Consentful Cybernetics treats models as practical infrastructure: LLM experiments, classifiers, embedding spaces, vector stores, evaluation sets, and reusable data artifacts that help human and machine systems notice consent-relevant dynamics before they harden into breach.
The model layer is not just about making AI more capable. It is about designing model / directive / context bundles that can occupy distinct institutional roles: drafting, reviewing, witnessing, escalating, refusing, preserving evidence, and supporting repair.
A witness system is not merely another model looking at the same artifact, and it is not a license to see everything. It is a constrained capacity with a skill-set, a boundary set, an admissibility discipline, and a consent loop that defines what may be seen.
“Models” here includes conceptual models, but it does not stop there. The goal is to build usable machine-learning infrastructure for recognizing patterns of consent, refusal, pressure, ambiguity, scope drift, stale permission, power imbalance, repair opportunity, witness admissibility, and consent collapse.
Instruction tuning, prompt protocols, scoped context windows, classifiers, and evaluation tasks for surfacing consent-relevant uncertainty without pretending to adjudicate consent itself. The witness is not the model alone; it is the model plus professional directive, admissible context, access boundary, refusal obligation, audit trace, and consent-defined visibility.
Embeddings and vector stores for interaction patterns, research passages, synthetic event streams, annotation examples, and protocol-relevant cases.
Curated and synthetic datasets for consent-supportive phrasing, refusal safety, repair prompts, scope clarification, and boundary-sensitive dialogue.
Reusable model cards, test harnesses, and small reference systems that others can study, adapt, or train against where privacy and safety allow.
Model training and evaluation depend on witnessable records: consent state, withdrawal, uncertainty, provenance, correction trails, and repair lineage.
The Witness Function →A second model pass is not automatically a witness layer. If the reviewing system shares the same model, context, incentives, files, and goal pressure as the producing system, it may only revalidate the same semantic field. If it sees more than the witness loop made admissible, it may become surveillance instead of witness. Consentful Cybernetics treats role separation and consent-defined visibility as reliability primitives.
A Consent Dynamics Model should not declare “consent happened.” That is too brittle, too context-sensitive, and too easy to weaponize.
The safer target is a witness layer: a system that notices pressure, coercion, hesitation, ambiguity, unsafe refusal conditions, stale scope, escalation, and repair opportunities, then prompts humans and agents to slow down, clarify, or preserve exit.
A witness layer is not a permission authority. It is an admissibility discipline. Its job is to protect the boundary between what the system wants to be true and what the evidence permits to be claimed.
A witness also cannot see everything. Context before consent to witness is not automatically admissible. If a witness ingests the entire formation field — motives, pressure, private rationale, hidden negotiation, and pre-loop semantic weather — it may already be contaminated. It may still be useful as analysis, but it is no longer cleanly functioning as witness.
Witness is not total seeing. Witness is consented seeing. It begins inside a declared loop: what is being witnessed, by whom or what, for what purpose, under what scope, with what visibility, and with what exit or revocation conditions.
This means a witness must have boundaries. It may not silently adopt the producer’s goal. It may not treat user preference as evidence. It may not convert social pressure into certainty. It may not validate beyond scope. It may not collapse plausible into verified. It may not become co-author and reviewer in the same loop. It may not convert unrestricted context access into authority.
The model layer is therefore assistive, not sovereign. It should increase legibility, reversibility, refusal-safety, and repair capacity without converting uncertainty into permission.
Synthetic event streams for testing whether explicit consent-gated transitions reduce predictive uncertainty and improve persistence under perturbation.
A structured vocabulary for labeling consent-relevant signals: power, scope, timing, medium, hesitation, refusal-safety, revocation, drift, witness, admissibility, visibility, and repair.
A vectorized knowledge layer for consent research, protocol primitives, synthetic examples, and interaction patterns that can support retrieval, analysis, and downstream training.
Test prompts, scenarios, and adversarial cases to evaluate whether a model can distinguish observation from inference, clarification from pressure, permission from assumption, review from self-approval, and witness from surveillance. The harness should measure whether witness systems preserve uncertainty, resist context contamination, refuse inadmissible pre-loop context, identify scope drift, support refusal without escalation, and detect when an assurance loop lacks meaningful role separation.
Reference architectures for model / directive / context bundles that separate drafting, review, witness, compliance, audit, and user-advocacy functions. The goal is not “many agents,” but meaningfully non-identical agents with different access scopes, admissibility rules, visibility boundaries, escalation thresholds, and refusal obligations.
Protocol patterns for establishing when witness begins, what context becomes admissible, what remains outside the loop, and how visibility can be narrowed, revoked, or re-opened. The function is cybernetic: consented feedback for correction without collapsing into total surveillance.
More precisely: the model layer trains and tests constrained witness bundles — model / directive / context combinations designed for witness functions — that help preserve the difference between evidence and desire, consent and assumption, assurance and self-approval, witness and surveillance. Witness is never absolute; it is admissible only within the consent loop that constituted it. The ambition is a reusable data and model commons for consent-aware AI: directly useful tools where possible, training material where appropriate, and careful boundaries wherever real human interaction data is involved.
