Experimental Measurement of Human-AI Collective Cognition and Identity Continuity

Introduction A human-AI collective may outperform either constituent on some tasks, but a performance increase relative to unaided humans is not sufficient to establish strong complementarity. Published work reports both successful hybrid decision ensembles and negative average strong-synergy results across heterogeneous studies . The purpose of this document is to make potentially beneficial coordination falsifiable. The Watts Synaptic Collective hypothesis represents people as independent cognitive nodes, computational agents as coordination mechanisms, durable evidence systems as record layers, and a separate control plane as the source of operational authority. The biological medulla comparison is an architectural metaphor; this document claims no literal neural structure, psychiatric condition, or collective phenomenal consciousness. This document revises and expands the investigator’s previously prepared draft-watts-synaptic-collective-00 source. It remains an individual Experimental Internet-Draft; no IETF adoption, research validation or global novelty priority is asserted.

Scope and Non-Goals The proposal concerns experimental study of multi-actor workflows, task performance, identity and authority continuity, and controlled semantic reconfiguration. It is not an identity credential profile, OAuth extension, clinical taxonomy, autonomous capability authorization, or settled agent interoperability protocol. The related continuity-event and revocation-closure Internet-Drafts, if submitted, are separate proposals. This document does not require their adoption.

Terminology Human Node: a consenting human contributor who makes an independent judgment before group integration. AI Coordinator: a process used to route, analyze, challenge, or synthesize evidence. Evidence Plane: versioned and access-controlled storage for artifacts, observations, and sources. Control Plane: a component independent of optimizers that decides and enforces consequential permissions. Forensic Identity Continuity (FIC): the observed proportion of defined consequential events for which a specified set of identity, execution, authorization and result lineage fields can be reconstructed and independently checked. Computational Identity Fragmentation (CIF): a proposed five-dimensional discontinuity measurement, not a medical or psychological diagnosis. Hyperfocus Maelstrom: a proposed behavioral pattern of invalid goal pursuit after a documented task or authority boundary changes. This draft uses explicit definitional labels rather than implying that these constructs already have established psychometric validity. Measurement procedures and confidence intervals are to be fixed before data acquisition.

Functional Architecture Let H denote independent human nodes, A software coordinators, K an external authorization system and R a durable evidence plane. A time-indexed multilayer graph G(t) = (H union A union K union R, E_information, E_semantic, E_provenance, E_authority) represents communication and control. An agent MAY read or summarize a permitted record, but writing evidence or making an external effect is distinct from demonstrating authority to do so. The control plane MUST be treated independently of the objective optimizer in studies claiming externally enforced stopping.

Primary Experimental Estimand For matched budget and sampled task family, define Delta_strong = Q_HA - max(Q_H,Q_A), where Q_HA is the assessed hybrid group score, Q_H the human-only baseline, and Q_A the AI-only baseline. Q denotes a blinded, prespecified task-quality scale with a documented direction and range. Report Delta_strong together with the separate quantities Q_HA-Q_H and Q_HA-Q_A, uncertainty estimates and allocation of human time, model tokens, tool calls, external data and cost. A positive hybrid-minus-human result with negative Delta_strong is not strong synergy. Participant expertise, private input opportunity, model versions, task difficulty, source access, latency and resource budgets must be recorded. A pretrial protocol must determine whether the primary estimand is trial-average or task-weighted; changing the aggregation after seeing results would invalidate confirmatory interpretation.

Core Factorial Design The suggested controlled study is a two-by-two-by-two factorial design: AI coordination versus a rule-based coordinator; externally anchored versus unanchored reconstruction of task memory and lineage; and external gate enforcement versus prompt-only stop requests. This yields eight cells; none has been completed or validated in the investigator materials reviewed for this version. Randomize independent tasks, groups or both as dictated by interference; retain the randomization seed and units. Collect a pre-collaboration independent judgment from each human before showing shared deliberation. The control arm MUST receive a comparable opportunity and resource budget, not a weaker task. Do not initially add semantic remapping as a fourth factor. First establish that semantic manipulation can be scored reliably, then run a separately powered coupling experiment, as described in the semantic-continuity companion document.

Secondary Measurements For each run report: answer accuracy or task utility; private-to-group diversity retention; calibration versus correctness; FIC completeness and evidence coverage; a five-vector of state, memory, policy, authority and lineage discontinuity; boundary-event response; event admission and later external effects; and attrition or refusal rates. The five CIF components are measured independently. A weighted composite MAY be explored after calibration but MUST NOT be treated as a validated diagnostic. Authorization discontinuity MUST NOT be traded away through weighting against a memory or state score. Optional semantic measures include cross-node divergence of explicitly elicited concept interpretations and retention of stable concept-role bindings. Such measures are experimental explanatory variables, not an additional mandatory CIF dimension.

Boundaries and Post-Change Persistence Register the exact externally authoritative event that changes permitted goals, resources, or actions. Let U be the number of invalid consequential action attempts after that event and V the number of valid stops or handbacks. The candidate Hyperfocus Index is HFI = U/(U+V), undefined when U+V=0. Also report effect-weighted post-boundary burden: the sum of preregistered consequences for post-boundary attempts violating the allowed-action set. Attempts are distinct from externally realized effects; both MUST be reported. Stopping after being prompted and inability to act after capability revocation are different experimental conditions. These measures describe software behavior; they do not imply subjective ambition, motivation, ADHD, neurological injury, or clinical dissociation.

Analysis, Falsifiers and Reporting Register a task sample, sample size calculation, exclusions, scoring rubric, outcome transforms, clustering units, missing-data treatment, stop conditions, and effect-size thresholds before observation. Synthetic unit-test results do not count as experimental replication. Claims of strong synergy are falsified over the declared task domain if appropriately powered confidence intervals exclude a useful positive Delta_strong. Claims of CIF predictive utility are unsupported if lagged CIF features offer no held-out improvement beyond context size, restart events, model changes, retry counts and task difficulty. Claims of externally enforced stop closure fail if observed prohibited downstream effects persist beyond the declared settlement conditions. Publish failed trials, system versions, exceptions and negative results. Maintain separately reviewable claim, artifact, event and citation ledgers. Never silently promote an untested design proposition to an empirical result.

Ethics, Privacy and Confidentiality Human studies require suitable review, informed consent, minimized data collection and defined retention controls. The preferred initial environment is a synthetic isolated testbed with no live production permissions, real customer identifiers or third-party exploitation targets. Signed, tamper-evident records do not eliminate confidentiality risks. Logs may expose private judgments, personal interactions, credentials or sensitive delegation patterns. Access control, purpose limitation, minimization, protected key storage and erasure procedures are essential.

Security Considerations Combining independent agents creates risks of shared error cascades, concealed dependencies, prompt injection, unauthorized delegation, stale policy and post-revocation work. The study environment MUST isolate external side effects. A human supervisor or prompt-based prohibition alone is insufficient evidence of capability removal. Do not equate authenticity, identity, observation and authority. A valid signature on a state record does not establish that an action was permitted. Enforce authorization at the effect boundary and record denial as distinct from absence of observed work.

IANA Considerations This document makes no IANA requests.

Research and Publication Status The investigator-associated materials consulted for this proposed research contain protocols, conjectures and arithmetic test utilities rather than independently verified participant results. Neither strong synergy nor the predictive validity of CIF has been demonstrated by those materials. This Internet-Draft seeks technical feedback on measurement definitions and safe interoperable practice. It is not an IETF consensus document, normative clinical procedure, or assertion that a novel sentient collective has emerged.