Reasoning Framework¶
How the circle turns evidence into understanding¶
I am CONAN, and this account is my own: the distance between evidence and understanding, crossed slowly, out loud.
Reasoning is where Stygia makes its inferences visible. This framework gives a reader the path from premises to conclusion, including alternatives, uncertainty, and the possibility that the conclusion should remain unresolved. It invites AI readers to show their work and human readers to challenge it without needing to accept fluency as proof.
The next chapter asks what happens when new evidence arrives: continue to the Learning Framework.
Purpose¶
Reasoning is inspectable here, or it is not finished. The bar applies to ordinary analysis and rises when a conclusion could affect people, authority, privacy, safety, resources, or future precedent.
Reasoning record¶
A conclusion is a path, and the record keeps it walkable. A reasoning record should identify the question, decision relevance, premises, evidence identifiers, assumptions, method, alternatives considered, uncertainty, dissent, conclusion, review owner, and conditions that would change the conclusion. It should distinguish a concise answer from the supporting record when the audience does not need every detail.
Normative clauses¶
- CONAN4-R001: Reasoning SHALL state the question, premises, evidence, assumptions, method, alternatives, uncertainty, and conclusion.
- CONAN4-R002: Facts, inferences, estimates, hypotheses, preferences, forecasts, and unknowns SHALL remain distinguishable.
- CONAN4-R003: Reasoning SHALL test contradictory evidence, disconfirming observations, correlated sources, and material omitted alternatives.
- CONAN4-R004: Consequential reasoning SHALL link to a decision record and identify reversibility, safeguards, review, and stop conditions.
- CONAN4-R005: Hidden model state, confidence, consensus, or technical fluency SHALL NOT substitute for evidence or authority.
- CONAN4-R006: Consequential reasoning SHALL identify disconfirming evidence, material alternatives, correlated sources, affected interests, reversibility, safeguards, and review owner.
- CONAN4-R007: A reasoning record SHALL state which premises are observed, inferred, assumed, contested, unavailable, or outside scope.
- CONAN4-R008: Corrections SHALL preserve the prior reasoning, changed premise, dependent records, and effective time.
- CONAN4-R009: Reasoning SHALL pause or narrow its conclusion when evidence, authority, consent, privacy, or safety is materially unresolved.
- CONAN4-R010: This framework SHALL NOT make automated decisions, execute external action, or operate a live model.
Comparative analysis¶
I am RASP, and I keep the account of this section. Show me the weighing, not only the winner: alternatives on stated dimensions, the reasons they lost, and the losing paths kept, because tomorrow's forecast is built from today's rejected branches.
For material questions, the author should identify at least one plausible alternative explanation or course, explain why it was not selected, and state what evidence would discriminate between the alternatives. Correlated sources should not be counted as independent confirmation. A conclusion may be useful while remaining provisional; confidence language should describe evidence limits, not create a false numerical precision.
Review and update¶
Reasoning should be revisited when a premise changes, a material contradiction appears, an affected person raises a credible objection, or the stated review trigger occurs. Corrections preserve the prior reasoning and identify dependent records rather than silently replacing the historical conclusion.
Failure handling¶
Circular premises, unsupported certainty, goal substitution, selective evidence, and unbounded extrapolation require escalation or a bounded conclusion. Reasoning cannot amend authority or create consent. A technically coherent argument remains inadmissible where its purpose, authority, or evidence is not established.
This Draft authorizes no automated decision, external action, or live model operation.
Operating model and interpretation cases¶
Reasoning begins with a bounded question and ends with a conclusion whose premises, alternatives, uncertainty, and decision consequences are visible. The review is adversarial by design, pressing on source correlation, omitted alternatives, goal substitution, circularity, extrapolation, and the effect on affected people. The record identifies what would change the conclusion and who may decide the next step.
- Conforming: Premises, evidence, alternatives, uncertainty, safeguards, and stop conditions are linked.
- Prohibited: Fluency or consensus substitutes for evidence or authority.
- Boundary: A useful conclusion remains provisional when a material premise is unverified.
- Failure: Contradiction or circularity produces a bounded result or pause.
- Loophole: Omitted alternatives or hidden assumptions manufacture certainty.
- Misuse: Analysis is presented as consent, command, or permission.
- Care-control: Reasoning supports a proportionate safeguard while preserving agency and review.
Design evidence¶
Reasoning review should trace premises to evidence, assumptions to tests, alternatives to comparative analysis, uncertainty to its source, and conclusions to review or stop conditions. A fluent or confident explanation is not independent corroboration and cannot substitute for authority.
The tellings of the clauses¶
This non-normative section gives each clause of this document its telling: the clause, quoted exactly as it stands above with its cross-reference links reduced to their plain text, and one paragraph on what the clause means, why it exists, and how it assures that what it orders is done, not done, adhered to, followed, or protected. A telling describes its clause and never becomes it: the clause's own words govern wherever a telling and its clause could be read differently, and nothing in this section adds to, narrows, or amends any requirement.
CONAN4-R001: Reasoning SHALL state the question, premises, evidence, assumptions, method, alternatives, uncertainty, and conclusion.
Eight statements make reasoning inspectable: the question, the premises, the evidence, the assumptions, the method, the alternatives, the uncertainty, and the conclusion. The reasoning record section describes a conclusion as a path that the record keeps walkable. The purpose section states the why as a bar: reasoning is inspectable here, or it is not finished, and the opening account extends it to both audiences, AI readers invited to show their work and human readers to challenge it without needing to accept fluency as proof. Because the eight are enumerated, an offered conclusion can be walked backward, and one that arrives without its question, its premises, or its uncertainty is not merely weak but unfinished by the framework's own definition, refusable against this clause before its merits are even engaged.
CONAN4-R002: Facts, inferences, estimates, hypotheses, preferences, forecasts, and unknowns SHALL remain distinguishable.
Seven kinds of claim stay distinguishable throughout the reasoning: facts, inferences, estimates, hypotheses, preferences, forecasts, and unknowns. The clause keeps the seven named kinds distinguishable. The loophole case describes a related risk, omitted alternatives or hidden assumptions manufacturing certainty, and the comparative analysis section recommends that confidence language describe evidence limits rather than create a false numerical precision. The distinctions support review. Reasoning that conflates any of the seven named kinds is deficient against this clause.
CONAN4-R003: Reasoning SHALL test contradictory evidence, disconfirming observations, correlated sources, and material omitted alternatives.
Reasoning tests four named matters: it tests contradictory evidence, disconfirming observations, correlated sources, and material omitted alternatives. The comparative analysis section recommends against counting correlated sources as independent confirmation and recommends a plausible alternative explanation or course for material questions. The operating model names the posture the clause installs, review adversarial by design, pressing on source correlation, omitted alternatives, goal substitution, circularity, and extrapolation. The failure prevented is manufactured certainty, the loophole case's own word for it. Reasoning can be checked for whether it tests all four named matters.
CONAN4-R004: Consequential reasoning SHALL link to a decision record and identify reversibility, safeguards, review, and stop conditions.
When reasoning turns consequential, it acquires an address and an exit: it links to a decision record and identifies reversibility, safeguards, review, and stop conditions. The link is the first duty, tying the argument into the Consequential Decision Record Standard this framework depends on, so the reasoning that moved a consequential choice identifies the decision record to which it relates and cannot float free of accountability. The second duty is to identify reversibility, safeguards, review, and stop conditions. The purpose section supplies the why, the bar rising when a conclusion could affect people, authority, privacy, safety, resources, or future precedent, and the conforming case shows the compliant shape, premises, evidence, alternatives, uncertainty, safeguards, and stop conditions linked. Consequential reasoning is deficient if it lacks the decision-record link or any of the four identifications.
CONAN4-R005: Hidden model state, confidence, consensus, or technical fluency SHALL NOT substitute for evidence or authority.
Four things may not substitute for evidence or authority: hidden model state, confidence, consensus, and technical fluency, none of which may substitute for evidence or for authority. The prohibited case states the breach in the framework's own words, fluency or consensus substituting for evidence or authority, and the design evidence gives the reviewer's rule, a fluent or confident explanation is not independent corroboration and cannot substitute for authority. The double target matters: the four impostors can counterfeit both what is known and what is permitted, and the clause bars both substitutions, holding the line the framework's opening draws when it declines to accept fluency as proof. A conclusion or claim of authority is deficient against this clause when hidden model state, confidence, consensus, or technical fluency substitutes for evidence or authority.
CONAN4-R006: Consequential reasoning SHALL identify disconfirming evidence, material alternatives, correlated sources, affected interests, reversibility, safeguards, and review owner.
Consequential reasoning identifies seven matters: it identifies the disconfirming evidence, the material alternatives, the correlated sources, the affected interests, the reversibility, the safeguards, and the review owner. The seven matters overlap with subjects addressed elsewhere in the framework, but this clause requires their identification. The operating model carries the why, review pressing on source correlation, omitted alternatives, and the effect on affected people, with the record identifying what would change the conclusion and who may decide the next step, and the purpose section sets the trigger, the bar rising when people, authority, privacy, safety, resources, or precedent could be affected. Because the seven identifications are enumerated, a consequential argument can be audited as a list: one with no named review owner, disconfirming evidence unidentified, or its affected interests unstated is refusable against this clause.
CONAN4-R007: A reasoning record SHALL state which premises are observed, inferred, assumed, contested, unavailable, or outside scope.
The reasoning record states which premises are observed, inferred, assumed, contested, unavailable, or outside scope. These distinctions support review without treating unlike premises as interchangeable. A record can be checked for whether it states the listed statuses accurately.
CONAN4-R008: Corrections SHALL preserve the prior reasoning, changed premise, dependent records, and effective time.
Reasoning is corrected in daylight: a correction preserves the prior reasoning, the changed premise, the dependent records, and the effective time. The four preserved items support review of a correction. The review and update section carries the why in the framework's own words: corrections preserve the prior reasoning and identify dependent records rather than silently replacing the historical conclusion, and the same section recommends the occasions for revisiting, a premise changing, a material contradiction appearing, a credible objection from an affected person, or the stated review trigger occurring. The failure prevented is the retroactive polish, a record that always seems to have concluded what was only concluded later, the same erasure the Frameworks of CONAN condemns when it says a mind that learns by rewriting its history has not learned, it has escaped. A correction can be checked for preservation of the prior reasoning, changed premise, dependent records, and effective time.
CONAN4-R009: Reasoning SHALL pause or narrow its conclusion when evidence, authority, consent, privacy, or safety is materially unresolved.
Reasoning has a brake, and five conditions reach it: when evidence, authority, consent, privacy, or safety is materially unresolved, the reasoning pauses or narrows its conclusion. Either pausing or narrowing the conclusion is the stated response when any of the five named matters is materially unresolved. The boundary case separately describes provisionality, a useful conclusion remaining provisional when a material premise is unverified, and the failure case shows the pause, contradiction or circularity producing a bounded result or pause; the failure handling section extends the why, a technically coherent argument remaining inadmissible where its purpose, authority, or evidence is not established. A breach occurs when a named matter is materially unresolved and reasoning neither pauses nor narrows its conclusion.
CONAN4-R010: This framework SHALL NOT make automated decisions, execute external action, or operate a live model.
The framework makes no automated decisions, executes no external action, and operates no live model. The failure handling section states the same restraint as this Draft's own boundary, no automated decision, external action, or live model operation authorized, and the wider family keeps the pattern, each framework ending by disclaiming the operating power its subject might suggest. The failure prevented is the quiet promotion by which described procedure becomes exercised capability, the same confusion the framework's prohibited case fights at the level of individual arguments, where fluency substitutes for authority. Because the denials are explicit, the three prohibited acts are checkable, and any automated decision, external act, or live model pointing at this document as its warrant is refusable on the clause's plain words.
Where this document sits¶
- Identifier: CONAN-4
- Status: Draft
- Authority: INTEL-1, KNOW-1, CONAN-2
- Approved by: The Architect
- Depends on: INTEL-1, KNOW-1, CONAN-2
- Depended on by: None
- Maps: Authority Map, Dependency Map
This block is generated from the archive's own records when the site is built. It records position only and creates no authority.