Crosswalk

CEM/SAI Consciousness Crosswalk

Author: Gergely Vámossy
Date: August 2026
Type: Crosswalk

CEM-001 / SAI-000: A Crosswalk to the Consciousness-Science and AI-Welfare Literature

Purpose: to locate the QSII consciousness and sentience modules (CEM-001, the evidence-aware consciousness model; SAI-000, the sentience-assessment infrastructure) within the existing science of consciousness and the emerging AI–moral-status literature — so that, if these modules are ever pursued as a contribution, they engage the work they must, and so their genuinely-recombinant moves are separated from their re-derivations. This is a positioning document, not an endorsement of CEM's or SAI's claims. It mirrors the REF primitive-basis crosswalk.

Citation caveat. Attributions below are given from established knowledge of canonical works. Verify each against the primary source before using in a submission; do not cite page numbers or exact years from this document without checking.

The one-paragraph situation

CEM-001 and SAI-000 independently converge on the indicator-property, theory-neutral, evidence-weighted approach to machine consciousness and sentience: enumerate the functions that competing scientific theories associate with consciousness, decline to crown any single theory, represent a system as a profile across those functions, and treat phenomenal experience and moral status as separate high-uncertainty inferences that are never granted automatically by any functional result. This is a real, currently-dominant methodology in the science of consciousness — but it is not new, and CEM/SAI cite essentially none of the work that established it. Most pointedly, the flagship 2023 report by Butlin, Long, and colleagues, Consciousness in Artificial Intelligence: Insights from the Science of Consciousness, uses the exact same method — derive "indicator properties" from neuroscientific theories, assess AI systems against them, keep the phenomenal question open — and the 2024 follow-on Taking AI Welfare Seriously (Long, Sebo, et al.) does for moral status what SAI-000 does. CEM/SAI's defensible novelty, if any, is not the framework; it is the governance wrapper (deterministic maturity classification, non-self-approval, adoption-vs-validation separation, the KM ladder) that the consciousness literature does not have. The crosswalk below is what a reviewer in the science of consciousness or AI ethics would expect these modules to engage.

Part A — CEM-001 (consciousness model), concept by concept

The whole method: indicator properties derived from theories, no single theory adopted. CEM's core stance — "consciousness is a family of distinguishable scientific questions that require separate evidence," represent a system across functional dimensions, and treat no theory as established — is precisely the method of Butlin, Long, et al. (2023), Consciousness in AI (the "indicator properties" report from a large multi-author group including Yoshua Bengio, Patrick Butlin, Robert Long, Jonathan Birch, Grace Lindsay, Megan Peters, Liad Mudrik, Anil Seth, and others). That report derives a checklist of indicator properties from the leading neuroscientific theories and assesses AI systems against them, explicitly declining to settle the metaphysics. CEM-001 is that report's method re-derived in-house. Gap to close: this is the single most important omission. CEM's L0–L7 layers and its C(t) variable vector must be positioned against the Butlin et al. indicator list, or it will read as an uncredited reinvention.

Level vs. content vs. state; wakefulness vs. awareness. CEM's separation of conscious level (capacity to support experience at all) from conscious content (what is experienced) is the standard clinical framework from the science of disorders of consciousness — Laureys and colleagues on the two-dimensional wakefulness × awareness model; Owen et al. on covert awareness in unresponsive patients; the locked-in dissociation CEM explicitly cites. Gap: cite the DoC literature (Laureys, Owen, Schiff); CEM's level/content/responsiveness dissociations are its findings.

Access vs. phenomenal consciousness. CEM's A (access) vs. P (phenomenal) split, and its "PX — Phenomenal Status Unknown" tag, is Ned Block's access/phenomenal distinction (1995, Behavioral and Brain Sciences) verbatim in substance. The insistence that report/global-availability may not exhaust conscious content is the live access-vs-phenomenal / "rich vs. sparse" debate (Block; Cohen & Dennett on the flip side). Gap: Block is the owner of this distinction; CEM cannot use A-vs-P without him.

I (integration) and D (differentiation). "Consciousness level is associated with differentiated and integrated large-scale dynamics" is Integrated Information Theory (Giulio Tononi; Tononi, Boly, Massimini, Koch) almost word for word — integration + differentiation is IIT's founding pair. Gap: IIT is the most direct match to CEM's two headline variables and must be engaged, including the well-known critiques (e.g. Aaronson's; the 2023 "IIT as pseudoscience" open letter controversy) so the engagement is even-handed.

Perturbational complexity as an indicator, not a measurement. CEM's "perturbational complexity measures can distinguish many conscious and unconscious conditions, but remain indicators rather than direct measurements" is the Perturbational Complexity Index (PCI) — Casali et al. (2013), Massimini & Tononi's Sizing Up Consciousness. CEM's epistemic caveat about it is exactly the field's own caveat. Gap: name PCI; it is the empirical anchor CEM is gesturing at.

G (global broadcast) and A (access). Global availability / broadcasting for report, reasoning, memory, and action is Global Workspace Theory — Bernard Baars — and its neural form, the Global Neuronal Workspace (Stanislas Dehaene, Jean-Pierre Changeux). CEM's G and A variables are GWT/GNW constructs. Gap: GWT is one of the two or three theories any indicator-property scheme must cite.

R (recurrent stabilization). "Recurrent processing … has empirical support for consciousness-related functions" is Recurrent Processing Theory — Victor Lamme. Gap: cite Lamme; R is his.

S (self-model) and the L5 perspectival layer. The perspectival self-model / representing states as owned by a subject is Attention Schema Theory (Michael Graziano) and, philosophically, Thomas Metzinger's phenomenal self-model (Being No One). Gap: AST and Metzinger own the self-model layer.

M/L6/L7 (metacognition, confidence, recursive awareness). Tracking confidence, error, and uncertainty about one's own states, and "representing that one is aware," is Higher-Order Theory (David Rosenthal; Hakwan Lau & Richard Brown) plus the metacognition/confidence literature (Stephen Fleming; Lau). CEM's L6 "metacognitive awareness" and L7 "recursive awareness" are HOT constructs. Gap: HOT is the third must-cite theory; L6/L7 are its predictions.

P (predictive/counterfactual) and E (embodied/environment-coupled). Predictive modeling and embodied regulation as consciousness-relevant is predictive processing / the free-energy view (Andy Clark, Surfing Uncertainty; Jakob Hohwy) and Anil Seth's interoceptive/"beast machine" account (Being You), with roots in enactivism (Varela, Thompson, The Embodied Mind; Thompson, Mind in Life). Gap: Seth and Clark own P and E.

U (uncertainty/evidence qualification) — where CEM is actually distinctive. Making evidence-qualification a first-class variable inside the model (the U component), and pinning the whole thing to a deterministic knowledge-maturity classification, is not standard in the consciousness theories above — it is CEM's genuine recombinant move, imported from the QSII governance layer. This is the part worth foregrounding as contribution rather than the L0–L7 ladder. Keep: this is the real novelty; see "what it would take."

The adversarial-testing remark. CEM's line that "recent adversarial testing has not established a single complete winner" refers to the Cogitate Consortium adversarial collaboration testing IIT vs. GWT (Melloni, Mudrik, Koch, Dehaene, et al.), results circulated 2023–2025. Gap: cite it directly; it is the evidence for CEM's own agnosticism.

Part B — SAI-000 (sentience assessment), concept by concept

The whole enterprise: sentience as a high-uncertainty moral-status claim under precaution. SAI's thesis — sentience is not read off performance; it requires convergent evidence, explicit alternative explanations, calibrated confidence, and precaution proportionate to possible harm — is the thesis of Jonathan Birch's precautionary framework (The Philosophy of Animal Minds lineage; "Animal sentience and the precautionary principle," 2017; The Edge of Sentience, 2024) transposed to AI, plus Long, Sebo, et al. (2024), Taking AI Welfare Seriously — which, like SAI, combines consciousness and agency markers to argue for a non-trivial credence in near-term AI moral patienthood and for precautionary institutional steps. Gap to close: this is SAI's single most important omission — the 2024 report is almost exactly SAI's remit; SAI must position against it.

Capability / intelligence / agency / autonomy / sentience / sapience / moral status / legal status — the mandatory distinctions. SAI's insistence on separating these is the standard move of Nick Bostrom & Eliezer Yudkowsky ("The Ethics of Artificial Intelligence," on moral status hinging on sentience and sapience), Eric Schwitzgebel & Mara Garza ("A Defense of the Rights of Artificial Intelligences"), and the animal-ethics baseline (Peter Singer; David DeGrazia on moral status). Gap: cite these for the distinctions SAI treats as primitive.

Sentience = valenced experience (pleasure/pain/distress). Defining sentience specifically as valenced experience rather than cognition is the animal-sentience / affective-neuroscience tradition — Jaak Panksepp (affective neuroscience, primary-process emotion); the sentience criterion in animal-welfare science (Marian Dawkins; Donald Broom); and the UK's formal recognition of animal sentience informed by Birch et al.'s cephalopod/decapod review. Gap: SAI's valence-first definition should be anchored here.

Precaution proportionate to possible harm; decisions under uncertainty. SAI's governance core — act precautionarily when potential suffering is material, never authorize irreversible deletion or suffering-inducing tests under uncertainty — is Birch's precautionary principle for sentience, Sebo's work on moral consideration and risk ("The Rebugnant Conclusion"; moral circle under uncertainty), and Metzinger's proposed moratorium on synthetic phenomenology (his call to avoid creating artificial suffering). Gap: these are the owners of "what is permissible under sentience uncertainty."

"Do not infer sentience from verbal claims alone; do not infer non-sentience from silence." This double caution maps onto the gaming/mimicry problem flagged in Butlin et al. and in Schwitzgebel's "the full rights dilemma" and on the animal side to the problem of behavioral read-outs; the "valence not proven by optimization signals" point is the RL-reward-≠-pleasure caution now common in AI-welfare discussions. Gap: cite the mimicry/gaming discussion so SAI's cautions are positioned, not just asserted.

Where SAI is actually distinctive. As with CEM, SAI's recombinant move is not the assessment content but the infrastructure: candidate-system registration, a fixed authority boundary ("may / may not" lists), mandatory escalation to a decision-and-validation layer (SDVI) and to human governance, preservation of superseded assessments and disagreements, and explicit non-self-certification of moral-status recommendations. The AI-welfare literature argues that institutions should do this; SAI specifies a machine-checkable governance object that does it. Keep: this is the contribution — an operational governance schema for sentience assessment, not a new theory of sentience.

Summary table

CEM/SAI concept or move	Nearest existing tradition	Key references (verify)	Status
Indicator properties from theories; no single theory adopted	Theory-based indicator approach to machine consciousness	Butlin, Long, et al. 2023 (Consciousness in AI)	Re-derivation (largest omission)
Level vs. content; wakefulness × awareness	Disorders-of-consciousness framework	Laureys; Owen; Schiff	Re-derivation
Access (A) vs. phenomenal (P); PX-unknown	Access/phenomenal distinction	Block 1995; Cohen & Dennett	Re-derivation
Integration (I) + Differentiation (D)	Integrated Information Theory	Tononi; Tononi & Koch (+ critiques: Aaronson)	Re-derivation; must engage critiques
Perturbational complexity as indicator	PCI	Casali et al. 2013; Massimini & Tononi	Re-derivation
Global broadcast (G), access (A)	Global (Neuronal) Workspace	Baars; Dehaene & Changeux	Re-derivation
Recurrent stabilization (R)	Recurrent Processing Theory	Lamme	Re-derivation
Self-model (S), perspectival L5	Attention Schema; phenomenal self-model	Graziano; Metzinger (Being No One)	Re-derivation
Metacognition/recursive awareness (M, L6, L7)	Higher-Order Theory; metacognition	Rosenthal; Lau & Brown; Fleming	Re-derivation
Predictive (P), embodied (E)	Predictive processing; enactivism; interoceptive self	Clark; Hohwy; Seth (Being You); Varela/Thompson	Re-derivation
"No single winner yet"	Adversarial collaboration IIT vs GWT	Cogitate / Melloni, Mudrik, et al.	Cite as evidence
Evidence-qualification variable (U) + maturity classification	(not standard in the theories)	— (QSII governance layer)	Genuine recombinant move
Sentience as high-uncertainty moral-status claim under precaution	AI-welfare; precautionary sentience	Long, Sebo, et al. 2024 (Taking AI Welfare Seriously); Birch 2017/2024	Re-derivation (largest omission)
Capability/agency/sentience/sapience/moral/legal distinctions	AI & animal moral-status theory	Bostrom & Yudkowsky; Schwitzgebel & Garza; DeGrazia; Singer	Re-derivation
Sentience = valenced experience	Affective neuroscience; welfare science	Panksepp; Dawkins; Broom; Birch et al.	Re-derivation
Precaution; no irreversible tests under uncertainty	Precautionary principle; synthetic-suffering moratorium	Birch; Sebo; Metzinger	Re-derivation
Machine-checkable sentience-assessment governance object	(not standard)	— (QSII/SAI infrastructure)	Genuine recombinant move

Honest assessment and what it would take

What CEM-001 and SAI-000 are, positioned honestly: competent, unusually well-organized in-house re-derivations of the indicator-property approach to machine consciousness (CEM) and the precautionary, evidence-weighted approach to AI sentience and moral status (SAI) — both of which are real, currently-active programs in the science of consciousness and in AI ethics. The reasoning is sound and the distinctions are the right ones. What is missing for either to be a contribution as consciousness science is not more structure; it is scholarship and a thesis. Right now they re-derive, without citation, a literature that already exists and is moving fast.

The trap to avoid is the same one that runs through all of this work: presenting CEM/SAI as a theory of consciousness or sentience invites immediate, fatal comparison to Tononi, Dehaene, Lau, Seth, Block, Birch, and the Butlin/Long reports — and loses, because it re-derives rather than advances them. Presenting the governance-and-assessment apparatus (the U variable, the KM maturity ladder, the non-self-approval and escalation structure, the "may/may not" authority boundary) applied to consciousness and sentience assessment is a different and defensible claim, because that apparatus is exactly what the existing reports call for but do not specify as machine-checkable objects.

To make it publishable, in order:

Position against the two anchor reports. Add explicit comparison to Butlin, Long, et al. (2023) Consciousness in AI (for CEM) and Long, Sebo, et al. (2024) Taking AI Welfare Seriously (for SAI). These are not optional; they are the same enterprise, and a submission that ignores them will be desk-rejected.
Engage the load-bearing theories behind each variable. CEM's C(t) vector is a map onto GWT (G/A), IIT (I/D), HOT (M/L6/L7), RPT (R), AST/Metzinger (S), predictive processing/enactivism (P/E), and Block (A-vs-P). Cite the owner of each, including the major critiques (especially the IIT controversy), so the agnosticism is earned rather than asserted.
State the novel, defensible thesis. Not "here is a model of consciousness" (contested, un-new) but: "a deterministic, non-self-approving governance object for consciousness/sentience assessment, with machine-checkable maturity classification and explicit evidence-qualification, that operationalizes the precautionary recommendations of the current AI-welfare literature." There the novelty is real — the apparatus, not the metaphysics.
Foreground the U variable and the maturity/authority layer as the contribution, and demote the L0–L7 ladder and the C(t) taxonomy to "a consolidation of existing indicators, used as the assessment substrate."
Keep the claim-withholding property visible. CEM states it is "not validated"; SAI "may not self-certify its own moral-status recommendations"; AAIP-007 "does not certify ASI" and "explicitly excludes consciousness, sentience, subjective experience, moral status." This structural refusal-to-self-certify is genuinely notable and is the same safety property as the REF non-self-approval example — it belongs in the methodology paper as evidence, and in any consciousness-module submission as the honest framing.

The recommendation that runs through this whole line of work applies once more: the credible contribution lives in the disciplined method and governance apparatus, and the totalizing framing — a house theory of consciousness, sentience, and superintelligence — should be kept well away from it. Engage the science as scholarship; publish the governance as the novelty.

Related research

Citation

Gergely Vámossy (2026). CEM/SAI Consciousness Crosswalk. https://vamossy.com/research/cem-sai-consciousness-crosswalk