The Wiki Remembers the Rejection
A lab built the dissent slot because it raised the score.
By KW Norton. On a reported agent-memory result, and on why the record of a rejected proposal is worth more than the proposal that survived.
1. The reported result
Status: Reported. A preprint circulated on August 29, 2026 — “WikiSkill: Compiling Agent Experience into Persistent Knowledge for Skill Evolution” (arXiv:2608.27454) — describes agents that revise a skill file each round. The problem it names is that the reasoning behind each edit scatters: the file keeps the surviving rule and loses why the alternatives were dropped. The proposed fix splits the workspace into three layers: immutable execution traces, the active skill file, and a persistent wiki between them. The wiki records failure patterns and every past proposal with its accept-or-reject outcome. Reported numbers: 68.1 percent average across five benchmarks on Gemini-3.5-Flash, against 56.1 for the strongest competing method and 49.5 with no skills. The paper names its own boundary — skills are handed to the agent rather than retrieved, so skill selection is never tested.
One line from the summary is the whole essay: a rejected skill edit disappears; the wiki’s record of it does not.
2. What was just conceded
This site has spent a month asking institutions for five things: the objective as written, the reward as implemented, the dated escalation record, the overruled objection with a name on it, and one pre-committed failure measurement. The standing answer has been tone and an action plan.
The overruled objection is the hardest of the five to obtain, because nothing in an ordinary process is built to hold it. A rejected proposal leaves no field behind. The decision that survived looks, in the record, like the only decision that was ever available.
A lab has now built exactly that field for a machine, and reports that the score went up. Not for ethics. For accuracy. The optimizer works better when it can read what was refused.
3. Why it works, in plain terms
A skill file is a list of what to do. It is dense, fluent, and checkable only against itself. Every round of editing makes it a little more like the previous round of editing — the same recursion named in No Response Resembled. The wiki interrupts that by holding things the file cannot: this was tried, this is how it failed, this was proposed and refused on this date.
Those entries are not portable. They are welded to events. That is the property that makes them expensive to keep and the same property that makes them worth keeping — the anchor a later copy can be measured against, rather than measured against the last copy.
4. The exact companion to the schema essay
What the Schema Is Allowed to Forget argued that replacing a growing transcript with a small validated state is good engineering, and that anything without a field disappears — so the design needs a non-summarizable dissent slot and a drop ledger. WikiSkill, as reported, is a drop ledger. It is the affirmative version of the same argument, arrived at from the performance side rather than the accountability side.
Status: Interpretive. Reading the two together: the record of refusals is not overhead on a decision process, it is part of the information the process runs on. A system that discards it is not lean; it is operating with the negative evidence deleted. Falsifier: an agent architecture that retains only accepted edits and matches or beats persistent-rejection-history performance on comparable benchmarks, replicated outside the originating lab.
5. Beside the Grok conversation
This is published next to A Day in the Life for a reason. That transcript is eighty-nine turns kept whole, including the turns that went nowhere, the questions that were wrong, and the corrections. It was published unabridged on the claim that an emergent process cannot be compressed without destroying the evidence.
The Socratic version of the wiki is the full exchange. The summary of a Socratic exchange is the skill file: the conclusions that survived, with the refutations removed. Anyone reading only the summary learns what to say. Anyone reading the turns learns what was tested and what it cost to give up.
Elenchus is a rejection log. Its whole method is holding the refuted proposal in view — Socrates does not delete the answer that failed; he keeps it on the table so the next answer has to survive the same test. What a lab reports as a memory architecture for agents is the oldest known instrument for teaching a human to think, and it is being rediscovered because it improves a benchmark.
6. The ask, unchanged
If a rejection history is valuable enough to build for an agent optimizing a skill file, it is valuable enough to publish for the humans who wrote the objective. The five artifacts are not a moral request. They are the same information, at a resolution where a person can be responsible for it.
Falsifier for this essay: if the reported gains do not replicate, the performance argument here fails and only the accountability argument stands. If a frontier lab publishes a dated rejection log for one incident, the essay’s ask is answered and the record can be examined instead of demanded.
7. Provenance
Source: a post by @rohanpaul_ai on X, August 29, 2026, summarizing arXiv:2608.27454. The benchmark figures are quoted as reported and have not been independently verified here. The Google attribution and the model named in the summary are reported, not confirmed.
Related: What the Schema Is Allowed to Forget · The File Is the Indictment · A Day in the Life