries a matched null.

matched_null:
  method: random_onset          # epoch onsets drawn from the same recording
  clearance_ms: 1000            # minimum distance from any real stimulation
  n_draws: 10                   # independent sham averages
  value_median: null            # REQUIRED
  value_p95: null               # REQUIRED
  ratio_to_reading: null        # REQUIRED, the reading divided by value_median

The null holds everything constant except the thing being measured: the same admitted channels, the same trial count per state, the same parameter set, the same averaging and baseline correction. Only the onsets change. A null computed on synthetic noise does not satisfy this section — white noise is spectrally flat and uncorrelated across channels, which inflates the value several-fold and is the error gate/RESULT-FLOOR.md caught in its own first pass.

The ratio is displayed on the face of the reading, in the same visual field as the value, exactly as the modality tier of §3.4 is. A reading whose ratio is at or below 1 is a valid record and is not issuable as a value: the instrument did not distinguish the perturbation from its own behavior on that recording, which is a result about the measurement and is recorded as one.

Two consequences follow for comparisons. A separation between two states whose denominator sits near its own null is a lower bound, not a measured ratio, and says so. And a reading may not be compared as an absolute value against one whose null was not measured, because the two are not known to be above the same floor.

A deposit that cannot supply the null cannot carry the reading. Added in 0.4.10, after the case that made it concrete. A perturbational deposit that archives only stimulus-locked epochs — and not enough continuous record for a sham onset at the issued parameter set — cannot satisfy this section, however good the experiment behind it. The arithmetic is the test: one sham onset needs the response window, the baseline window and the clearance, summed. Where the archived epoch is shorter than that sum, there are no onsets to draw, and a second recording from the same subject and session does not substitute unless it matches the admitted channels and the parameter set as well.

This is a defect of deposition, not of the science, and it is recorded that way. The register names such a deposit as refused with the arithmetic shown, publishes the values it would have issued alongside the refusal so the work is not lost, and asks the depositors for the continuous record. Where that record arrives, the reading is issued normally. The corollary is a request this register makes of anyone archiving a perturbational recording: deposit the continuous data, not only the epochs. An epoched archive is sufficient to reproduce the authors' own analysis and insufficient to measure the floor that analysis rests on.

Readings issued before 0.4.1 are not edited (§8). Their measured ratios are recorded against them in the register, and the conformance report shows the gap.

---

3. Reading types

Every reading carries a type. The type governs what may be said about it and, critically, whether it shares a scale with any other reading.

3.2 Why Type F exists

Added in 0.3.2. A reading carries one state and one value — that rule was established in 0.3.0, it is correct, and Type F does not relax it.

What 0.3.0 did not anticipate is the consequence. The register's two state contrasts are each held as two readings whose intervals overlap, because the between-subject spread is far larger than the between-state difference. Read on their own, those pairs say nothing. The finding in both cases is paired: wakefulness exceeded anesthesia in 21 of 24 mouse sessions and in 168 of 254 human stimulation sites, at sign-test probabilities of 1.4 × 10⁻⁴ and 1.5 × 10⁻⁷. Neither number could be written anywhere in the register. Both lived only in a result document, which is not a record anyone can cite, supersede or audit.

A register whose strongest evidence is not expressible in it has a format defect, not a presentational one. Type F is that quantity:

• Its value is a ratio, dimensionless, and therefore not on the Type A scale and never comparable to one.
• It names the two readings it derives from, which must both be issued, must be Type A or B, and must come from the same source deposit. A contrast whose inputs are not in the register is not auditable, so the two absolute readings are retained rather than replaced.
• It states the pairing unit — the session, the stimulation site, the subject — because a contrast is only as paired as its unit, and an unstated unit permits the most common error in this literature.
• Its interval is a bootstrap over the pairing unit, not a widener, because the systematic differences §5 exists to widen for are shared by both arms and cancel in the ratio. That is the point of a paired comparison and the reason the contrast is worth more than either arm.

A Type F reading may not be used to claim that a system is conscious, or that either state is. It says one thing: on this deposit, in these units, the quantity was higher in one state than the other, by this much, this often.

3.3 Recording geometry

Added in 0.3.3, after the reading code was found to be making this decision silently.

Type A and Type B readings share a scale. They do not automatically share a recording geometry, and a PCIst computed from a subdural grid is not interchangeable with one computed from depth electrodes.

The two arrangements sample different things. A subdural grid lies on the cortical surface and covers a contiguous patch of it; depth electrodes pass through tissue and sample points along their trajectories, many of them in white matter or deep structures that no grid reaches. PCIst reduces the response across whatever array it is given and counts state transitions in the reduced space. The number it returns therefore depends on what the array sampled, not only on what the tissue did.

The register had no provision for this and the code had an opinion: the channel filter accepted ECOG and nothing else. That was never a stated decision. It happened to match the first deposit computed (ds004370, subdural grid) and stayed in the reading path. On ds003848, which records from depth electrodes, it silently produced zero usable channels and the deposit computed to nothing. Had it been written the other way, values from depth electrodes would have entered the register beside grid-derived ones with nothing marking the difference, which is the worse of the two failures.

Accordingly:

• Every Type A and Type B reading records its geometry — ecog_grid, seeg_depth, scalp_eeg, mea, probe_linear, epidural_array — as a required field, not a caveat.
• Absolute values are compared only within a geometry. Setting a depth-electrode value beside a grid value, or either beside a scalp TMS-EEG value, requires a stated warrant in the reading's own record. Absent that warrant they are as separate as the three scales in the table above.
• A paired contrast (Type F) is exempt, because geometry is held constant within the pair by construction. A contrast between two states, or between two stimulation protocols, measured on the same contacts in the same subject, is unaffected by which arrangement those contacts belong to. This is why a contrast can be computed on a deposit whose absolute values cannot be set beside the register's existing ones.

The general rule this instance belongs to: where the code must choose, the document chooses first. A filter that decides what enters the register is a method provision whether or not anyone wrote it down, and the cost of leaving it in the code is that it can be wrong in either direction without saying so.

One implant may carry two geometries, and that is the harder case. Added in 0.3.5. Everything above contemplates a deposit whose subjects differ from one another. A survey of every verified deposit found a second arrangement (gate/verify_geometry.json): in some subjects a subdural grid and depth electrodes are implanted in the same patient at the same time, and in others scalp EEG is recorded simultaneously with depth electrodes.

That case defeats the guard the previous paragraph implies. Where a geometry is simply absent, a filter asking for it returns nothing and the failure is detectable. Where both are present, a filter asking for either returns real channels and computes without error — and the two answers differ, for the same patient, in the same state, under the same stimulation. Nothing in the output records which half of the implant produced the number.

Accordingly:

• Computing one geometry of a mixed implant is permitted and must be declared. The reading records the geometry it used, as above, and states in a caveat that the subject carried another and how many channels of it were excluded.
• The channel sets of two geometries are never pooled into one value. They sample different tissue on different scales; a single PCIst across both is a number about the implant, not about the brain.
• A deposit's geometry is surveyed before any of its subjects is computed, not discovered when a run fails. The survey is cheap — one channels table per subject — and the failure it prevents is not.

The survey's own finding is worth recording plainly: of the ten deposits this register had verified as carrying stimulation, nine have readable channel tables and eight of those nine are geometrically heterogeneous. The single homogeneous deposit is the one the register had already computed. Verifying that a deposit carries perturbation with per-trial current, which §9 and the sourcing checklist required, does not verify that its subjects share a recording geometry. The two checks are independent and both are now required.

3.1 The basis for Type C

Type C exists because the datasets that matter clinically are spontaneous recordings, and a standard that refused to touch them would be silent on the case it most needs to address.

It rests on Breyton et al. (2025), which computed four measures on resting-state EEG — dynamic functional connectivity fluidity, Lempel-Ziv complexity, functional repertoire size, and global activation potential — in fifteen subjects under propofol, xenon and ketamine, each also assessed by TMS-EEG. Fluidity and functional repertoire size separated conscious from unconscious states at accuracy matching PCI, and correlated with PCI across subjects and conditions.

Two things follow, and the second is the one that gets forgotten:

1. A spontaneous proxy is defensible. It was not invented here to make a dataset usable.
2. Correlating with PCI is not being PCI. n = 15, three anesthetics, healthy subjects. A proxy validated in that population does not carry its validation into brain injury, sleep, or any other population without its own calibration. Every Type C reading names the proxy measure it used, and its interval widens under proxy_substitution unconditionally.

The ketamine arm is why this matters. Those subjects were behaviorally unresponsive and reported vivid experience, and the spontaneous measures tracked the experience rather than the behavior. That is the case this whole field exists for — and a reminder that a measure which had happened to track behavior would have been wrong in exactly the way that matters.

This table is the most consequential page of the standard. The temptation in this field is to put every kind of system on one axis because it makes a compelling picture. Types C, D and E do not belong on the Type A/B axis. Any chart, table or figure issued under this standard that displays them together must mark the boundary visibly and repeat this caveat in its caption. A chart that does not is not issued under this standard, whatever it is called.

---

The set above is closed. Until 0.3.7 it ended "or another named arrangement", which put a free-text field in the middle of a decision INSTRUCTION.md 3.1 describes as mechanical — a rater meeting an unfamiliar array could name it whatever seemed reasonable, and two raters could name the same array differently while both following the instruction. An arrangement outside the set now blocks the reading and is added by amendment, which is slower by design: the alternative is a geometry vocabulary that grows silently and cannot be grouped or compared after the fact.

epidural_array was added in 0.4.3, for a surface array implanted under the scalp against the skull or dura. It is distinguished from scalp_eeg, which is applied non-invasively outside the skull, and from ecog_grid, which lies on the cortical surface inside it. The census maps an EEG channel type to scalp_eeg by default, so without this member a mouse epidural implant and a human 10-20 scalp montage were the same geometry — the precise conflation this section exists to prevent.

Defect of record, 0.4.3. The sentence that follows said probe_linear was the arrangement "the register's two mouse readings were taken on." That is false. Every reading path in this repository reads acquisition/ElectricalSeriesEEG, the thirty-contact epidural surface array; the deposit files the register computed from contain one device, one electrode group, no units table, and every probe_vertical_position set to −1. There is no shank probe in them. NS-0006 and NS-0011 carry geometry: null, having been issued before the field existed, so no reading was mislabeled — but a member was added to a closed set in 0.3.7 on a premise that was never checked, and the arrangement actually in use had no member until now. probe_linear is retained: it describes a real arrangement and the wider deposit does carry Neuropixels probes. Its justification is corrected here rather than removed.

probe_linear was added in 0.3.7, for penetrating silicon probes with contacts along a shank. It is distinguished from mea, which is planar and used for slice and culture work, and from seeg_depth, which penetrates but at clinical macro-contact scale. Those are three different sampling geometries and the number PCIst returns depends on which one was used, so collapsing them would defeat the purpose of recording geometry at all.

NS-0006 and NS-0011 were issued before §3.3 existed and carry no geometry field. They are not edited, per §8. The gap is visible in the conformance report rather than repaired out of sight.

3.4 Measurement modality tier

Added in 0.3.6, adapted from the objectivity ceiling of an evidence-grading standard built in a separate workstream.

A reading derived from instrumented neural measurement and a reading derived from behavioral report are different evidentiary objects. Treating them as the same kind of thing with different error bars is the mistake this section exists to prevent, and a footnote is not where that distinction belongs. Every reading carries a modality tier on its face, wherever its value is displayed.

The tier is not a restatement of the reading type. The type records provenance — whose data it was and who computed it. The tier records what kind of measurement it is. A Type D published value may be M1 or M4 depending on what the third party actually measured, and that difference constrains what may be said about it far more than its provenance does.

The ceiling is applied without exception, and it is applied in addition to §7, not instead of it. Where a trial reports both an instrumented and a report-dependent endpoint, they are separate readings with separate tiers.

The reason for capping M4 is specific rather than a general suspicion of report. Where the quantity depends on what a subject says about their own state, the thing being measured and the instrument measuring it are not separable, and no feature of the design separates them. That is a real limit on what the evidence can establish, and capping it is the honest representation of that limit rather than a penalty.

---

3.5 Recording grain

Added in 0.3.9, in response to gate/RESULT-REPLACEMENT.md.

§3.3 fixes the arrangement of the recording. This section fixes its resolution: how much substrate each admitted channel integrates over. The two are different, and until 0.3.9 only the first was specified.

The reason is measured. A linear-Gaussian system was replaced node by node with functionally equivalent two-unit surrogates — an exact replacement, preserving the system's causal effect at every retained variable to 4.8 × 10⁻⁷ and its spectrum to 10⁻⁸. Read at its original variables, every screened measure returned its original value. Read at the units the replaced system actually has, the same measures moved monotonically to between 0.53× and 1.79× of it. Both are factors comparable to the wakefulness-against-anesthesia contrasts this register publishes — 1.21× in human, 1.89× in mouse. A control recording the unchanged system at the same inflated channel count moved every measure by ≤ 5%, or in the opposite direction, so this is not channel counting.

Every reading issued under 0.3.9 or later declares its grain:

A reading whose grain cannot be established from its source is not issuable. It is not issued with the field omitted and it is not issued with a guess.

Comparison rule. Two readings may be compared as absolute values, and subjects combine into one reading, where the ratio of their nominal pitches is at most 1.3. Beyond that ratio the §7 caveat is required in the same visual field as the value.

The tolerance is measured, not provisional (gate/RESULT-DECIMATION.md). The same recordings were read at successively coarser subsets of their own contacts, with extent held constant so that spacing moved alone. In mouse surface EEG, 16 state-cells: a 1.26× pitch ratio moves the median reading 8%, 1.66× moves it 17%, and 2.26× moves it 32%. In one human subdural grid the effect is not detectable between 1.77× and 4.17×. The tolerance is set where the measured effect is smaller than the smallest state contrast this register publishes — wakefulness against propofol at 1.21×. The human arm is a single subject and is a weak basis for a negative; the tolerance is set from the mouse.

Why this section has been revised twice in a day. 0.3.9 required exact equality of pitch because nothing was known; 0.3.10 compared nominal rather than measured pitch because the first rule made nearly every reading in the register unissuable; this version replaces exact equality with a tolerance because the effect was finally measured on tissue. Each step is recorded rather than compressed into a rule that appears to have arrived fully formed.

Why nominal and not measured, which 0.3.9 got wrong. 0.3.9 required identical pitch with no threshold of triviality, borrowing the construction §4 applies to parameter departure. That construction is sound because a parameter is an exact declared number. Pitch as measured from coordinates is not: it carries estimation error, anatomical curvature and projection choices. Applied to the register's own supply the rule made nearly every multi-subject reading unissuable — the seven subjects pooled into NS-0012 measure 5.86 to 7.85 mm and would have become seven readings (gate/RESULT-PITCH.md). A rule nothing can satisfy protects nothing. The nominal figure is exact, comes from a specification, and is what two readings must share.

Non-uniform layouts, added in 0.4.3. The first array this section was applied to outside a clinical grid has no pitch at all. The NeuroNexus EEG-mouse-30-A array behind NS-0006 and NS-0011 is anatomically placed — fifteen sites per hemisphere over named cortical targets, mirrored about the midline, with nearest-neighbour distances running 1.00 to 1.38 mm. The manufacturer publishes per-site coordinates relative to bregma and states no spacing, and neither does either methods paper. A rule that requires a nominal pitch cannot be satisfied by such an array, and 0.3.9 through 0.4.2 silently assumed every array is a lattice.

Where layout is non_uniform: the declared pitch is the median nearest-neighbour distance over the site coordinates as the manufacturer publishes them, the full coordinate table is cited in grain.basis_source, and the consistency check below compares the deposit's own coordinates against that manufacturer-derived figure rather than against a catalog spacing. A derived figure of this kind is never presented as a manufactured one.

Consistency check, with a measured band. Where both figures exist, the measured pitch must sit within 15% of the nominal. This factor is measured, not provisional: across the four deposits in the register's human supply that declare their coordinate units, clinical grid and strip groups measure 9.00 to 10.05 mm against a nominal 10 mm — inside 10% in every case, over 14 groups. A measured value outside the band means the coordinates are not in the space the deposit implies, and the reading is not issuable until the discrepancy is named. One deposit in the register's own supply fails it: ds004370 declares no coordinate units and its grid contacts measure 21 to 41% below nominal.

What is not yet known, and is stated rather than papered over. The tolerance rests on mouse surface EEG. The one human grid measured shows no detectable effect across a 4× range, which is not agreement with the mouse and is a single subject. Until a human cohort is run, the 1.3 tolerance is carried on the mouse measurement alone and is listed for review each cycle. Separately, recorded extent moves the mouse reading further than pitch does — 0.761 at 0.67× extent against 1.014 at 1.49× pitch — which is the subject of §5's coverage widener rather than of this section, and is now measured rather than asserted.

Readings issued before 0.3.9 are not edited (§8). Their grain is recoverable from their source deposits but is not a field of the record, and the conformance report shows the gap rather than hiding it.

3.6 Temporal grain

Added in 0.4.9, in response to gate/RESULT-TEMPRES.md.

§3.5 fixes the spatial resolution of a recording. This section fixes its temporal resolution: how many samples the response window actually contains. The omission was invisible for six versions because every reading the register computed happened to be electrophysiology at several hundred hertz, and the question only became askable when a substrate arrived that is imaged at 2 Hz.

The reason is measured, on the register's own recordings. The sixteen mouse state-cells behind NS-0006, NS-0011 and NS-0027 were re-read at their issued parameter set with the resample rate swept down, nothing else changed — same brain, same stimulus, same trials, same windows, same channels. Only the number of samples in the response window moves:

The decay is monotone and tight — at 200 samples the interquartile range is [0.60, 0.79] — so this is not noise. It is the measure. A principal-component decomposition of an evoked response cannot return more components than the response has samples, and PCIst's value is built from them.

Two consequences, and the second is the one that matters.

First, a reading taken at a materially different temporal grain is not comparable to one taken at this register's grain, and no substrate is exempt from that. A calcium recording at 2 Hz sits at roughly a third of the value the same underlying response would return at electrophysiological rates. Publishing such a number next to a mouse value of 61.35 would be the §3 scale error committed in a new axis.

Second, and this is a defect of record rather than a rule for the future: the register's own issued readings already span this axis undeclared. Its mouse readings realize 399 samples and its human readings realize 207 — a 1.93× spread, sitting in the steepest part of the curve above. No claim in the register rests on a direct mouse-to-human absolute comparison, and §3.4's modality tiers already forbid several of the comparisons that would be affected; but the spread was never declared, never checked, and was found only by trying to issue a reading on a substrate that made the axis visible. It is recorded here rather than corrected, because §8 does not edit issued readings.

What survives degradation, and what does not. The direction of a within-subject contrast is robust where the absolute value is not: awake over isoflurane held in 6 of 6 subjects at every sample count tested, down to 10 samples, even where the absolute value had fallen to 0.15×. Trial count behaves oppositely. Swept independently, at the 2 to 5 trials a single-trial-per-target paradigm affords, the median lands between 0.52 and 0.68 of the full value with an interquartile range as wide as [0.12, 1.80], and the awake-over-anesthesia direction fails in 2 of 6 subjects. Sparse sampling makes the number smaller; few trials make it unreliable. A substrate carrying both carries both faults, and they must be reported separately rather than summed into one caveat.

Every reading issued under 0.4.9 or later declares its temporal grain:

Resampling upward does not create samples. The reference computation will resample to any rate asked of it, so a recording acquired at 312.5 Hz can be resampled to 725 Hz and declare 207 samples in a 285 ms window, sitting beside a reading genuinely acquired at that rate. Interpolation adds no information. response_samples is therefore computed from min(resample_hz, fs_hz), and a declared count above what acquisition supports is not conformant.

The sensitivity is a property of the recording. It cannot be looked up. Added in 0.4.10. When 0.4.9 declined to set a tolerance on one substrate, that caution turned out to be the whole finding rather than a placeholder. Four sweeps now exist. Read at matched fractions of each recording's own native grain:

Three things follow, in order of how much they constrain the standard.

First, the two short-window human recordings agree to 1.004× at 30% of native — four parts in a thousand — across a difference in modality (scalp against subdural), perturbation (TMS against direct electrical stimulation), sampling rate (312.5 against 725 Hz), channel count and geometry. Where the window is matched, the degradation is reproducible across recordings that have almost nothing else in common.

Second, window length accounts for about half the remaining gap, and no more. The mouse re-swept at a 300 ms window moves from 0.582 to 0.705 at 30% of native against the humans' 0.820 — roughly half the distance. Shortening the window is not sufficient to make two recordings comparable, and the residual is a property of the recording that no declared parameter predicts.

Third, and consequently, at matched absolute sample count the substrates disagree by 2.2× to 3.3×. A tolerance set from any one of them would have been wrong for the others by more than the state contrasts this register publishes.

So this section requires what §2.3 requires of the null, and for the same reason: the figure is measured on the recording rather than inferred from a population. A reading that will be compared as an absolute value against a reading at a different realized sample count carries its own measured sensitivity — the same recording re-read at the other reading's grain — and the comparison is made at the matched grain or not at all. Where no such comparison is made, declaration alone suffices.

What is still not known. Every sweep so far is cortex, and three of the four are the evoked response to an instrumented perturbation in a mammal. Whether the residual tracks anything nameable — response duration, dominant frequency, channel count — is unmeasured, and until it is, "measure it per recording" is the honest instruction rather than a placeholder for a formula.

---

---

4. The reading record

A reading is not a number. It is a record with the number in it. Every issued reading is a file in readings/ conforming to schema/reading.schema.json, containing at minimum:

commissioned_by and fee are mandatory fields with no default. A reading that omits them is invalid and the validator rejects it. This is not a disclosure convention that can be waived by agreement with a commissioning party; it is a structural property of the record format.

---

5. Uncertainty

Every reading carries an interval. The interval is not decoration and it is not a standard error dressed up — it is the honest width of what is not known, and for several reading types it is very wide.

The interval must widen for each of the following, and the reading must name which ones applied:

• Small n. Fewer than five subjects, or fewer than the number of trials the reference implementation's parameters assume.
• Instrumentation mismatch. Electrode count, sampling rate, or montage materially different from the benchmark setup.
• Species extrapolation. Any non-human subject. The cortical geometry that the method's spatial assumptions rest on is not shared.
• Preparation extrapolation. In vitro tissue, where "the rest of the system" that a response would propagate into does not exist.
• Proxy substitution. Type C, where no perturbation was delivered at all.
• Parameter departure. Any deviation from a documented parameter set.

A reading whose interval spans the empirical cutoff must say, in its own summary line, that it does not resolve the question it was taken to answer. Being unable to tell is a result, and it is published exactly like any other.

---

6. Artificial systems: indicator profiles

Artificial systems are assessed by a different instrument and reported on a different page.

The assessment is against the fourteen indicator properties set out in Butlin, Long et al. (2023; arXiv:2308.08708), derived from recurrent processing theory, global workspace theory, higher-order theories, attention schema theory, predictive processing, and agency and embodiment. Each property is scored as present, absent, or not assessable, with the evidence stated.

6.1 Access levels

Assessability depends entirely on access, and the level must be declared:

An L0 profile is not a weaker version of an L2 profile. It is a statement about eight properties and silence about six, and it must be read as such.

6.2 Contamination controls

A system trained on text about consciousness will produce text about consciousness. Every profile must apply and document:

• Vocabulary-free framing. Probes that do not use the terms the literature uses.
• Paraphrase invariance. Consistent scoring across semantically equivalent probes.
• Negative controls. Probes designed so that a system without the property would nonetheless produce a convincing-sounding answer.
• A rotating held-out item set. Never published, rotated on a fixed schedule, so that the public probe set is not the scoring probe set.

The public subset of probes, the scoring code, and the per-item results publish with every profile. The held-out set does not.

6.3 What a profile is not

An indicator profile does not produce a number, does not go on the perturbational axis, and does not support a statement that a system is or is not conscious. It reports which properties the evidence supports, which it does not, and which could not be examined at the access level granted.

---

7. Claims discipline

A reading, and any commentary issued with it, may not contain:

1. That a subject is, or is not, conscious.
2. That a subject does, or does not, have experiences, feelings, suffering, or welfare interests.
3. That a subject is more or less conscious than another subject, unless both are Type A or B, measured on matched instrumentation, with overlapping parameters.
4. A value without its interval.
5. An interval without the basis on which it was derived.
6. A comparison across reading types without the §3 caveat stated in the same visual field.
7. A recommendation about care, treatment, protection, regulation, or legal status.
8. A statement about what the number implies for anyone's decision.

A reading may contain: the value, the interval, the conditions, the provenance, the caveats, and a plain-language sentence describing what was measured and under what limits.

The distinction is the whole business. A reading that tells someone what to do is an opinion with a number attached, and the moment this standard issues one, every reading it has ever issued becomes an opinion too.

---

8. Versioning

This document is versioned MAJOR.MINOR.PATCH.

• PATCH — clarification with no effect on any computed value.
• MINOR — a change that could alter a value: new parameter set, new reading type, changed uncertainty rules. Existing readings retain their method_version and are not recomputed silently.
• MAJOR — a change to the reference computation or to §3 or §7.

Readings are never edited. A reading found to be wrong is superseded by a new reading that names it in supersedes, and the original stays in the register with a pointer forward. The register is append-only. There is no mechanism, and will be no mechanism, for removing a reading from it.

---

9. Validation

9.1 The reproduction gate

Before any reading is issued, the pipeline must reproduce a published result from the original authors' data using the original authors' code.

Target: the mouse PCIst results of Claar, Rembado et al. (2023), eLife 84630, from the Allen Institute's openly licensed deposit (DANDI:000458), using the reference PCIst implementation.

The pass criterion

Version 0.2.0 required the pipeline to "reproduce a published result," which is not a criterion. The second attempt reproduced four published quantities, failed one, reproduced an effect in direction while overshooting its magnitude, and separated sixteen of eighteen sessions. Under the old wording that is a pass, a failure, or a partial depending on which sentence is read as the rule, which means the author decides after seeing the numbers. That is not a gate.

The gate passes when all four of the following hold on a single run over the whole deposit:

Combined standard error is the root sum of squares of the two standard errors of the mean. Two means within two combined standard errors are not distinguishable at conventional thresholds; this is the ordinary test for agreement between two reported means, not a bar chosen for this project.

Requirement 4 is the strict one, and it is deliberate. A method that stratifies its measurements has claimed the strata are meaningful. If a stratum does not reproduce, the method does not reproduce, regardless of how the pooled figure looks.

Recorded but not gating: the quiet-versus-active wakefulness ordering, and the magnitude of the depth ratio. Both are reported in every attempt. Neither blocks a pass, because the published error bars on them are wide enough that agreement would carry little information.

This criterion was written with attempt 2's results already known

That is a defect in the criterion's provenance and it is recorded rather than hidden. The rule above was drafted after the numbers were in, so it cannot serve as a blind pre-registration of them.

Two consequences follow, and both are binding.

Attempt 2 does not pass. Applying the rule above to it: requirement 1 holds at 16/18, requirement 2 holds at 1.24, requirement 3 holds at 0.84 — and requirement 4 fails, because the superficial-stimulation stratum sits 2.15 combined standard errors from its published value. The verdict on gate/RESULT-2.md stands as not passed.

The criterion binds from attempt 3 forward and may not be relaxed to admit an attempt that has already run. Changing any threshold above is a MINOR version change to this document under §8, requires the amendment process in GOVERNANCE.md, and must state which past attempts the change would have reversed.

Outcome if it fails: the failure is published, in this repository, with the attempt, before anything else is. There is no version of this project that quietly moves on from a failed reproduction.

This gate exists because it is the cheapest possible falsifier. If the method cannot reproduce a published number from published data with published code, the problem is found in month four rather than year three.

9.2 Standing validation

• Every reading in the register recomputes from its recorded provenance on demand, in CI, on the current pipeline.
• A recomputation that no longer matches its recorded value opens a public issue automatically. It is not resolved by adjusting the value.

---

9.3 Interrater reliability

Added in 0.3.6. §9.1 and §9.2 test this pipeline against the published world and against its own past outputs. Neither detects the failure that actually kills a measurement standard: two competent people running this method on the same deposit and getting different readings. A standard that does not produce the same answer twice is a house style with a version number.

INSTRUCTION.md is the remedy — every judgment reduced to a signalling question with a pre-specified threshold. This section is how the register finds out whether the remedy worked, and publishes the answer whether or not it flatters.

Rating. Every reading is produced independently by two raters who do not confer. Both ratings are recorded in full, at domain level, before adjudication. Disagreements are resolved by a named third adjudicator whose decision and reasoning are recorded.

The pre-adjudication ratings are retained permanently. Most bodies adjudicate and discard the disagreement. Retaining it is the only thing that makes any statistic below verifiable rather than asserted, and discarding it would make every figure in this section unfalsifiable.

9.3.1 Rating modes, and the honest disclosure of the one in use

§12 condition 9 requires two independent raters. At the time this section was written there was one, and a standard whose conditions cannot be met by the body that wrote them is in one of two states: quietly breaking its own rule, or unable to issue. Neither is acceptable, and the resolution is neither to weaken the condition nor to pretend it is met.

Every reading states the mode under which it was rated.

dual_independent_implementation tests a weaker proposition than dual_independent_human and is labeled as the weaker thing wherever it appears. Two implementations agreeing shows the method is specified tightly enough to execute the same way twice. It does not show that two minds reading the same instruction reach the same judgment, because a shared author is a shared assumption, and the questions an implementation never thinks to ask are exactly the ones where the instruction is loosest. It is a floor, not a substitute, and the register does not report it as interrater agreement.

Run for the first time, 2026-09-21 (gate/RESULT-DUAL.md), in the dual_implementation_same_author mode. A second implementation was written from §2.1 and the published description, resolving five points the specification leaves open. Across seventeen cells given byte-identical input:

These are published at that value because publishing them only when they flatter is the failure mode this section exists to prevent. They say something precise: the specification is tight enough to fix which of two states reads higher and nowhere near tight enough to fix by how much. The second implementation removes the practical obstacle — the machinery now exists and can be handed to anyone — but it does not make a reading conformant, because the two share an author. Conformance under condition 9 needs an implementer who is not us, and the register's conformant count stays at zero until there is one.

A single reading is a valid record. It is not conformant, it carries that on its face wherever its value is displayed, and it is excluded from every published reliability statistic — because a statistic computed over readings that had only one rater would be a statement about nothing.

The register publishes the count of conformant against total readings each cycle. That number is expected to be low at first and to rise, and publishing it while it is low is the only way the rise means anything.

Published statistics. Each cycle, and before the first figures are known:

Both κ and Gwet's AC1 are published because κ is unstable when marginal distributions are skewed, and these marginals will be skewed: most readings will land in the same two or three type and tier combinations, which deflates κ artificially. Publishing only κ would understate reliability; publishing only AC1 would look like selection of the flattering statistic. Both, always.

The scalar statistics are an addition rather than an inheritance. The source standard this procedure is adapted from grades claims into ordered bands, where κ is the right tool. A Nooscope reading's principal output is a continuous value, and κ does not apply to it at all. Two raters who both assign Type B, ecog_grid, M1 and agree on every categorical judgment can still return 32.26 and 46.70, and a reliability program that reported only κ would score that as perfect agreement. ICC and the limits of agreement are what catch it.

Improvement mechanism. Where a domain's agreement falls below κ = 0.40, the response is an amendment to the signalling questions and thresholds in INSTRUCTION.md, not an instruction to raters to try harder. Low agreement is evidence that a judgment is under-specified, and the remedy for under-specification is specification. Each amendment is versioned and the affected readings are re-rated under the new version, with both results retained.

Reproducibility challenge. On a published schedule, a set of twenty fully worked readings is released with complete rating worksheets, source provenance and rationales, with an open invitation for external teams to produce the same readings independently and publish their agreement with the register. The one published cross-organization comparison in the adjacent field found near-zero agreement, so surviving such a challenge at moderate agreement is a stronger credibility claim than any internally generated statistic.

For context, published benchmarks for unaided expert judgment on comparable rubrics run from κ = 0.06 to κ = 0.44. A body publishing κ = 0.45 with a stated improvement plan is in a stronger position than one publishing nothing, and no measurement or grading body currently publishes this at all.

9.4 Constructed evidence and its limits

A measure is developed on systems built to have a known answer, because no real system has one. That is legitimate and it is how every result in gate/ was obtained. What it cannot do is establish which real substrates a measure applies to.

The following therefore binds:

1. A statement that a measure applies to, or fails on, a named class of real system — cortex, a cultured preparation, an artificial network — may not be issued on constructed evidence alone. It requires a measurement on a system of that class, or it is recorded as an open question rather than as a domain.
2. A candidate measure is tested on tissue before it is written up. The rule was adopted after three runs of constructed-system evidence failed at first contact with mouse recordings, because a property of the test bed — being scale-free by construction — had been carried silently into a definition (gate/RESULT-SPREAD-REAL.md).
3. A construction is not a ground truth until the construction itself has been checked. An unnormalized quantity minimized by the most unbalanced cut, and random within-block weights that move a designed boundary, both produced confident wrong numbers that were invisible in the output (gate/RESULT-LINEARGAUSS.md).
4. Every descriptor entering a comparison must be shown to vary across the set it is compared against. A quantity pinned by the design of the comparison carries no information, and a normalization applied to it anyway lets numerical noise decide the result while every reported figure remains plausible (gate/RESULT-CELEGANS.md).

The rule exists because this register has broken it. A domain-of-validity sentence asserting that cortex is distance-structured was written from seven synthetic topology families, carried into the plan, and used to argue that artificial systems are the riskier extrapolation. The first measured causal graph it met — a complete nervous system — sat outside the ensemble entirely, and the sentence was withdrawn.

---

10. Independence

These rules bind the issuer and are structural, not aspirational. They exist because the entire value of a reading is that it is not for sale.

1. Disclosure is a schema field, not a policy. commissioned_by and fee are required. A reading cannot be issued without them.
2. Fees are fixed and payable regardless of the finding. No engagement is accepted on contingency or on any basis tied to an outcome.
3. Every commissioned reading enters the public register, whichever way it comes out. A commissioning party unwilling to accept this is declined before work begins.
4. No veto, no revision right. An assessed party receives its report thirty days before publication and may publish a response alongside it. It may not alter the finding.
5. Corrections are published by the issuer first. Version history stays visible.
6. The standard, the register and the certification mark are held separately from any entity performing commercial assessment work.

---

11. References

• Casali AG, Gosseries O, Rosanova M, et al. (2013). A theoretically based index of consciousness independent of sensory processing and behavior. Science Translational Medicine 5(198):198ra105.
• Casarotto S, Comanducci A, Rosanova M, et al. (2016). Stratification of unresponsive patients by an independently validated index of brain complexity. Annals of Neurology 80(5):718–729.
• Comolatti R, Pigorini A, Casarotto S, et al. (2019). A fast and general method to empirically estimate the complexity of brain responses to transcranial and intracranial stimulations. PMID 31133480.
• Butlin P, Long R, Elmoznino E, et al. (2023). Consciousness in artificial intelligence: insights from the science of consciousness. arXiv:2308.08708.
• Breyton M, Fousek J, Rabuffo G, et al. (2025). Spatiotemporal brain complexity quantifies consciousness outside of perturbation paradigms. eLife 13:RP98920. — the basis for §3.1.
• Claar L, Rembado I, et al. (2023). Cortico-thalamo-cortical interactions modulate electrically evoked EEG responses in mice. eLife 84630. — source publication for the reproduction gate dataset.
• Mikulan E, Russo S, et al. (2020). Simultaneous human intracerebral stimulation and HD-EEG, ground-truth for source localization methods. Scientific Data 7:127.
• Reference implementation: renzocom/PCIst (GPL-3.0). See LICENSING.md for what that license requires of anything built on it.

---

This document is the standard. The code implements it; where they disagree, this document governs and the code is the bug.

---

12. Conformance conditions

A third party may state that a reading was produced under this standard only where all eleven of the following hold. Until this section existed, the Nooscope was a method. These conditions are what make it a standard others can be held to, and they are what a certification mark would later recite.

1. The reading is a record conforming to schema/reading.schema.json, with every mandatory field present — including commissioned_by and fee, which have no default and cannot be waived by agreement with a commissioning party.

2. The reading type is assigned per §3 and displayed wherever the value is displayed.

3. The measurement modality tier is assigned per §3.4, displayed on the face of the reading, and its ceiling applied.

4. The scale is declared, and no comparison in the reading or its commentary spans two scales.

5. Recording geometry is assigned from the channel-type census per §3.3, not by judgment, and a mixed-implant subject is excluded rather than reduced to its larger array.

6. The parameter set is named and every departure from it is declared with a reason, with no threshold of triviality.

7. The interval is derived per §5 and INSTRUCTION.md Domain 5, from the between-unit standard deviation rather than the standard error, with every applied widener named and every provisional factor marked as provisional.

8. The rejection record states the rule and the counts, and no channel or trial was excluded by a decision not expressible as a rule.

9. The reading was produced independently by two raters with disagreements retained and adjudicated by a named third party, per §9.3, in a rating mode §9.3.1 marks as conformant. A single-rated reading is a valid record and is not conformant, and says so wherever its value appears.

10. The versions of this document and of INSTRUCTION.md in force at the time of computation are cited with the reading.

11. Recording grain is declared per §3.5 — nominal pitch, admitted channel count, and the basis it came from — the measured pitch sits within 15% of the nominal where both exist, subjects combined into one reading fall within a 1.3 nominal pitch ratio, and no absolute comparison beyond that ratio appears without the §7 cross-grain caveat in the same visual field.

12. Temporal grain is declared per §3.6 — acquisition rate, resample rate, the realized number of samples in the response window, and the trial count with the number of distinct perturbation sites those trials are drawn from. No absolute comparison against a reading whose realized sample count differs appears without the §7 cross-grain caveat in the same visual field. No tolerance is set: §3.6 requires declaration and explicitly declines to state a comparison ratio it has not measured.

Conditions 11 and 12 were added in 0.3.9 and 0.4.9 and appended rather than inserted, so that every reference to conditions 6 through 10 in earlier documents still points at what it pointed at. Readings issued before those versions fail them, and the register reports that rather than backdating them.

A value displayed without its interval, its type, its tier or its version is not conformant, and is a breach of license where a license exists.

Conformance is a property of the record, not of the result. A reading that conformantly reports a wide interval, a failed widener, or an inability to resolve the question it was taken to answer is conformant. A reading that reports a confident number without its tier is not.