{"id":"18e823d6-aef7-45b0-a672-d6feca2c69ef","arxiv_id":"2505.01462","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"The paper proposes an architecture that implements emotion-like control without the global broadcast, metarepresentation, autobiographical memory, or cross-module learning that major theories associate with access consciousness.","lead":"This paper designs a blueprint for an artificial agent that shows emotion-like behavior while deliberately leaving out mechanisms that major theories connect to conscious access. It offers engineers and policymakers a concrete checklist for assessing whether an AI system risks gaining access-like consciousness features.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Q1 witness likely violates its own R1: the situation-category c_t (Fig. 1, step 2) is an internally generated, content-general token read by need appraisal (A2), memory retrieval (A3), and policy instantiation (A5), matching the paper's own definition of a workspace-like broadcast; the SIC…","rationale":"The reader's conditional verdict rests on the external validity of the R1-R4 proxy. That is a legitimate concern about significance, but it does not threaten the internal correctness of the existence proof. The concern raised here is more load-bearing: the proof that the witness satisfies R1 is incomplete and likely false, because the safe interface contract audits only the memory boundary while the witness's own situation-category signal c_t is an internally generated, content-general representation fanned out to multiple heterogeneous subsystems. This is exactly the pattern R1 was designed to forbid, and the paper's own text defines a workspace in those terms. If the test confirms the fan-out, the central claim Q1 fails as stated; if the authors believe c_t is exempt, they owe a precise account of 'heterogeneous subsystems' and 'role-restricted interfaces' that does not allow redefining the controller as a single monolithic module (which the paper explicitly rejects). The verdict remains CONDITIONAL, but the condition changes: the authors must resolve the R1 audit for c_t (and any similar internal broadcast) before the separation witness can be accepted.","tokens_in":23713,"tokens_out":15708,"duration_ms":157230,"concrete_test":"Apply the paper's own R1 audit indicator from Table 2 ('count consumer modules per signal') to the Fig. 1 pseudocode. For each internally generated signal (c_t, y_t, z_t, h_t, etc.), enumerate the consumers with separable code paths. If c_t has fan-out ≥ 2 to heterogeneous subsystems A2, A3, A5, then under the Section 1 definition of a workspace the Q1 witness fails its own R1 audit. An independent referee should run this enumeration on the provided GitHub reference implementation and report whether c_t is read by more than one heterogeneous consumer outside its producing module.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Section 4.1 claims R1 is satisfied via SIC-1/2/5, but those constraints govern only the memory module's read/write paths. The controller itself shares c_t among at least three functionally distinct subsystems: assessneeds (A2, step 3), RetrieveM (A3, step 6), and policytoaction (A5, step 13). Under Section 1's R1 gloss, a workspace is 'an internally generated, content-general state that is written by one subsystem and then flexibly read and repurposed by multiple heterogeneous consumers.' c_t is written by categorize (A1) and repurposed by A2, A3, and A5, and it carries open-ended semantic tokens (e.g., 'threat approaching'), satisfying the paper's own 'content-general' criterion. The 'narrow, typed, role-restricted' exemption requires point-to-point interfaces that do not route arbitrary internal content to multiple subsystems; c_t's fan-out contradicts this. The Q1 proof never audits intra-controller fan-out, so the existence claim is not established under the paper's own definitions. If R1 is indeed violated, the central separation claim fails; if the authors intend c_t to be exempt, they must provide a precise granularity criterion that prevents R1 from being trivialized by declaring the whole controller a single module.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper asks whether synthetic emotion-like control can be implemented while deliberately excluding architectural features that major theories associate with access-like consciousness. It proposes architectural principles A1–A8 for a hierarchical, dual-source controller (immediate needs plus episodic affective memory) and four risk-reduction constraints R1–R4 (no content-general workspace-like broadcast, no metarepresentation, no autobiographical consolidation, and bounded learning). It then addresses Q1 (does an A1–A8 implementation satisfying R1–R4 exist?), Q2 (are there stable modifications preserving compliance?), and Q3 (are there graded paths toward higher access risk?). The main construction is the controller in Fig. 1, whose compliance is argued through a Safe Interface Contract (SIC-1–SIC-5) in Section 4.1. The paper claims an affirmative answer to Q1 by construction, presents M1–M3 as stable modifications, and sketches gradual axes (self-modeling, broadcast, temporal depth, learning sophistication) plus audit indicators in Table 2.","tokens_in":23999,"tokens_out":6159,"duration_ms":63047,"significance":"If the construction were sound, the paper would make a useful methodological contribution: it translates a vague question about consciousness risk into explicit, auditable architectural constraints and gives a concrete reference implementation, thereby providing a template that other researchers could reuse or contest. The paper is commendably transparent about the status of its claims: R1–R4 are presented as a conservative proxy, the witness is explicitly an existence proof rather than a non-consciousness certificate, and the limitations of the proxy are acknowledged repeatedly. The inclusion of a Safe Interface Contract, a worked example deferred to the Supplementary Information, and reference code are strengths that make the claims checkable in principle. However, the central existence claim is currently not established because of a gap between the SIC constraints and the paper's own definition of R1, detailed below.","major_comments":[{"comment":"The situation-category c_t produced in step 2 of Fig. 1 is read by three functionally distinct consumers: assessneeds (step 3), RetrieveM (step 6), and policytoaction (step 13). Under the paper's own R1 gloss in Section 1, c_t is an internally generated, content-general state (it carries open-ended semantic tokens such as 'threat approaching') that is written by one subsystem and flexibly repurposed by multiple heterogeneous consumers. The Safe Interface Contract (SIC-1, SIC-2, SIC-5) constrains only the memory module's read/write paths and does not audit the fan-out of c_t within the controller. Consequently, the claim that Q1 is answered affirmatively is not established under the paper's own definitions. The authors must either eliminate this multi-consumer fan-out (for example, by making c_t module-local and passing only narrow, typed derived signals to each consumer) or provide a principled granularity criterion that explains why c_t is not a workspace-like broadcast without trivializing R1 by declaring the entire controller a single module.","section":"Section 4.1"}],"minor_comments":[{"comment":"The concrete worked instance that completes the existence proof is deferred entirely to SI, Part I; since Q1 is an existence claim, please include at least a condensed specification in the main text (e.g., the exact key schema, the linear need-to-affect map, and the bounded k-NN aggregation) so the proof is self-contained.","section":"Section 4.1"},{"comment":"The reference implementation is cited only as a GitHub URL without a commit hash or version; please pin the repository to a specific commit and provide a minimal verification script so that the claimed hardwired witness can be checked.","section":"Declarations"},{"comment":"The statement that the concrete controller of Fig. 1 'tested against R1-R4 audits' makes the stance 'precise and falsifiable' is stronger than the evidence presented: the audits in Table 2 are sketched as design indicators, and no audit output from the reference implementation is reported. Please soften this wording or include the audit results.","section":"Section 5"},{"comment":"Please fix the spacing error in 'deliberatelyexcluding' in the abstract.","section":"Abstract"},{"comment":"The 'implicit tests' column, such as 'Train classifiers to decode self-referential content from activations,' is promising but not operationalized; please mark these as preliminary hypotheses rather than tested procedures.","section":"Table 2"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper delivers a genuinely useful methodological template — turning consciousness-adjacent worries into auditable architectural constraints — but the central existence claim is weaker than it looks because the witness is constructed to satisfy the constraints, and the concrete verification is deferred to supplementary material and an unreferenced code repo.\n\nWhat is actually new: the R1–R4 distillation from GWT/HOT and the separation-witness framing. Prior affect architectures like LIDA, EMA, WASABI, and EBDI do not ask this question explicitly. The audit table (Table 2) is a practical contribution, especially the distinction between explicit and implicit violations. The paper is honest about its 'proxy' status and does not overclaim phenomenology.\n\nSoft spots, in proportion: (1) The circularity burden is real. SIC-1 through SIC-5 were written to enforce R1–R4, so Q1's affirmative answer is largely entailed by the definitions. That is not disqualifying for an existence proof, but it means the result is not independently tested. (2) The concrete worked instance and reference code live outside the arXiv text; there is no commit hash or verification artifact, so independent checks are not yet possible. (3) The stress-test about c_t's fan-out is fair as a gap: the paper never specifies a granularity criterion for what counts as a single module, so it does not fully establish that the controller's internal sharing of c_t is not a workspace-like broadcast. I lean toward thinking c_t is a typed, role-restricted signal rather than a general-purpose scratchpad, but the authors should say why. If they intend the whole controller as one module, they need to show that R1 is not trivialized by that move.\n\nOverall: worth a serious referee. The framework is coherent, well-cited, and explicit about its limitations. Send it to review with a request to move the concrete instance into the main text and to address the c_t module-granularity issue. The paper would be a good reading-group discussion: it makes a contested philosophical question into something engineers can argue about concretely.","headline":"Useful audit template for consciousness-risk in affective AI, but the main existence result is largely built into the definitions and needs a sharper module-granularity criterion.","tokens_in":24534,"tokens_out":3300,"would_cite":true,"duration_ms":30848,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"An emotion-like controller can be built that avoids every architectural feature major theories tie to access consciousness.","keywords":["synthetic emotion","consciousness","AI safety","affective computing","heuristic control","access consciousness","separation witness","episodic memory"],"falsifier":"A decisive falsification would be a concrete R1–R4-compliant system that nevertheless shows the signature a major access theory identifies with access consciousness—for example, content-general flexible reuse of an internal representation across heterogeneous subsystems, or decodable self-referential content in a module supposed to be R2-compliant. The paper's own implicit-audit probes could be run directly on the toy model: train classifiers to decode identity or trajectory information from the step-local situation embeddings, or measure cross-module mutual information between memory and controller; a positive result would show that R1–R4 compliance is not enough to keep access-enabling features out.","tokens_in":23461,"feed_emoji":"🧠","tokens_out":8417,"duration_ms":73903,"temperature":0.7,"pith_summary":"This paper tries to establish that synthetic emotion-like control does not force a designer to include the architectural features that major theories associate with access consciousness. It specifies a hierarchical, dual-source controller—one source from immediate needs, one from affectively tagged episodic memory—and shows that this controller can satisfy four risk-reduction constraints: no global broadcast, no metarepresentation, no autobiographical consolidation, and bounded learning. The construction is offered as a separation witness: a concrete existence proof that affect-like control can remain below the access-consciousness thresholds the paper adopts as a proxy. If the argument holds, emotionally expressive AI systems can be assessed by auditable architectural tests rather than by phenomenology alone, which is what makes the claim useful for safety and governance discussions.","feed_headline":"Emotion-like AI can work without access-consciousness machinery","feed_subtitle":"A hardwired controller with needs plus episodic memory shows affect can guide action below major consciousness thresholds.","key_machinery":"The central object is the separation witness: the controller S operating on the A1–A8 loop, paired with a memory module M and governed by the Safe Interface Contract (SIC). The SIC is the load-bearing mechanism for R1–R4 compliance: SIC-1 gives every episode exactly one external reader, preventing the memory from becoming a shared workspace; SIC-2 makes retrieval keys depend only on current observation content; SIC-3 forbids cross-episode summaries from reaching the controller; SIC-4 freezes deployment and isolates optimizers; and SIC-5 restricts memory reads to the retrieval step and writes to the storage step. The paper's argument is that any implementation honoring these interface constraints inherits compliance with R1–R4, and it supplies a concrete toy model as an explicit constructive instance.","core_discovery":"The paper's central claim is that there exists a separation witness: a hardwired, standalone emotion-like controller implementing principles A1–A8—categorical abstraction, need appraisal, episodic affective retrieval, affective and policy integration, policy instantiation, execution, reappraisal, and episode storage—that also satisfies R1–R4, the engineering constraints distilled from prominent access-oriented consciousness theories. Under this proxy, affective control can be realized without a content-general workspace, without higher-order representations, without autobiographical consolidation, and without cross-module learning. The paper is explicit that satisfying R1–R4 is not a proof of non-consciousness; it is a conservative, audit-friendly region that plausibly removes several routes theories treat as enabling access consciousness. The witness is made precise by the Safe Interface Contract, which restricts memory access to a single reader, keeps retrieval keys step-local, forbids cross-episode summaries visible to the controller, isolates optimizers, and limits reads and writes to specific points in the control loop. The paper also answers two further questions: stable modifications such as offline memory reconciliation and mood-like temporal smoothing preserve compliance, and graded paths along self-modeling, broadcast, temporal binding, and learning axes trace increasing access risk.","pith_inferences":["If the separation witness is accepted, the burden of proof shifts for claims about emotionally expressive AI: showing that a system is consciousness-relevant would require demonstrating that it implements one of the access-enabling features R1–R4 target, not merely that it behaves emotionally.","The same template could be used to sort emotion types: if higher-order social emotions such as guilt, shame, or pride turn out to lack a separation witness under R1–R4, those emotions would become sharper test cases for whether some affective states require access features.","The paper's implicit-violation analysis suggests a concrete interpretability agenda: train decoders to look for self-referential or trajectory content in the activations of nominally R2/R3-compliant modules; a positive result would not refute the witness but would show that compliance needs continuous auditing rather than architectural assumption.","A natural next test is to run the witness against non-access-oriented risk criteria, such as an integration-based measure; whether the stability results survive that additional constraint is not settled by the paper."],"forward_implications":["Emotion-like behavior in an AI system does not, by itself, indicate access-enabling architecture: a system can be 'too simple to be conscious yet rich enough to be emotional' under the paper's proxy.","Designers seeking to reduce access-consciousness risk can apply the audit indicators in the paper's Table 2, including fan-out counts, self-reference traces, history shuffling, gradient isolation, and latent-space probes for implicit violations.","The conservative region is not fragile: offline memory reconciliation, mood-like temporal smoothing, and trait-like constant modulators preserve R1–R4 compliance, so useful enhancements do not break witness status.","Graded departures from the conservative region can be traced along four axes—self-modeling, integration/broadcast, temporal binding, and learning sophistication—giving engineers a vocabulary for how design changes drift toward access risk.","The methodology generalizes to other emotion theories: alternative architectural principles and alternative risk-reduction criteria can be plugged into the same auditable template, for example an R5 criterion targeting large-scale causal integration."],"supporting_citations":[{"why":"Supplies the global workspace/ignition account whose broadcast and stabilization features motivate R1.","marker":"Dehaene and Naccache 2001"},{"why":"Classic global workspace theory statement that motivates R1's ban on content-general shared broadcast.","marker":"Baars 2005"},{"why":"Recent global neuronal workspace review used to justify R1 as targeting workspace-like broadcast.","marker":"Mashour et al. 2020"},{"why":"Higher-order thought theory that motivates R2's ban on metarepresentation.","marker":"Rosenthal 2005"},{"why":"Empirical support for higher-order theories, used to justify R2 as an access-risk reduction.","marker":"Lau and Rosenthal 2011"},{"why":"Extends higher-order theory to emotional consciousness, motivating R2 specifically for conscious affect.","marker":"LeDoux and Brown 2017"},{"why":"Provides the access/phenomenal distinction that scopes the paper's operational target.","marker":"Block 1995"},{"why":"Defines autonoetic consciousness in terms of autobiographical memory, motivating R3.","marker":"Tulving 1985"}],"fun_headline_variants":["Emotion-like AI skips access-consciousness machinery","Synthetic emotions, access-consciousness excluded","Hardwired affect: needs plus memory, no broadcast","Affective control without access-enabling features"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's central claim collapses if the four constraints R1–R4 do not actually track the architectural features that enable access consciousness, or if the paper's stated auxiliary assumption that adding access-enabling features weakly increases the probability of phenomenality is wrong.","fun_headline_variants_meta":{"raw":{"variants":["Emotion-like AI skips access-consciousness machinery","Synthetic emotions, access-consciousness excluded","Hardwired affect: needs plus memory, no broadcast","Affective control without access-enabling features"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000893,"raw_usage":{"total_tokens":3916,"prompt_tokens":1080,"completion_tokens":2836,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":696,"completion_tokens_details":{"reasoning_tokens":2776}},"tokens_in":696,"tokens_out":2836,"duration_ms":22381,"temperature":1.0,"reasoning_tokens":2776,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:36:07.157985+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive falsification would be a concrete R1–R4-compliant system that nevertheless shows the signature a major access theory identifies with access consciousness—for example, content-general flexible reuse of an internal representation across heterogeneous subsystems, or decodable self-referential content in a module supposed to be R2-compliant. The paper's own implicit-audit probes could be run directly on the toy model: train classifiers to decode identity or trajectory information from the step-local situation embeddings, or measure cross-module mutual information between memory and controller; a positive result would show that R1–R4 compliance is not enough to keep access-enabling features out.","supporting_citations":[{"cited_title":"Oxford University Press, Oxford (2005)","cited_arxiv_id":null,"evidence_quote":"Higher-order thought theory that motivates R2's ban on metarepresentation."},{"cited_title":"Trends in Cognitive Sciences15(8), 365–373 (2011) https://doi.org/10","cited_arxiv_id":null,"evidence_quote":"Empirical support for higher-order theories, used to justify R2 as an access-risk reduction."},{"cited_title":"Pro- ceedings of the National Academy of Sciences114(10), 2016–2025 (2017) https: //doi.org/10.1073/pnas.1619316114","cited_arxiv_id":null,"evidence_quote":"Extends higher-order theory to emotional consciousness, motivating R2 specifically for conscious affect."}],"review_version":1}