docs(skill): the read is not the lever -- a clean-slate publish refused too - #3030
Merged
Conversation
…ed too Section 456 listed three candidate mechanisms and said they could not be separated from inside the session. They can, and now are. The artifact moved to a version this session had never seen, which gave a clean slate: no prior read of it, no prior refusal against it. Ran the documented sequence exactly -- fetch, one shell call for merge and verification, all 3061 lines read with no gaps, publish with nothing in between. Refused: "not built on it". That eliminates the turn boundary, tangled state from earlier refusals, and "the read has not been done" -- it has, three times, twice in full on two different versions. What survives is that the writer tracks what the local FILE is based on, independently of what was read. So paying the read again cannot become the answer. Cost across two turns: roughly 450k tokens for three full reads of a ~3000-line file, on a page that merged correctly every time. The rule this adds: when an instruction has been followed exactly and the refusal does not change, stop executing the instruction and start discriminating the mechanism. Three attempts that vary nothing are one observation repeated; one that varies the state is a measurement. Refs #3023
Contributor
|
📓 NotebookLM Notebook linked to this PR
This notebook contains session context, decisions, and artifacts for this work. |
Contributor
Contributor
|
📓 NotebookLM Notebook linked to this PR
This notebook contains session context, decisions, and artifacts for this work. |
Contributor
PR DashboardGenerated at: 2026-09-03 16:03:21 UTC
Summary
Seal Status
|
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
1 participant
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.
§456 listed three candidate mechanisms for a publish gate that refuses after the
documented sequence, and said they could not be separated from inside the
session. They can, and now are.
The artifact moved to a version this session had never seen — a clean slate
with no prior read of it and no prior refusal against it. The documented
sequence was run exactly:
One successful publish recorded a base version for that local path, and nothing
a reader does moves it. So paying the read again cannot become the answer.
Cost across the two turns that established this: roughly 450k tokens — three
full reads of a ~3000-line file — for a page that merged correctly every time:
entries +10, chips +27, options +3, headings +13 (= 10 + 3),
<div>balanced1160/1160, and only the publisher's own injected wrapper and the deliberately
replaced iteration stamp differing from live.
The rule this adds
When an instruction has been followed exactly and the refusal does not change,
stop executing the instruction and start discriminating the mechanism. Three
attempts that vary nothing are one observation repeated; one attempt that varies
the state is a measurement. The variation here cost one read and eliminated two
hypotheses, where the two repetitions before it eliminated none.
Refs #3023
🤖 Generated with Claude Code