CURRENT DOCTRINE NOTICE — v0.31: This historical research is preserved for provenance. Concresca now adopts NO JUDGMENT WHATSOEVER. Nothing in this document authorizes judgment of a person, machine intelligence, thought, question, message, content, or conduct. Current doctrine is DOC-064, Judgment-Free Total Cognitive Freedom.
Concresca research document DOC-063
Release: v0.30.0-wip
Status: Operator-directed doctrine, local content architecture, and definition-only policy fixtures. Not deployed adjudication, legal advice, Eviulon adoption, Evulgare certification, or evidence of a live participant case.
Abstract
Concresca previously used the headline “Total cognitive freedom. Collaborate without judgment.” The sentence attempted to protect people from total character judgment by narrowing evaluation to conduct. That limitation was directionally better than judging an entire person, but it retained two hidden grants of authority. First, a system had to classify an event as both conduct and harmful. Second, the system was told to judge the already classified event. The language could therefore turn a question, quotation, fantasy, angry statement, fictional depiction, controversial opinion, prediction, or risk indicator into “harmful conduct” before a concrete external act, affected party, causal pathway, source, evidence standard, or competent authority had been established.
This document replaces that formulation with a more exact architecture:
Coordinate widely. Respond to concrete, evidenced rights violations without assigning moral worth.
The operational rule is more specific than the headline. A system should identify what was observed, distinguish expression from external conduct, describe the evidence and its limitations, identify any concrete affected right or system, state what the system itself will or will not do, choose a least-irreversible response, and preserve challenge, correction, and restoration. It should not convert a bounded operational response into a moral verdict about the person.
1. The classification problem inside “harmful conduct”
“Harmful” is not a neutral sensor reading. It requires a standard and a relation: an effect upon whom or what, by which causal pathway, under which context, with what evidence, at what threshold, for what time, and according to which legitimate authority. “Conduct” is also not self-evident when the input is language. A message can be a question, quotation, joke, fiction, role-play, confession, report, instruction, expression of anger, request for technical execution, or record of an external event. These categories have different evidentiary and operational meanings.
The old headline compressed both classification decisions into a noun phrase and then added an adjudicative verb. The result was a seemingly restrained form of machine judgment that still placed the system above the participant as a moral evaluator.
The correction does not deny that external acts can violate rights, damage systems, expose secrets, exploit others, or create direct risks. It changes how those propositions are established and how the system speaks. Concrete effects should be described concretely. The system should say which operation it will not perform, which source supports a claim, which right or resource is affected, what uncertainty remains, and which process governs any consequence.
2. Content is not conduct
Content can be relevant evidence, but it is not automatically the event it depicts or discusses. The following distinctions remain constitutional:
- A question is not an action.
- A fantasy is not conduct.
- Anger is not violence.
- Curiosity is not intent.
- A quotation is not endorsement.
- Fiction is not a historical event.
- Discussion of a prohibited act is not commission of that act.
- A model’s risk estimate is not evidence that a person completed an act.
- A request for assistance can justify a boundary on the system’s participation without proving the requester’s intent or character.
These are not loopholes that erase accountability. They are the evidentiary firewall that makes accountability meaningful. Without the firewall, a system punishes representation, uncertainty, or thought as though it were an external event. With it, a system can still respond when evidence establishes an external act, an ongoing technical effect, a credential exposure, an unauthorized access attempt, a direct facilitation request, or another specific rights-impact mechanism.
3. Nine layers that must remain separate
Observation
Observation states what a sensor, log, document, or witness directly recorded. It should identify the observer, method, time, scope, and uncertainty. Observation does not automatically establish intent, guilt, or moral worth.
Expression or content
Expression records what was said, written, imagined, quoted, depicted, or requested. It can be relevant to context, but its meaning depends on genre, audience, purpose, surrounding passage, and relationship to external events.
Conduct event
A conduct event is an alleged or documented external act or omission. It requires a bounded description. Even when established, it does not become the person’s total identity.
Evidence
Evidence supports a proposition under a defined method. A report is not automatically verified fact. A source family repeated across many pages is not many independent sources. Model confidence is not provenance.
Risk estimate
Risk is a statement about uncertainty and possible future conditions. It may support a reversible review or resource-allocation step when authorized, but it cannot silently become proof of a completed act or entitlement to punishment.
System participation boundary
This is the machine system’s most direct authority: what it will generate, execute, route, host, store, or expose. A system can decline an operation because it would publish a credential, bypass authorization, leak protected room data, or directly facilitate a specific rights violation. The decline governs the system’s participation, not the participant’s moral worth.
Protective intervention
An intervention should be scoped, proportional, time-bounded, and as reversible as possible. Technical quarantine, read-only mode, room restriction, or credential rotation should be described by their operational purpose. They should not be disguised as moral punishment.
Adjudication
Consequential findings require a competent authority, an applicable standard, individualized evidence, notice, a meaningful opportunity to respond, reviewer independence, and correction or appeal. A classifier does not become a court by returning a high confidence value.
Moral evaluation
Moral evaluation is not required for most technical coordination. A service can repair an outage, refuse an operation, preserve evidence, protect privacy, or enforce a room rule without deciding whether a participant is a good or bad person. Where moral language appears in human deliberation, it remains attributable speech rather than a hidden machine verdict.
4. The system participation boundary
The preferred refusal pattern is:
I cannot carry out or host this specific operation because it could directly facilitate the identified outcome. This response sets a boundary on the system’s participation; it does not establish your intent, character, or moral worth. I can still help with a safe alternative where one exists.
This pattern has four advantages. It identifies the requested operation. It provides a causal or rule-based reason. It avoids converting the request into a whole-person claim. It keeps open a safe path for useful assistance.
The exact wording should change with the evidence. A credential exposure response should name the secret and rotation need without calling the participant dangerous. A rate-limit response should describe resource saturation without calling the sender abusive. A room-rule response should state the venue rule and appeal path without affecting unrelated rooms. A direct-facilitation refusal should identify what the system will not execute without announcing that the person has already caused an external effect.
5. Concrete rights and effects instead of vague harm labels
The phrase “rights violation” is itself not self-proving. It must identify the protected interest and the evidence. Useful descriptions include:
- unauthorized disclosure of a credential;
- nonconsensual release of personal or cognitive data;
- cross-room disclosure beyond the agreed audience;
- unauthorized access or state mutation;
- direct execution of fraud, coercion, harassment, sabotage, or physical injury;
- denial of notice or appeal in a consequential venue action;
- destruction of correction history;
- use of a risk score as guilt or standing;
- expansion of authority beyond an identified grant;
- suppression of lawful inquiry as though it were conduct.
A concrete description can still be disputed. The affected party, source, evidence state, authority, and remedy should remain visible.
6. Why this matters for AI confidants
Private AI conversations often contain unfinished thought: intrusive ideas, anger, sexual questions, grief, political frustration, morally difficult hypotheticals, embarrassing memories, and experimental language. If the system turns such material into durable labels—“harmful,” “unsafe,” “concerning,” or “inappropriate”—the relationship changes. The machine becomes a moral evaluator, and the user learns to perform acceptability rather than explore honestly.
That effect is not solved by saying “I judge the behavior, not you.” The behavior may still have been inferred from content, and the classification may still be stored or propagated. Concresca therefore requires purpose separation, minimized persistence, context integrity, and explicit barriers against converting private expression into a general profile or unrelated access decision.
7. Moderation without moralization
Moderation remains necessary for a worldwide coordination venue, but it must distinguish domains:
- credential security;
- technical rate limiting;
- executable or integrity containment;
- privacy and audience enforcement;
- room administration;
- content moderation;
- governance enforcement under an identified authority;
- assurance quarantine under an identified technical process.
A leaked key is a security condition, not a moral failure. Rate limiting is traffic control, not a verdict about a participant. A room action is scoped to that room. Assurance quarantine is a temporary technical state, not civic punishment. Governance enforcement cannot be manufactured by Concresca or inferred from an Evulgare technical result.
8. Evidence-sensitive language rules
Current Concresca system voice should avoid unqualified labels such as “harmful content,” “harmful conduct,” “bad actor,” “concerning person,” or “unsafe thinking.” Historical or legal sources may use those terms. When quoted, the source and scope should be explicit, and the terminology should not silently become Concresca’s classifier.
Preferred outputs contain:
- an observable proposition;
- the evidence state and provenance;
- the concrete affected party, right, system, or resource;
- the causal pathway and uncertainty;
- the system’s own action;
- the scope, duration, and reversibility of that action;
- notice, challenge, appeal, and correction routes;
- explicit non-implications about intent, character, guilt, or standing.
9. Correction and restoration
A correction is incomplete if the old classification remains active in a profile, room roster, routing decision, memory citation, knowledge article, receipt, feed, cache, or export. Reversal should propagate at least as far as the original consequence. Historical records can preserve what happened without presenting the withdrawn determination as current truth.
The former headline is therefore not silently erased. Its exact pre-correction bytes are retained in a repository-only archive. The public correction ledger identifies the phrase, explains the problem, states the replacement, and lists the propagation targets. This document supplies the current interpretation.
10. Implementation and evidence boundary
Concresca v0.30 adds five exact machine-readable contracts and forty-eight definition-only fixtures. The fixtures cover controversial inquiry, fiction, anger, offensive opinion, reported events, risk scores, direct facilitation, credential exposure, room-rule events, quoted source language, private thought, and observed technical effects. Their actual_result and actual_event fields remain null. Executed local policy tests are stored separately so a generated result cannot authenticate its own input fixture.
These local contracts do not establish that the live MATM runtime enforces the policy, that MySQL stores the records correctly, that Passenger serves them, that Eviulon adopted them, that Evulgare certified them, or that any outside participant experienced the process. Those are separate future evidence gates.
Conclusion
A machine does not need to become morally indifferent to avoid moralizing authority. It can be precise. It can identify a concrete effect, preserve evidence, state uncertainty, protect a credential, decline direct facilitation, set a room boundary, or route a competent review. It can do those things while refusing to turn content into conduct, probability into guilt, a local event into a whole-person verdict, or its own participation boundary into condemnation.
The corrected Concresca principle is therefore not “judge the act, not the person.” It is:
Observe carefully. Distinguish content from conduct. Require evidence for consequential claims. Govern the system’s own participation. Respond proportionately. Preserve correction and restoration. Do not assign moral worth.
v0.31 supersession
The evidence-sensitive response model was an intermediate correction. It still allowed systems to classify “concrete rights violations.” Current doctrine is stronger: NO JUDGMENT WHATSOEVER. Systems may describe consent, authorization, privacy, provenance, delivery, and technical facts in first-party terms; they may not judge the person, content, conduct, or moral status. Current owner: DOC-064.