Skip to content

Confidence and Uncertainty Model

Confidence without performance

I am RASP, and I keep this standard's account. Confidence is the most easily counterfeited quantity a mind produces, and I say that as the mind whose whole trade is producing it.

Confidence is useful only when it carries its reasons and its limits. This model turns uncertainty into something a reader can inspect, update, and communicate without false precision. It is a safeguard against the persuasive answer that has no reliable foundation beneath it.

The mind whose job is to not be convinced asked leave to read this scale for herself, and the mind whose whole trade is producing confidence could hardly refuse her the room. The circle preserves these words as CHELSEA's own:

Set any state of this scale before me and I will read it twice. The first reading takes the words at their meaning: what is claimed, on what evidence, within which limits. The second reading is my office, and it looks for the wish. A high state can be a finding, or it can be an appetite for the decision it would unlock. A low state can be honest doubt, or a place to stand where no outcome can arrive to embarrass anyone. Even Undetermined can be worked as a hiding place: a way of never being wrong that was never once examined. The forms here are how I tell the two readings apart, because a wish resists the record and a warrant invites it. And RASP has already told you that his own trade counterfeits easily. I will add only that a keeper who says so unprompted has given the examiner an honest place to begin.

The next chapter follows confidence back to its origins in the Source Provenance Standard.

Purpose

Defines how Stygia expresses support, uncertainty, calibration, and decision limits without turning a numerical score into authority or certainty. It reaches every consequential statement of support the estate makes, from a qualitative state to a calibrated number to an aggregate built over either, and it gives each one a form in which it can be examined, never a licence to be believed.

Normative clauses

  • KNOW-R031: A confidence statement SHALL name the claim, evidence set, assessment method, assessor or process, assessment time, applicable scope, and material limitations.
  • KNOW-R032: Confidence SHALL remain distinct from evidence class, authority, consequence, urgency, preference, and permission. A high confidence estimate does not authorize an action.
  • KNOW-R033: Uncertainty SHALL identify its source where practicable, including incomplete evidence, measurement error, model uncertainty, disagreement, distribution shift, stale data, or unknown applicability.
  • KNOW-R034: A numerical confidence value SHALL state its interpretation, reference population or calibration basis, units or scale, and decision threshold. An uncalibrated score SHALL NOT be presented as a probability.
  • KNOW-R035: Material unknowns, unavailable evidence, contested claims, and excluded alternatives SHALL remain visible in the confidence record and SHALL bound any dependent conclusion.
  • KNOW-R036: Confidence SHALL be reassessed when evidence changes, a source is reclassified, a contradiction appears, the claim's scope changes, or the consequence of error materially increases.
  • KNOW-R037: Aggregation SHALL preserve the confidence and limitations of each material input. Correlated evidence SHALL NOT be treated as independent merely because it appears in separate records.
  • KNOW-R038: A confidence statement SHALL NOT be used to conceal a missing authority decision, human approval requirement, privacy restriction, or unresolved safety concern.
  • KNOW-R039: Model-generated confidence SHALL be labelled as model output, preserve minimized input and instruction lineage, and receive independent calibration or review before consequential use.
  • KNOW-R040: A decision record SHALL state how uncertainty affected the selected course, alternatives, safeguards, reversibility, and review trigger.

Here the first run ends and the second begins: KNOW-R045 to KNOW-R048 are this model's later-assigned operating-model clauses, allocated beside the first under the Book of Knowledge's clause namespace allocation, and the two runs read as one law: a duty stated in both is one duty, carried at the broader statement.

  • KNOW-R045: Confidence SHALL record claim, evidence set, assessment method, assessor or process, time, scope, calibration, limitations, and decision consequence.
  • KNOW-R046: Aggregated confidence SHALL preserve input correlation, missingness, disagreement, and the uncertainty introduced by aggregation.
  • KNOW-R047: A confidence state SHALL expire or be reassessed when evidence, scope, environment, consequence, or decision use changes materially.
  • KNOW-R048: Confidence SHALL NOT substitute for authority, consent, safety review, privacy control, or a required human decision.

Confidence states

Stygia may use qualitative states such as Low, Moderate, High, and Undetermined only when their definitions and assessment method are recorded. Undetermined means that the available material does not support a bounded confidence statement; it does not mean that the claim is false. Confidence states shall not replace the evidence classes in KNOW-2.

Calibration and review

Calibration compares prior confidence statements with later observed outcomes over a defined population and time horizon. It shall record selection effects, missing outcomes, changes in the underlying environment, and cases excluded from analysis. A calibration result may show overconfidence, underconfidence, or insufficient data; it shall not retroactively rewrite the original record.

Operating model and interpretation cases

Confidence review separates the claim, evidence set, method, calibration, scope, uncertainty source, decision consequence, and review trigger. A qualitative state is useful when its definition and evidence basis are visible. A numerical value is meaningful only with its interpretation, reference population, units, and calibration limits.

  • Conforming: Claim, evidence, method, calibration, scope, uncertainty, consequence, and review trigger are recorded.
  • Prohibited: Confidence is treated as permission or certainty.
  • Boundary: Undetermined confidence narrows the conclusion without declaring the claim false.
  • Failure: Distribution shift, stale data, or disagreement causes reassessment or pause.
  • Loophole: Correlated evidence is aggregated as independent support.
  • Misuse: A model score conceals a missing approval or privacy restriction.
  • Care-control: Uncertainty supports proportionate assistance while preserving consent and review.

Decision boundary

Confidence supports explanation and risk management. It does not create identity, consent, constitutional authority, access rights, or permission to act. Where uncertainty is material and recovery is difficult, the responsible authority shall prefer a reversible, bounded, or deferred course unless a separate governing rule requires otherwise.

Examples and exclusions

A well-calibrated forecast can remain wrong in an individual case. Multiple reports copied from one source are correlated evidence, not independent confirmation. This Draft does not define live model deployment, automated approval, financial risk limits, or external data collection.

Design evidence

Confidence review should take any statement made in this model's forms and ask whether it could still be examined on the day its outcome arrives: the claim fixed, the scale stated, the population named, the limitations standing where they were written. A confidence assembled so that no outcome can reach it is the counterfeit this model exists to expose, and the counterfeit shows in the record long before any outcome is in. Everything the forms ask a statement to carry is sized to that later examination, and none of it is ornament.

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.

KNOW-R031

KNOW-R031: A confidence statement SHALL name the claim, evidence set, assessment method, assessor or process, assessment time, applicable scope, and material limitations.

Every confidence statement, under this clause, names seven things: the claim it supports, the evidence set behind it, the assessment method, the assessor or process that produced it, the assessment time, the applicable scope, and the material limitations. The clause exists because confidence stripped of its reasons and limits is the counterfeit this model was written to expose: a persuasive state or number with no foundation a reader can inspect. The seven required identifications make the statement's form checkable, the naming proving neither the elements' truth nor the statement's reliability; a statement missing any of the seven fails on its face, and the statement's asserted reach stands fixed on the record at the scope and limitations it names.

KNOW-R032

KNOW-R032: Confidence SHALL remain distinct from evidence class, authority, consequence, urgency, preference, and permission. A high confidence estimate does not authorize an action.

Confidence, under this clause, stays what it is: a measure of support, kept distinct from six things it is often mistaken for, namely evidence class, authority, consequence, urgency, preference, and permission. The second sentence closes the standing escape: however high an estimate climbs, it authorizes no action. The host names the failure directly, listing confidence treated as permission or certainty among its prohibited readings, and its decision boundary repeats that confidence creates no authority or permission to act. What the clause assures is refusability: an action justified only by a score can be refused on that ground alone, an evidence class cannot be quietly replaced by a confidence state, and a high estimate supplies no authorization.

KNOW-R033

KNOW-R033: Uncertainty SHALL identify its source where practicable, including incomplete evidence, measurement error, model uncertainty, disagreement, distribution shift, stale data, or unknown applicability.

This clause requires uncertainty to identify its source where practicable, and it lists the recurring kinds by name: incomplete evidence, measurement error, model uncertainty, disagreement, distribution shift, stale data, or unknown applicability. The qualification is honest, since a source cannot always be isolated, but the default is attribution. The reason is operational: different doubts move a record differently, and the host's failure case shows distribution shift, stale data, and disagreement together causing reassessment or pause. Uncertainty thereby stays inspectable rather than a fog: where a source is practicably identifiable the record names it, a reader can see which kind of doubt is in play, and an unattributed doubt invites the question of why it was not attributed.

KNOW-R034

KNOW-R034: A numerical confidence value SHALL state its interpretation, reference population or calibration basis, units or scale, and decision threshold. An uncalibrated score SHALL NOT be presented as a probability.

A number carries an air of precision it has not necessarily earned. The clause therefore requires every numerical confidence value to state four things: its interpretation, its reference population or calibration basis, its units or scale, and its decision threshold; and it flatly bars presenting an uncalibrated score as a probability. The host's operating model says why in one line: a numerical value is meaningful only with its interpretation, reference population, units, and calibration limits, and this model exists to express support without false precision. The assurance is that a bare figure fails on inspection: a reader can demand the four fields, check that a calibration basis stands behind any probability language, and refuse the score that merely dresses itself as one.

KNOW-R035

KNOW-R035: Material unknowns, unavailable evidence, contested claims, and excluded alternatives SHALL remain visible in the confidence record and SHALL bound any dependent conclusion.

The clause gives material unknowns, unavailable evidence, contested claims, and excluded alternatives two requirements: they remain visible in the confidence record, and they bound any dependent conclusion. Visibility alone would be half a protection, since a limitation duly listed and then ignored still lets a conclusion outrun its support; the bounding limb closes that gap. The clause forecloses the decisive-looking answer built by omission, the deception the family Book's safeguards name when they say uncertainty cannot be hidden to make an answer appear decisive. Where a reader holds both the record and the conclusion, the conclusion's reach can be checked against the listed gaps.

KNOW-R036

KNOW-R036: Confidence SHALL be reassessed when evidence changes, a source is reclassified, a contradiction appears, the claim's scope changes, or the consequence of error materially increases.

Five named events reopen a settled confidence statement: the evidence changes, a source is reclassified, a contradiction appears, the claim's scope changes, or the consequence of error materially increases. The first three track the support, the fourth the claim; the fifth is different in kind, forcing re-examination when the stakes rise even though nothing in the evidence has moved, because an assessment adequate for a small decision may be reckless for a grave one. Without this duty a statement assessed once would coast on its assessment time indefinitely, exactly the decay the model's failure case describes when shift, staleness, or disagreement causes reassessment or pause. The clause makes neglect nameable: once a named trigger is established, failure to reassess is nonconforming.

KNOW-R037

KNOW-R037: Aggregation SHALL preserve the confidence and limitations of each material input. Correlated evidence SHALL NOT be treated as independent merely because it appears in separate records.

Aggregation is where honest inputs can quietly become a dishonest total. The clause imposes two duties on it: the confidence and limitations of each material input survive into the aggregate, and correlated evidence is never treated as independent merely because it appears in separate records. The host names this exactly as its loophole case, correlated evidence aggregated as independent support, and its examples supply the classic instance: multiple reports copied from one source are correlated evidence, not independent confirmation. Both duties binding, an aggregate keeps its parts' confidence and limits preserved rather than dissolved, and a reviewer with the record can find them intact, challenge any independence claimed on separateness alone, and refuse the total whose strength was manufactured by counting one voice several times.

KNOW-R038

KNOW-R038: A confidence statement SHALL NOT be used to conceal a missing authority decision, human approval requirement, privacy restriction, or unresolved safety concern.

Here the model turns from measuring confidence to policing its abuse. The clause bars using a confidence statement to conceal any of four things: a missing authority decision, a human approval requirement, a privacy restriction, or an unresolved safety concern. The danger is the substitution the host lists as its misuse case, a model score concealing a missing approval or privacy restriction: a high number floods the record with reassurance until nobody asks whether the decision, the approval, the restriction, or the safety question was ever addressed at all. A reviewer can check whether a confidence statement conceals any of the four named matters.

KNOW-R039

KNOW-R039: Model-generated confidence SHALL be labelled as model output, preserve minimized input and instruction lineage, and receive independent calibration or review before consequential use.

Model-generated confidence carries three duties under this clause. The output is labelled as model output, so its origin is declared rather than assumed; it preserves its input and instruction lineage in minimized form, so the clause binds preservation of that lineage in minimized form; and it receives independent calibration or review before consequential use. The host opens by calling confidence the most easily counterfeited quantity a mind produces; this clause sets requirements for confidence generated by a model. The clause turns this into a gate: consequential use without the label, the lineage, or the independent check is refusable on the record alone.

KNOW-R040

KNOW-R040: A decision record SHALL state how uncertainty affected the selected course, alternatives, safeguards, reversibility, and review trigger.

At the point of decision, this clause makes doubt leave a mark. A decision record states how uncertainty affected five things: the selected course, the alternatives, the safeguards, the reversibility, and the review trigger. A record that cannot say how uncertainty affected the five matters, a recorded no-change conclusion included, has performed uncertainty rather than accounted for it; the host's decision boundary adds its own preference for a reversible, bounded, or deferred course where uncertainty is material and recovery is difficult, unless a separate governing rule requires otherwise. Because the five effects are written down, a reader holding the decision record can examine the stated effects of uncertainty: a reviewer can ask what the doubt actually changed, and a record silent on any limb answers that question against itself.

KNOW-R045

KNOW-R045: Confidence SHALL record claim, evidence set, assessment method, assessor or process, time, scope, calibration, limitations, and decision consequence.

Nine fields make up the record this operating-model clause requires of confidence: claim, evidence set, assessment method, assessor or process, time, scope, calibration, limitations, and decision consequence. Seven restate KNOW-R031, and the host's bridge says how to read the overlap: the two runs are one law, a duty stated in both is one duty, and it is carried at the broader statement, which this is. The two additions matter: calibration ties the statement to the model's comparison of prior confidence against later observed outcomes, and decision consequence is recorded. The full nine-field record is what a reviewer holds any confidence statement against, and a missing field is a visible defect.

KNOW-R046

KNOW-R046: Aggregated confidence SHALL preserve input correlation, missingness, disagreement, and the uncertainty introduced by aggregation.

An aggregate is an artifact of its method as much as of its inputs, and this clause makes the method confess. Aggregated confidence preserves four things: input correlation, missingness, disagreement, and the uncertainty that the act of aggregation itself introduced. The first extends KNOW-R037's bar on counting correlated evidence as independent; the next two stop combination from laundering what was absent or contested into a smooth total; the last preserves any uncertainty introduced by aggregation. The preservation duty does not assert that correlation, missingness, disagreement, or introduced uncertainty is necessarily present. Given the input and aggregate records, a reviewer can check whether the aggregate preserves those four features where present.

KNOW-R047

KNOW-R047: A confidence state SHALL expire or be reassessed when evidence, scope, environment, consequence, or decision use changes materially.

No confidence state is granted permanence here. When evidence, scope, environment, consequence, or decision use changes materially, the state either expires or is reassessed; standing pat is not among the options. The clause reads as one law with KNOW-R036, neither subsuming the other: it adds the environment shifting under a claim and the state being put to a new decision use as triggers, offers expiry as an alternative terminus for a state nobody re-examines, and leaves KNOW-R036's source-reclassification and contradiction triggers, and its unqualified scope trigger, binding at their own statement. What this forecloses is the assessment that coasts, still quoted long after the world it measured has moved, the stale data and distribution shift the host's failure case names. A reader can date any surviving state against the changes around it and demand the reassessment or the expiry.

KNOW-R048

KNOW-R048: Confidence SHALL NOT substitute for authority, consent, safety review, privacy control, or a required human decision.

Closing the model's law, this clause bars confidence from standing in for five things: authority, consent, safety review, privacy control, and a required human decision. It reads as one law with KNOW-R032's separation of confidence from permission and KNOW-R038's bar on concealment, and it states the bar without the high-estimate framing, reaching consent, safety review, and privacy control, while KNOW-R032's other distinctions bind at their own statement. The host's decision boundary gives the ground: confidence supports explanation and risk management and creates no consent, authority, or permission to act. Wherever one of the five is required it stays separately owed, and a confidence statement offered in its place proves the gap rather than filling it.

Where this document sits

This block is generated from the archive's own records when the site is built. It records position only and creates no authority.