CatQualia/discriminator-negatives

CatQualia discriminator negatives

96,330 rows used to train a discriminator that separates a real cross-domain structural mapping from one that merely looks like one. Each row is a candidate mapping together with the judge's verdict on it.

This corpus is interesting because of what it encodes: an operational definition of a bad structural mapping, with the failure mode named.

Read this before using it: the filename overstates the negatives

Despite the name, most rows are not negatives. Measured distribution over the real grade field:

grade	Rows	Contains a failure class?
GENERATIVE	80,070	no — failure_class is absent/None on all of them
STRIP	16,257	mixed — see below

So the corpus is best understood as a graded candidate pool with a negative subset inside it, not as a pure negative set. Anyone using it as a clean negative class should filter on failure_class first.

Measured failure_class distribution across all 96,330 rows:

failure_class	Rows
(none — absent/None)	80,070
phantom	13,990
typical	2,263
decorative	4

13,990 + 2,263 + 4 = 16,257 rows carry an actual failure class. The rest do not.

The alien_or_typical field tracks this: alien 80,070, typical 13,990. A row marked typical is one the judge judged to be an ordinary, non-distinctive mapping — which is itself a failure mode for this pipeline, whose whole purpose is to find structure that is not obvious.

Schema

Fields of the first record, read from the file in this repository:

Field	Type	Meaning
ts	string	Timestamp of the judgement
grade	string	GENERATIVE or STRIP
alien_or_typical	string	alien / typical — distinctiveness verdict
mechanism	string	The mechanism being mapped
candidate_artifact	string	The software artifact the mapping proposes
isomorphism	string	The claimed structural correspondence
why	string	The judge's reasoning
failure_class	string / None	How the mapping fails, when it does
floor_checked	(varies)	Whether a quality floor was applied
provenance	string	Where the row came from
seed_dataset	string	The seed corpus the row was generated from

The three failure classes are worth stating plainly because they are the useful part of this dataset:

• phantom — the mapping asserts a correspondence that is not there. This is the apophenia failure: shared vocabulary without shared structure.
• typical — the mapping is real but ordinary; the system was built to find the non-obvious, so a mapping anyone would make is a failure for this purpose.
• decorative — the mapping is true but produces nothing buildable; it is a description, not an artifact.

Measured facts

Property	Value	Command
Rows	96,330	wc -l < quality_judge_negatives.jsonl
Bytes	68,114,930	stat -c%s quality_judge_negatives.jsonl
Distinct failure_class values	4 (incl. none)	full-file pass

Duplication in the author's tree

This exact file (md5 b21f78ecc87271773b0c93d30ba2e831) exists under four to five other paths in the author's tree: 06_forge/forged_quality_judge_negatives_seed.jsonl, 61_watermark_splits/quality_judge_negatives_clean.jsonl, 70_leverage_remaining/vaginal_atlas_watermarks_split_quality_judge_negatives_clean.jsonl, and inside the dated snapshot directory. A watermarked variant also exists at 68,173,336 bytes with the same row count. This repository is the canonical published copy — the others are the same bytes under different names.

How it was produced

By the author's quality-judge pipeline: candidate mappings generated by the isomorphism engine were graded, and the judge's reasoning and failure classification were recorded. The method is described at https://catqualia.com/isomorphism/.

Known limitations

• Self-judged. Every verdict here comes from the author's own judge, not independent human review. The failure taxonomy is the system's own.
• Not a balanced classification set. 83% of rows carry no failure class; do not compute accuracy on the raw file without filtering.
• The why field is judge-generated prose and may rationalise a verdict rather than explain it.
• English only.
• floor_checked is not consistently populated and its semantics are not documented in the file; treat it as unreliable.

Licensing and attribution

CC BY 4.0. Attribution: Christopher Betances, catqualia.com.

Citation

@misc{betances_catqualia_discriminator_negatives,
  author       = {Betances, Christopher},
  title        = {CatQualia discriminator negatives: quality-judge corpus with failure classes},
  year         = {2026},
  howpublished = {\url{https://catqualia.com/isomorphism/}},
  note         = {96,330 graded rows; 16,257 carrying an explicit failure class}
}