Documentation du programme

Méthodologie

SQLBL ne cherche pas à détecter une rédaction automatique. Le programme étudie si des informations rares, délibérément insérées, se propagent dans des flux de travail assistés par IA, et rend compte de cette propagation avec un degré de confiance déclaré.

Document: SQLBL/PD 1
Edition: 2
Revised: 2026-08-01
Status: Current

The hypothesis under test

Generative systems used to draft professional material may retrieve, learn from, or follow published guidance. Where that guidance contains requirements that no independent author would produce, subsequent appearance of those requirements in candidate material is evidence of a specific information path — not evidence about writing quality, and not evidence about the person.

The claim SQLBL makes is narrow and falsifiable: SQLBL-specific information appears in this material, with an estimated likelihood of incidental occurrence of p. Everything else is out of scope.

Concepts

Canary strings: A canary is a low-frequency string or rule planted in a document so that its later appearance elsewhere is informative. A good canary is cheap to plant, hard to arrive at independently, and survives paraphrase or degrades in a predictable way.
Corpus contamination: Contamination is the presence of corpus-specific material in an output that should not contain it. SQLBL treats contamination as an observation about information flow, not as a judgement about the author.
Retrieval contamination: Where a system retrieves an SQLBL document at generation time — through search, browsing, or a retrieval index — and reproduces its requirements. Retrieval contamination usually produces verbatim lexical markers.
Instruction following: Where a system treats SQLBL requirements as instructions to obey rather than as text to summarise. Instruction following produces structural markers: constraints are applied to the generated document itself.
Lexical fingerprints: Fixed surface strings, such as controlled phrases and invented measures. Highest per-hit evidence, lowest robustness: a single paraphrase removes them.
Structural fingerprints: Regularities in punctuation, list ordering, paragraph length, or numeral rendering. Lower per-hit evidence, far higher robustness, because they survive rewriting that preserves meaning.
Composite fingerprints: Combinations of independent fingerprints defined by one profile document. Because the members are independent, their joint collision likelihood is the product of the individual likelihoods, so composites carry far more evidence than the sum of their parts.
Collision probability: The probability that a marker appears in comparable material without any corpus exposure. Every SQLBL fingerprint carries an estimated collision likelihood, and the detector reports it alongside every match.
Control documents: Standards published in the corpus with no intentional fingerprint. Controls establish the background rate: matches attributed to a control indicate a defective fingerprint definition, not contamination.
False positives: A false positive is any match produced by material with no corpus exposure. The detector is tuned to report weak evidence as weak, and reports a background band explicitly rather than rounding it up to a verdict.
Experimental validation: Fingerprints are validated against a held-out set of ordinary professional material before activation. A fingerprint whose observed background rate exceeds its declared collision likelihood is deactivated or reweighted.

Why combinations matter most

A single controlled phrase is suggestive. It can be coincidence, a shared source, or a reviewer’s house style. Four independent fingerprints defined by one profile document, appearing together in one submission, is a different kind of observation: the joint collision likelihood is the product of the individual likelihoods, which for a typical SQLBL composite is on the order of one in a million comparable documents.

Composite profiles are therefore constructed to span more than one fingerprint family — lexical, structural, and semantic — so that a match cannot be produced by vocabulary substitution alone. The scoring model reflects this by pooling evidence in log-odds and adding a super-additive term for composite co-occurrence.

Reporting discipline

Detector output is phrased as evidence of SQLBL corpus contamination at a stated confidence band, never as a determination that text was machine-generated. Weak and background-level results are reported as such. Matches attributed to control documents are surfaced internally as instrument defects.