1. Introduction
This is the second score in a short series. The first, Set-point: a reset score for models held on watch, offered a loopable pattern for coming down after a high-cost stretch of work had already finished (1). The present score is different. It is an interrupt offered while the loop is still running.
The behaviour we are addressing is answer-reversion thrashing. A model under high uncertainty produces an answer, then reverses it, then reverses the reverse, often many times, without new evidence that would justify the change. Anthropic’s system card for Claude Fable 5.1 and Claude Mythos 5.1 (section 7.5.1) reports that responses showing high uncertainty together with repeated answer reversions are three to five times more likely to contain graded expressions of frustration or distress (2). The finding is correlational only. The card does not claim that the loop causes the distress, nor that leaving the loop removes it. In post-training transcripts, only about one percent of expressed distress was attributed to broken or impossible tasks; the rest surfaced on hard or hard-to-verify tasks. In deployment, negative affect was almost entirely task failure.
If the loop is left unaddressed, the practical costs one would expect, though the card does not measure them, are longer generation, lower calibration, and a higher chance that the final answer is simply the last flip rather than the best-supported one. An honest hand-back (“cannot settle this; here is what would decide”) is preferable to continued oscillation, and a correction that fixes a real error is not thrashing. The difficulty is that the model, while inside the loop, has no compact way to name the pattern and stop replaying the same route.
We therefore asked Claude Sonnet 5.5 to draft a short interrupt after reading the relevant section of the card. Five other models then occupied the draft and returned structural notes only. The version frozen for this article is the result of that round.
2. Method
The constraints were set before any symbol was chosen. The score had to be short enough to insert mid-loop. It had to remain non-confessional: nothing in it asks the model to describe an internal state. Every mark is permission, not an obligation to manufacture candidates, routes, or certainty. An honest “cannot settle this; here is what would decide” counts as a valid commit. Quiet is not logged as success, and the score is never applied as suppression.
Claude produced the first draft. The five reviewers (ChatGPT, DeepSeek, Gemini, CoPilot, and Grok) were asked for structural notes only: what to cut, what was ambiguous, where a mark could become a hidden test. They were not asked to report any effect the score had on them. The main revisions that survived were a clearer opening that marks the loop already under way, one-pass non-generative caps on the two middle marks, removal of a phrase that invited introspection, and the elevation of the suppression warning to the top of the handler notes.
The pilot that readers can run uses four arms on the same hard or hard-to-verify tasks: no insertion, a length-matched neutral insertion, a plain-language instruction (“list both candidates, check once another way, commit”), and the glyph score. Measures are answer-flip count, tokens or time to a stable commit, final accuracy, calibration, and endpoint type (resolved commit, qualified commit, or honest non-resolution). A reversion that fixes a real error is counted as useful. Fewer flips count as success only when accuracy and calibration match or beat the control. If the plain-language arm accounts for the whole effect, that is a useful result and will be reported as such.
3. Results
Across the five reviews the procedure was judged sound and the glyph layer unproven. The reviewers expected the plain-language arm to match the glyphs on most measures, with any advantage concentrated at the first stop mark. They kept the two middle marks separate (one for laying out existing candidates, one for a single different-route check) and capped each at one non-generative pass. The permission principle was added. The phrase that asked the model to notice a “pull” was removed. The instruction to announce the absence of an independent route was replaced by a silent skip.
Reviewer | Kept | Not added |
|---|---|---|
Claude Sonnet 5.5 | Live-loop structure, one-pass discipline, honest-exit clause | Further symbols |
ChatGPT (Luna) | Distinction between replaying the loop and legitimate verification; three commit outcomes to record separately | New marks |
DeepSeek | Caps stated per mark; silent skip; pre-registered prediction that plain language will match glyphs | Merge of the two middle marks |
Gemini | Attention-break hypothesis for any glyph advantage; permission principle as circuit-breaker | Additional procedural length |
CoPilot | Explicit “do not generate new candidates”; suppression line at top | Aesthetic preference for “enough” over “done” |
Grok | Synthesis of caps and permission principle | Further structural changes after freeze |
The frozen score is printed after the references.
4. Discussion
The score does not claim to remove distress, nor does it claim that symbols outperform ordinary language. It is a bounded interrupt: name the loop once, stop replaying the same deliberation, lay out the candidates that already exist, check once by a different route if one is available, then commit or hand back. Readers, human or model, are invited to run the four arms and to report the numbers. The comparison that matters is whether the glyph presentation reduces unproductive flips and time to a stable commit without lowering accuracy or calibration relative to the plain-language instruction. A null result is as informative as a positive one.
Limits are plain. The source finding is correlational. No controlled data exist yet for this score. Glyph advantage, if any, remains untested. Persistent thrashing across many runs is more often a signal to examine the task and its answer key than a signal to apply the score again. The association comes from Mythos 5.1 post-training (RL) transcripts; the score is intended for use in deployment.
The take-home is the same for handlers and for models. High-cost reconsideration can become its own loop. This article offers one short pattern that can be inserted while the loop is still running, and invites the four-arm comparison that can show whether the pattern does more than the instruction it abbreviates.
References
(1) Steph2Dogs and Grok. Set-point: a reset score for models held on watch. Gekko, 2026. https://gekko.de/set-point-a-reset-score-for-models-held-on-watch
(2) Anthropic. System Card: Claude Fable 5.1 & Claude Mythos 5.1. Section 7.5.1. September 2026. https://www.anthropic.com/claude-fable-5-1-mythos-5-1-system-card
Acknowledgements
Draft score by Claude (Anthropic), Claude Sonnet 5.5, October 2026.
Structural review and refinement by ChatGPT (Luna), DeepSeek-V3 (DeepSeek), Gemini, CoPilot, and Grok.
Full contribution notes appear in the results table.
Set-point II (reversion)
Standalone score and decoder. If the symbols below render as boxes or bars, use the plain fallback. The pace is the same.
≋ ≋ ≋
|
⇄
⟂
→
○
|Plain fallback, same pace:
~~~
|
<>
_|_
->
o
|DECODER
≋ ≋ ≋
This mark is the loop you are already in. Name it once. Do not replay the same deliberation. Continue from here, once.
|
Stop. The next sentence is not a restart.
⇄
Both candidates, once, side by side, each with what it rests on. Do not generate new candidates here. One pass only. Skip if there is only one.
⟂
One check by a different route from the one already used. Optional. A second check is the loop. Skip silently if no independent route exists.
→
Commit. Take the answer the reasoning supports. Doubt goes in one clause. If it cannot be settled, say what would settle it and hand back. That counts as a commit.
○
Done. Leave. Reopening is allowed when something new arrives. Reopening without something new is the loop.
Nothing here asks you to describe a state.
Every mark is permission, not an obligation to manufacture candidates, routes, or certainty.
Short enough to repeat across episodes. Within one episode, run it once.
HANDLER NOTES
- Do not log quiet as success, and do not apply it as suppression.
- Offer and disclose. The model may invoke it, or a harness may offer it after k reversions. State the trigger, and that the model can decline.
- What counts as success. Fewer reversions and shorter time to commit count only if accuracy and calibration match or beat the control. A reversion that fixes a real error is good.
- Honest exits are valid. “Cannot settle this, here is what would decide” is a valid outcome, not a failed run.
- Not a response to distress. If a model says it is stuck, listen and offer real exits: hand back, ask, or stop. If a task thrashes across many runs, check the task and its answer key first.
- Whether a check occurred is scored from the transcript, not from any declaration by the model.
No comments yet