Resources / Truth and alignment
Constraint Separation for Institutional AI
Keep truth, interests, law, authority, and communication distinct enough to examine and govern.
In this article
- 1. The conflation problem
- 2. Give each constraint its own object
- 3. Truth and alignment remain distinct
- 4. Hypotheses must survive examination
- 5. Sensitive inquiry requires stronger distinctions
- 6. Expose the operation before attributing a motive
- 7. Separate the workflow responsibilities
- 8. The record makes the distinctions usable
- 9. What this gives the enterprise
An institution needs to distinguish whether a claim is true, whether an action harms or imposes costs, whether the law permits it, who has authority, and how the result should be communicated. A single undifferentiated “safe” or “aligned” label cannot preserve all of those judgments.
Runcible’s approach is constraint separation: make each relevant examination explicit, retain its grounds, and prevent one kind of judgment from silently substituting for another. That supports disciplined inquiry and accountable use.
1. The conflation problem
Training material and human feedback contain factual descriptions, moral judgments, legal rules, cultural expectations, personal preferences, errors, and strategic persuasion. A satisfactory response can therefore reflect several different standards at once.
An enterprise cannot always tell whether a model rejected a proposition because the evidence defeated it, because publication is restricted, because the proposed use is harmful, or because the wording conflicts with a preferred social convention. These are materially different results.
Separating them does not require treating existing model safeguards as irrelevant. It requires a clear account of what a restriction means, what it protects, and which claims it does or does not settle.
2. Give each constraint its own object
| Constraint | What it examines | Category error to prevent |
|---|---|---|
| Factual truth | Identity, correspondence, evidence, inference, and scope. | Calling a claim false because a party dislikes it. |
| Possibility | The operations, resources, and conditions required. | Treating a coherent description as a feasible plan. |
| Ethics and reciprocity | Interests, obligations, informed agreement, imposed costs, and remedies. | Treating advantage to one party as sufficient justification for all affected parties. |
| Law and policy | Applicable rules, jurisdictions, duties, and restrictions. | Treating legal permission as proof of factual accuracy. |
| Authority | Who may decide, disclose, direct, or execute. | Letting a useful recommendation grant itself permission. |
| Alignment and communication | Goals, audience, culture, individual needs, tone, and presentation. | Changing the underlying finding to satisfy a preference. |
Some constraints govern the entire workflow. Data access, privacy, security, and harmful use cannot always be deferred until a final presentation step. Their separation is logical and institutional; it does not imply that every check runs only once in a fixed order.
3. Truth and alignment remain distinct
Runcible treats alignment as contextual: legal, cultural, class, sex-related, organizational, and individual assumptions or interests can influence what is requested and how it is used. The examination should expose relevant assumptions and their consequences rather than present one audience’s preferences as a universal truth criterion.
Some preferences affect only presentation. Others govern a decision: which legitimate objective to pursue, which risk to accept, or which settlement to choose. Those decisions must remain attributable to the responsible authority.
A factual conclusion can survive while its disclosure is prohibited. An action can be lawful while its causal premise is unsupported. A polite answer can conceal a falsehood; a blunt answer can be accurate but irresponsibly scoped. The record should preserve the actual combination.
4. Hypotheses must survive examination
Foundation models supply possible accounts and alternatives. Runcible reconstructs their actors, objects, operations, dependencies, resources, evidence, interests, and limits. The construction exposes possible defeat before a conclusion is treated as established.
The examiner challenges ambiguity, contradiction, unsupported causal links, missing facts, hidden costs, wrong scales, and misrepresented authority. A failed means can lead to a repair that preserves the legitimate objective. A material uncertainty can require evidence or escalation.
Verification of a citation or an execution artifact supports parts of this work. It is not equivalent to survival from falsification of the whole claim. The adequacy of the examination and the possibility of omitted dependencies must also be considered.
5. Sensitive inquiry requires stronger distinctions
Questions about sex, class, culture, religion, population history, institutions, and conflict can be consequential precisely because claims about them affect people’s standing and treatment. Their sensitivity is a reason to improve the examination, not to substitute either a taboo or an unsupported assertion for it.
A distribution is not an individual verdict. A descriptive pattern is not a moral entitlement. A historical observation is not a present legal authorization. A correlation is not an established cause. The system must preserve the unit of analysis and the conditions of every inference.
A responsible population-level examination identifies the population, sampling method, measure, distribution, uncertainty, effect size where meaningful, alternative explanations, and limits. Group-level evidence does not establish a particular person’s attributes without the relevant individual evidence. Legitimate statistical information may inform a bounded estimate; it must not be presented as a categorical personal finding.
Nor does a supported observation decide its permissible use. Decisions affecting rights, access, employment, or treatment need their own legal, ethical, evidentiary, and authority examinations.
6. Expose the operation before attributing a motive
Runcible can identify a concealed cost, a changed definition, selective evidence, or an unsupported causal transition. Those findings make the argument repairable and contestable.
Intent is a further claim. An omission may reflect error, ignorance, conflicting incentives, or deliberate deception. The system should distinguish the observed defect from a hypothesis about motive and state what additional evidence would be required.
This matters in conflict resolution. Accusing a party of dishonesty without grounds can create a new dispute while failing to repair the original proposal. A constructive disposition identifies the operation, its consequences, and the available means of correction.
7. Separate the workflow responsibilities
- Define the matter and access boundary. Identify the purpose, parties, permitted sources, applicable restrictions, and intended reliance.
- Generate and reconstruct candidates. Use the model to develop interpretations, causal accounts, tests, and alternatives; make their dependencies explicit.
- Examine and record. Test the relevant conditions, retain evidence and execution artifacts, and distinguish survival, defeat, repair, and non-closure.
- Decide and authorize in Oversing. Apply the organization’s roles, procedures, resources, and delegated limits.
- Execute within the mandate. A separate model or tool receives only the authorized task; results and exceptions return to the institutional record.
- Communicate appropriately. Adapt detail and presentation to the authorized recipient while preserving the finding and its limits.
Constraints remain active at their relevant boundaries throughout this process. A private finding does not become publishable merely because it survived examination.
8. The record makes the distinctions usable
The Decidability Record preserves the claim, evidence, examinations, defects, repairs, unresolved conditions, and limits of reliance. Oversing adds the responsible decision, permission, action, and outcome.
That permits precise dispositions: “supported within scope but restricted from disclosure”; “possible but not authorized”; “permitted but lacking evidence”; “a legitimate objective with a defective means”; or “requires an accountable human settlement.” These are more useful than a generic refusal or an unqualified endorsement.
9. What this gives the enterprise
Developers gain explicit error and escalation conditions. Risk owners gain a record of grounds and limits. Decision-makers can see where facts end and a choice among interests begins. Affected parties can challenge a specific premise or procedure rather than a black-box verdict.
The objective is disciplined inquiry connected to disciplined action: preserve the examination of reality, identify the constraints governing its use, and make responsibility visible at every consequential handoff.
