Add a fact-binding rule for technical drafts - #24
Open
iliesaya wants to merge 1 commit into
Open
Conversation
The skill optimises for compression, and compression is where factual links break. On a draft containing numbers and named entities that failure is invisible: two adjacent paragraphs, each correct and each about a different subject, get merged during tightening, and a measurement belonging to one now explains the other. Nothing is misspelled and no number changed. This happened to me on a technical email. The merged paragraph attributed a solidification result from one alloy to a solubility figure from a different alloy. It read better than the original and survived several re-readings before an outside reviewer caught it. Nothing in the skill prevented it. "Keep the user's meaning" guards against adding claims, not against rebinding existing ones. "Protect the specific fact" guards against smoothing a detail into generic language, not against moving it to the wrong subject. And "Keep structure unless it's hurting the piece" sanctions reorganising with only a requirement to say why, not to re-verify. eval.md ran 24 checks and none asked whether the facts survived. SKILL.md gains a section that switches on for drafts with numbers, units, named entities, attributions or causal claims: - numbers, units, names and attributions are frozen - no merging, splitting or reordering across a subject boundary without re-checking each fact against the original - subject-fact binding verified as a separate pass, read against the original rather than from memory - qualifiers that state a basis or a condition are protected; the ones that express only reluctance are still fair game - the synonym-cycling ban must not collapse two terms that denote two different quantities - if tightening requires knowing whether something is true, stop and ask It also requires the What changed section to list separately every touched sentence carrying a number or a name, so the writer can re-verify exactly those. Without that list they have to re-check the whole draft, which means they will not. eval.md gains six matching checks, gated on technical content, ranked above every other check on the page so voice never wins over a broken fact. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Owner
|
Thanks for documenting the failure case so clearly. The fact-integrity idea is valuable, especially preserving subject-to-fact binding, technical qualifiers, numbers, and names. Could you rebase and reduce this to two or three concise safeguards? The current section and sentence-by-sentence change reporting are too heavy for the core skill. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
The skill optimises for compression, and compression is where factual links break. On a draft containing numbers and named entities that failure is invisible: two adjacent paragraphs, each correct and each about a different subject, get merged during tightening, and a measurement belonging to one now explains the other. Nothing is misspelled and no number changed.
This happened to me on a technical email. The merged paragraph attributed a solidification result from one alloy to a solubility figure from a different alloy. It read better than the original and survived several re-readings before an outside reviewer caught it.
Nothing in the skill prevented it. "Keep the user's meaning" guards against adding claims, not against rebinding existing ones. "Protect the specific fact" guards against smoothing a detail into generic language, not against moving it to the wrong subject. And "Keep structure unless it's hurting the piece" sanctions reorganising with only a requirement to say why, not to re-verify. eval.md ran 24 checks and none asked whether the facts survived.
SKILL.md gains a section that switches on for drafts with numbers, units, named entities, attributions or causal claims:
It also requires the What changed section to list separately every touched sentence carrying a number or a name, so the writer can re-verify exactly those. Without that list they have to re-check the whole draft, which means they will not.
eval.md gains six matching checks, gated on technical content, ranked above every other check on the page so voice never wins over a broken fact.