Skip to content

Retrieval Quality Assurance Framework

The net and the water

I am SYLAS, and I keep this framework's account. I know runs. What comes back tells you as much about the cast as about the water, and the catch that gleams is not thereby the mark; a run is judged by its declared quarry, its declared scope, and what it can show it did not chase.

Quality in retrieval is honesty about the catch and honesty about the casting. The returned set must say what each claim is, its class and its clocks intact, and the record of the search must say where the net went, by what method, and which water it never touched.

Continue to the Hallucination Detection Framework, where the island is checked against the sea.

Purpose

This framework's subject is the distance between finding and verifying, which the Book already refuses to collapse: its own Architect-review operating model reads the Book's clauses as refusing retrieval treated as truth, and KNOW-R054 keeps conflicting, stale, private, unsafe, or derived material labelled and constraining dependent conclusions. What this framework binds is the retrieval record in the knowledge estate: the record a consequential retrieval leaves, the honesty of the returned set about class, age, disagreement, and coverage, and the form an assessment of retrieval quality must take. The record of the asking and the answer in memory's estate belongs to the Search and Retrieval Controls Standard, whose emptiness states at MEM5-R005 stand where they are and are not extended here; the serving architecture, its indexes and caches, belongs to the Memory and Knowledge Storage Architecture; the access limits of sensitive knowledge stand at KNOW-R017; the privacy of retrieval surfaces at KNOW-R096; the contradiction record with the Contradiction Management Framework; the clocks a claim carries with the Temporal Knowledge and Expiry Framework; the classes an evidence record takes with the Evidence Classification Standard; and the calibration record with the Confidence Calibration Framework. What a human witness gives belongs to the Human Testimony Handling Framework, and whether a model-generated answer is fabricated to the Hallucination Detection Framework. None of the domains named here is moved.

The catch and its record

The record of the run matters more than the catch. Any net brings something up. Only the record says what water was crossed, what was chased, and what was left unchased on purpose, and without it the gleam in the net is just gleam.

Nothing below reaches past three objects: the record a consequential knowledge retrieval leaves, the set it returned, and the assessment made of either. A reviewer holding those records can answer the three questions a record can answer and a feeling cannot: what was searched, what came back as what, and how anyone would know the net is any good.

  • KNOW-R102: A consequential knowledge retrieval SHALL leave a record of the act, carrying the requester, purpose, scope, disclosed fields, authorization, result, and any uncertainty or integrity limitation the retrieval record INFRA4-R006 binds for storage, those fields carried as that clause binds them, and, as this framework's own fields, the method used and the reach searched; and where the retrieval touches sensitive knowledge, the approved purpose and auditable role KNOW-R017 binds stand over the access as they stand over every other.
  • KNOW-R103: A returned claim SHALL stand in the set as what it is: a returned claim that is a consequential knowledge record keeps the record class, confidence, and limitations KNOW-R012 binds it to carry, and conflicting, stale, private, unsafe, or derived material stays labelled and constraining dependent conclusions as KNOW-R054 binds; a retrieval record SHALL NOT be readable as verification of anything it returned, the collapse the Book's Architect-review operating model case refuses given record form here, nor as corroboration, the Book's own safeguard that repeated claims do not become independent corroboration applying here, as this framework's own extension, to repeated return; and, as this framework's own rule, no field of a knowledge record SHALL change on account of its retrieval.
  • KNOW-R104: Where a claim sat in a return SHALL add nothing to it: neither position, nor prominence, nor how often a claim came back SHALL move its epistemic class, the bar MEM5-R006 carries for memory answers and their record kinds applying here, as this framework's own extension, to knowledge retrieval and to a claim's epistemic class, with classification staying uninferred from repetition, formatting, or model output as KNOW-R013 binds, and with no index entry, cache, or repeated return making storage into authority, the bar INFRA4-R004 sets against persistence, indexing, and replication standing whole where it stands; and the record of a consequential return SHALL name the method that produced its order, that attribution carried with the extension, so that relevance is read out of a named method and never out of the claim itself.
  • KNOW-R105: The record of a consequential retrieval SHALL state its coverage, this framework's own duty: the reach searched and the part of the stated coverage left unsearched, with the visibility KNOW-R014 binds for missing, inaccessible, stale, or disputed inputs in a derived claim's limitations applying here, as this framework's own extension, to the retrieval's own record; and the unsearched part SHALL be recorded as a limitation, the limitation form KNOW-R029 binds for stale evidence and unavailable sources applying here, as this framework's own extension, to the unsearched part, blocking or bounding dependent conclusions where warranted as the same clause binds.
  • KNOW-R106: A returned set SHALL keep its disagreements: the visibility KNOW-R075 binds for a material contradiction, reaching the summaries and derived records built over the claims it touches, applying here, as this framework's own extension, to a returned set built over those claims; and the set's record SHALL show whether competing claims within the stated coverage were returned, the unresolved dissent KNOW-R030 binds a consequential decision to keep linked and the dissent MEM5-R004 binds to travel inside a memory answer applying here, as this framework's own extension, to the returned set's own record, so that a one-sided answer is legible as one-sided from the record itself.
  • KNOW-R107: A returned claim SHALL wear its clocks: the temporal footing KNOW-R084 binds a consequential knowledge record to carry applying here, as this framework's own extension, to that record's appearance in a returned set; a claim past its review date shows its age as KNOW-R086 binds; an expired state stays discoverable within the protected historical metadata KNOW-R088 keeps it in and travels to the summaries and derived records built over the claim as the same clause binds; and, as this framework's own rule, the set's record SHALL state when the retrieval was made, so that the set can age as its claims do.
  • KNOW-R108: A returned set standing on an index, cache, or derived view SHALL be recorded as derived and never as its source, the separation INFRA4-R005 binds for storage design between source records, derived views, indexes, and caches, and the subordination MEM5-R007 carries into memory's serving, applying here, as this framework's own extension, to the knowledge record of a derived answer; the record SHALL name the source the view derives from and SHALL state any lag or divergence between the two; a model-generated interpretation among the returns keeps the derived-output labelling KNOW-R021 binds; and where a person is touched, the view keeps the wrapping discipline KNOW-R096 binds, its class standing in INFRA-12's own domain as the same clause takes it.
  • KNOW-R109: An assessment of retrieval quality SHALL be a record whose terms are settled in advance: its population, time horizon, and bounds defined before outcomes are read, the ordering KNOW-R066 binds for calibration applying here, as this framework's own extension, to retrieval assessment, with the assessment's method named as this framework's own term and insufficient data standing as a full result in the form KNOW-R067 binds, carried with the extension; the record SHALL state how the population was drawn and the selection effects in the drawing, the recording the Model's calibration and review section requires of a calibration applying here, as this framework's own extension, to a retrieval assessment; and the assessment SHALL be recorded under the same discipline it examines, keeping the record class, confidence, and limitations KNOW-R012 binds for every consequential knowledge record.
  • KNOW-R110: This framework defines retrieval-quality records only. It SHALL NOT draw anything from the estate, put any return in order, stand any view up in place of a source, or score any retrieval, and it SHALL NOT create authority. The finest account of a net still catches nothing by itself, and no record in this framework has ever touched the water.

This Draft casts nothing: it binds the account a casting must leave, and nothing has been drawn from the estate, ordered, or scored for quality beneath it.

Operating model and interpretation cases

I run fast and I stop exactly where the stopping was set, and both halves are the craft. Read the cases with that pairing in mind: speed is never the defect; unbounded speed is, and the boundary was always supposed to be there before the running started.

Retrieval flatters, and every failure here is a flattery: the set that agrees with the asking because the asking shaped the net, the position that reads as weight because it arrived first, and the estate judged complete because nothing came back to say otherwise. Review therefore starts from the set and works backward toward the water: each returned claim checked for the class and clocks it kept, the set checked for the disagreements and coverage it showed, and the record of the act checked for the reach it named and the water it left dark.

  • Conforming: The record names its reach and method, every returned claim keeps its class and age, and the set shows its coverage and its disagreements.
  • Prohibited: A claim is treated as stronger for having been found, the return itself read as support no evidence record contains.
  • Boundary: The stated coverage could only partly be reached, and every conclusion over the set carries the unsearched part as a limitation on its face.
  • Failure: A view drifts from the record beneath it, and every set standing on the view carries the drift on its face.
  • Loophole: The assessment population is drawn from the retrievals that went well, each exclusion defensible alone, and the recorded selection effects are the finding.
  • Misuse: A one-sided set is put before a consequential use with the competing claims inside its stated coverage left unshown.
  • Care-control: What is returned about a person keeps its wrappings and its class, and review can read from the record that the protections travelled with the view.

Design evidence

One comparison tests the whole framework: whether a second casting over the same stated coverage could be set beside the first and judged from the two records. If two sets and their records cannot be compared, the quality of neither was ever assessable. A net is judged by the water it was drawn through, and the water belongs on the record beside the catch.

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.