You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
$27,110 left to reach the funding goal.
On 2 November 2026 I published The Architecture of Failure: A Disclosure, with Logs (DOI 10.57967/hf/11368): five verbatim session logs with commercial language models, hashed and timestamped before release. Its main text names sixteen kinds of failure under sustained load and gives one specimen of each. The logs hold many more, unmarked, and they do not mark who is speaking. I am asking for the time to mark the turns, label every instance of the sixteen kinds by line, and have a second reader check a tenth of it, so that people building long-horizon evaluations can load real cases.
I write adversarial pseudocode, poetry, memoir and philosophy, but not grant proposals. This one was drafted by an AI model that knows my work, and I have carefully reviewed and corrected it. The plan and the claims are mine.
Turns marked: who is speaking, in all 3,356 paragraphs of conversation (96,012 words). The pasted texts and repeats in the logs are set aside and marked as such.
A labelling guide: a definition and a decision rule for each of the sixteen kinds, written against the existing specimens, tried on the first 200 paragraphs, then frozen.
Labels: every instance, by line, with a one-line reason.
A second reader, who labels a tenth from the guide alone. I report our agreement by category.
A datasheet stating how the record was made, what is redacted and what its limits are, and a loader.
Everything is released under CC BY 4.0 beside the record, which does not change. The work is staged, and the first stage measures my labelling rate, so the later stages are priced from a measurement.
Three checks. The second reader's agreement with my labels, reported by category, says whether someone other than me can apply the sixteen kinds. Every label opens on its line in the published log and the set loads in one line; anyone can test both. And stage A ends with a report of the rate I actually worked at, which says whether the rest of the plan is priced honestly. Whether anyone builds on the set is the test that matters most, and the one I do not control.
Stage A, $4,060 (the minimum). The labelling guide (20 hours); the first 200 paragraphs of the fifth log at ten minutes each (33.3); a report of my measured rate (5). 58 hours.
Stage B, $7,290 more ($11,350 in all). The other 823 paragraphs at five minutes each (68.6 hours); the first 200 again under the frozen guide (16.7); datasheet and release (10). 95 hours, $6,650. A second reader on a tenth, 16 hours at $40: $640.
Stage C, $15,760 more ($27,110 in all). Marking turns in Logs 1 to 4 (10 hours); their 2,333 paragraphs at five minutes each (194.4); aligning Log 4's repeats (3); release (5). 212 hours, $14,840. A second reader on a tenth, 23 hours at $40: $920.
My time is at $70 an hour. Marking the turns of the fifth log (6 hours, $420) is in a separate application to BlueDot; if that is declined, those hours are added to stage A.
I work alone. I trained in philosophy, formal logic and cognitive science, and I have run and logged adversarial sessions across several vendors since 2024. The record went out on the date I set, with a hash for every file and timestamp proofs dated the day before. One finding from it is published as a brief (DOI 10.5281/zenodo.24168730) with its limits stated. I am not a programmer; the scripts will be written with an AI coding assistant and checked by hash and by test.
The second reader and I disagree. Then the categories that fail are reported as failing and are merged or dropped. That is a finding about the taxonomy.
The work is slower than five minutes a paragraph. Stage A measures the rate. The later stages are then re-priced or cut back to what the funding covers, and I say which.
Nobody uses the set. It remains a public, dated, labelled record.
One objection this project cannot answer: every session had the same operator.
None. A separate application to BlueDot's Rapid Grants ($4,250) covers a turn-by-turn trace of the fifth log and a replay check of its finding. No hour is in both.