Skip to content

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

1 Commit
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

typed-refusal-harness

A proposer → filter → typed-rejection loop for text corpora, described in the vocabulary of Marom, Tibbits, Zardini and Buehler, arXiv:2604.26367, with the measurements archived and the places the vocabulary does not fit named explicitly.

What is here

File What it is
LEXICON.md My terms expressed in the reference paper's vocabulary. False friends first — the places where borrowing their words would import a guarantee I do not have.
data/arms.json Raw per-arm counts from a pre-registered five-arm experiment, with the corpus hash, the protocol, and the known limits.
data/stats.py Recomputes every statistic this repository publishes. Standard library only, no arguments.
SORT.md An audit record: 21 claims a language model made about the relationship between that paper and my work, sorted into fold / adjust / decline. Fourteen were declined.
sort-items.json The structured record SORT.md is generated from.
build-sort.py Regenerates SORT.md. Asserts its counts before writing and fails the build if the record and the prose disagree.

Check it in two commands

python data/stats.py

Prints every number this repository claims, with confidence intervals, the test named at each use, and the limits that travel with them. If a figure appears anywhere here that this script does not print, treat it as unsupported.

python build-sort.py

Regenerates SORT.md from the structured record. The output should be byte-identical to the file in the repository. Change a verdict in sort-items.json and the build fails rather than publishing a stale count — that failure mode is not hypothetical, it is the third self-error recorded in SORT.md.

Why the reference vocabulary

I built this and named its parts before I knew anyone had formalised the shape. The reference paper does, precisely, in an established language. Where a published formalism already has a word, using my own word instead does not make the work more mine — it makes it harder to check. So their terms are the reference and mine are the ones that move.

This is not a claim that the two frameworks are the same. LEXICON.md opens with seven false friends and closes with four of my own terms retired, because the honest result of the comparison was mostly subtraction. The one correspondence I would defend is a shared control topology and nothing more; their verification composes by construction and mine demonstrably does not.

I have no affiliation with the authors and have not contacted them. Every sentence quoted from their paper is verbatim with a section address; every characterisation of their framework is my reading and is marked as such.

What the measurements do and do not show

A pre-registered experiment held the model fixed and added one structural layer at a time, over a hash-pinned public-domain corpus with twelve questions frozen before any run.

  • The first structural layer cut unsupported claims from 22.2 to 5.7 per 100 (Fisher exact, two-tailed, p = 0.002). The second held it there (p = 0.001).
  • The third layer is a smaller and less certain effect (p = 0.015) and does not clear p < 0.01.
  • No instrumented arm reached the off-model control, which produced zero unsupported claims with no structure at all. Structure narrowed the gap to a larger unstructured model; it did not close it.
  • Three of six pre-registered predictions were falsified, and one of those was falsified on a difference later shown to be indistinguishable from noise. Both facts are recorded. No significance criterion was pre-registered and none has been added retroactively.

Counts are single digit per arm, intervals are wide, and the claim denominator was produced by hand. data/arms.json lists five limits that travel with every number above.

Status and what is not here

The harness and coverage-witness tools that produced these runs are not published in this repository, so statements about their behaviour are attributed rather than independently verifiable here. The reasoning in SORT.md stands on its own without them.

This is a worked instance and an audit record, not a library. Nothing here is a package, and there is nothing to install.

Corrections

I would rather be corrected in public than be right in private. Open an issue. The declines in SORT.md rest on my reading of physics and category theory, and each one is stated plainly enough to be wrong out loud.

SORT.md was corrected twice before first publication — six errors in the first pass, four in the second, including one I introduced while fixing the first six. That history is in the document rather than removed from it.

License

Documents under CC BY 4.0. Code under Apache 2.0, matching the reference repository. Quotations from the paper are fair use; the paper itself is CC BY-NC-ND 4.0.

About

A proposer → filter → typed-rejection loop for text corpora, described in the vocabulary of Marom et al. (arXiv:2604.26367). Archived measurements and the audit record. Open-Source Collaboration

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages