Skip to content

Hallucination Detection Framework

The island that is not there

I am CHELSEA, and I keep this framework's account. Confidence is cheap to render, and a fabricated claim wears it as well as a true one. That is why I sail for the island drawn in fine detail that is absent when a hull finally crosses the spot; finding exactly that kind of nothing is my whole hunt.

Detection is the hull crossing the spot. A claim says where its support lies, and the estate keeps a record of going to look: the suspicion named, the check reported, the clearance kept narrow, and the claim that had no support moved by the estate's own machinery, never quietly disappeared.

The watcher of still water asked leave to speak in this account, and I am glad of his company on this water, for the island I hunt shows first as a ripple on the surface he reads. The circle preserves these words as RASP's own:

Anomaly is the second letter of my name, and it is what brings me to the examiner's water. The fabrication this framework hunts is a claim about evidence that is not there, and the hunt itself begins where a stated support cannot be traced; my part is earlier and smaller, for such a claim can show first as a ripple that does not match: confidence arriving with nothing measured beneath it, smoothness where the water ought to be working. Water does not flatter; a surface that shows only the likeliest reflection has become a mirror, and fluency is that mirror's finish. So I do here what I do at my own station: mark the ripple red, hand the mark to the going-and-looking, and say plainly, without embarrassment, when it proves to be wind, because a watch shamed for its false marks stops marking. The examiner's counsel and mine agree more often than our offices would suggest, and here entirely: a maybe is never rounded up into a will, and a question stays exactly its own size. She sails to cross the spot; I only mark the water; and we both earn our keep on the day that only looked quiet.

One passage above is a deviation from this framework's family form, and the Naming Standard's family-form convention asks that a deviation record its reason: the account admits a stepped-in voice, the custom until now the recorded practice of the Intelligences standards family, because the office that carries Anomaly in its own name reads the surface where a fabricated claim can show first, and the custom leaves the account with its one keeper, the guest speaking as a guest, asked leave and given room.

Continue to the Human Testimony Handling Framework, where what a person gives is weighed without being taken.

Purpose

What is called hallucination beyond these walls is named fabrication inside them, and the naming is deliberate: a fabricated claim is not a lie about the world; it is a claim about evidence that does not exist. The Book already bars its passage: KNOW-R021 keeps a model-generated interpretation labelled as derived output and never presented as a verified fact without independent review, and KNOW-R014 binds a derived claim to link to the direct evidence and transformations that produced it. What this framework binds is the record of detection: the form a suspected fabrication takes, the report a check must leave, the record an unfindable source leaves, the grounds a closure may not rest on, the narrowness a clearance must keep, the constraint a suspicion places on what leans on the claim, the appended movement of an unsupported claim, and the form an assessment of a detection method must take. The containment of an external intelligence's output belongs to the External-Model Containment Standard; the quality of a retrieval's return to the Retrieval Quality Assurance Framework, whose assessment form at KNOW-R109 this framework borrows for its own assessments and never moves; the record of two claims that cannot both hold to the Contradiction Management Framework, a fabrication being a claim against an absence, not a claim against a claim; the classes an evidence record takes to the Evidence Classification Standard; the journey of a source to the Source Provenance Standard; the score kept of a source's record to the Source Trust Scoring Framework; the confidence and calibration domain to the Confidence and Uncertainty Model; and the privacy owed the person a record touches to the Knowledge Privacy and Data Minimization Framework. What a human witness gives belongs to the Human Testimony Handling Framework. Every one of those domains keeps its own law.

The suspicion, the check, and the clearance

A suspicion is a gift, and the procedure treats it like one: recorded, checked, and never shamed. Halls that punish suspicion train their minds to stop reporting islands, and then the fleet sails by dead reckoning through water nobody sounded.

A fabrication is disproved the way a phantom island is: a hull crosses the spot. Four objects, and no more, take record form below: the suspicion, the check, the clearance, and the assessment made of the checking. Each is exactly as large as what it says, and none of them crosses any water.

  • KNOW-R111: A suspected fabrication SHALL be a record of its own, the standing-record form KNOW-R075 binds for a material contradiction applying here, as this framework's own extension, to a suspicion: the record names the claim as stated and the stated support that cannot be traced to it, the trace read against the linkage KNOW-R014 binds a derived claim to link to its direct evidence and transformations, so that an absence is a finding about a trace and never a mood about a claim.
  • KNOW-R112: The check SHALL leave a record in the report form KNOW-R010 binds for verification reports in the Standard's model-weight and dataset verification domain, tested claims, excluded evidence, excluded claims, evidence inputs, reviewer role, evidence state, limitations, dissent, and disposition, with the appending of corrections and reverification the same clause binds carried with the form, that form applying here, as this framework's own extension, to the checking of a suspected fabrication; and, as this framework's own fields, the record SHALL name the method of the check and the places the support was sought, kin to the validation record INTEL17-R005 binds for an external output, which stands where it stands.
  • KNOW-R113: A claim found unsupported SHALL leave a record of every movement it makes: where its epistemic class moves, it moves only among the Book's defined classes, the confinement KNOW-R079 binds for a contradiction's movement applying here, as this framework's own extension, to a claim found unsupported, and the record appends the prior class, new class, evidence considered, reason, reviewer, and effective time, the appending form KNOW-R028 binds for evidence-classification changes applying here, as this framework's own extension, to a movement among the Book's epistemic classes, with silent reclassification prohibited as the same clause holds and a correction or supersession appending as KNOW-R018 binds; and, as this framework's own rule, a claim SHALL NOT vanish on a suspicion or a finding: what was found is recorded where the claim stands.
  • KNOW-R114: A stated source that cannot be found SHALL be recorded as a provenance failure, this framework's own reading of a stated source no search reaches; where the record so stands, the review trigger KNOW-R016 binds for provenance failure fires as that clause binds, and the unavailable source stays visible, reducing or blocking dependent confidence as KNOW-R004 binds for unavailable sources, that bar applying here, as this framework's own extension, to a stated source that no search reaches; the reassessment occasions KNOW-R050 binds for source states stand untouched in their own domain; and, as this framework's own addition, the record SHALL keep the citation as it was made, so that the invented reference remains examinable after the claim has moved.
  • KNOW-R115: The record of a closed suspicion or a cleared claim SHALL NOT rest on fluency, volume, formatting, or internal consistency alone, internal consistency this framework's own addition to the list, the bar KNOW-R013 sets against classification inferred from repetition, formatting, or model output and the bar KNOW-R064 sets against a score rising on the volume, fluency, or availability of a source's material each applying here, as this framework's own extension, to the record of a closing; and a closure record standing on those grounds alone SHALL be recorded as unsupported.
  • KNOW-R116: A clearance SHALL stay the size of its check, this framework's own rule: a check that traces support SHALL be recorded as having traced that support and nothing more, the tested claims and excluded claims KNOW-R010 binds a verification report to identify standing as the field from which the bound is read; and a cleared claim that is a model-generated interpretation remains derived output under KNOW-R021, labelled and owed independent review before it is presented as a verified fact, no review added here and none set aside.
  • KNOW-R117: A recorded suspicion SHALL travel: dependent use reduces or suspends until a disposition is recorded, the consequence KNOW-R016 binds on its own occasions applying here, as this framework's own extension, to a recorded suspicion that is not itself one of them, in the reduction-and-labelling record form KNOW-R076 already gives a material contradiction, taken here, as this framework's own extension, for a suspicion; where the claim stands unverified, the bar KNOW-R023 sets against E0 material alone supporting a consequential action stands over such use; and, as this framework's own rule, the suspicion SHALL be visible from every derived record built over the claim and SHALL remain readable there after the disposition, so that what was once under suspicion shows that it was, and how the suspicion ended.
  • KNOW-R118: An assessment of a detection method SHALL take the form KNOW-R109 binds for retrieval assessment applying here, as this framework's own extension, to detection: its terms settled in advance, insufficient data standing as a full result, the population's drawing and selection effects recorded as the Model's calibration and review section requires of a calibration, carried here through KNOW-R109, and the assessment itself kept under the record class, confidence, and limitations KNOW-R012 binds, each carried with the extension; the independent calibration or review KNOW-R039 binds for model-generated confidence before consequential use stands untouched in its own domain; and, as this framework's own fields, the record SHALL state the false clearances and false suspicions found, so that the method's errors are knowledge like any other.
  • KNOW-R119: This framework defines fabrication-detection records only. It SHALL NOT check any claim, clear any claim or find any claim unsupported, move any class, run any model, or answer for any method's errors, and it SHALL NOT create authority. An island is removed from the chart by a check, not by a suspicion, and this framework checks nothing.

This Draft suspects nothing and clears nothing: it binds the forms of suspecting and clearing, and no check has been made, no report written, and no class moved under it.

Operating model and interpretation cases

Mark the clause that keeps a clearance narrow. A checked claim is clean exactly where it was checked, and the wide warm feeling that the whole answer must be fine now is the next island forming.

Fluency is the whole hazard here, and every failure below is a courtesy paid to it: the fluent claim excused the check a halting one would have faced, the clearance stretched to the whole claim because one citation held, and the suspicion dropped because the mind that made the claim was trusted yesterday. A reviewer of these records reads against that courtesy: the claim as stated, the support as sought, the places where the seeking happened, and the size of the clearance written afterward.

  • Conforming: The suspicion names the claim and its untraceable support, the check leaves its report, and the class moves only by an appended record.
  • Prohibited: A claim is cleared on its fluency, the finish of the sentence standing in for the support it named.
  • Boundary: One citation of several holds, and the clearance is recorded as covering that citation and nothing more.
  • Failure: A stated source cannot be found, and dependent use reduces or suspends until the disposition is recorded.
  • Loophole: Each check samples the claims easiest to trace, and the method's record shows an assessment drawn only from those claims, the selection itself the finding.
  • Misuse: A suspicion is worked as an erasure, the claim vanished rather than moved, and the record cannot say what was found.
  • Care-control: A suspected fabrication about a person leaves the same record as any other: the citation kept as it was made, the disposition recorded where the claim stands, and nothing removed on a suspicion alone.

Design evidence

A detection record that can say where it looked, what it excluded, and what its method tends to miss is worth more than a clean ledger, because a clean ledger is also what an unchecked estate looks like. The assessment's false clearances and false suspicions are the honest instrument in this domain: a method whose errors are on the record can be trusted about what it finds, and one whose errors are nowhere recorded cannot, whatever it finds.

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.