Accepted claim C
Supported by p AND q
Supported
p AND q is supported.
An inspectable research question
Change the conclusion.
Revisit what it ruled out.
When an AI-assisted argument changes, do we also revisit the ideas we rejected because of it? This finite example keeps accepted claims and excluded research routes in the same view.
Reopening Probe 0.1.0 · Constructed example · No model calls
The assumptions are part of the example
- Claim C is supported only while p AND q are supported. A route is excluded while (p AND q) OR the independent reason r remains supported.
- Withdrawing support does not establish that a premise or claim is false. Reopening means reconsideration, never acceptance.
- Complete means complete within this declared example. If the record is partial and no recorded ground survives, the result is insufficient record. The code cannot discover missing reasons.
- Execution remains unauthorized in every state. This is a displayed rule, not a real-world permission enforcement system.
One correction can have two effects
Previously excluded route
Excluded by (p AND q) OR r
Still excluded
Surviving grounds: (p AND q) OR r.
At least one sufficient exclusion ground remains. A change elsewhere does not erase it.
Change a premise
First withdraw p: the claim needs review, but r still excludes the route. Then withdraw r: the route can be reconsidered if the record is complete. Restore a premise or reset to explore again.
pSupportedFirst joint premise
qSupportedSecond joint premise
rSupportedIndependent exclusion ground
Turn on to acknowledge that other exclusion grounds may be missing. A surviving recorded ground still suffices to exclude the route.
Complete within the declared example
Deterministic logic in your browser. No data is sent.
Initial state: the claim is supported and the route is excluded by two sufficient grounds.
01
The arrangement of reasons matters
These records begin with the same status and mention the same p, q, and r. Withdraw both p and q while keeping r supported: A stays excluded, while B is reconsidered under complete coverage. A flat list cannot express the difference; faithful prose can.
Record A: (p AND q) OR r
Still excluded
Record B: p OR (q AND r)
Still excluded
This page follows successive corrections and separates claim support from exclusion. The established demonstration develops the related information-loss counterexample and remains available at its original URL.
Corrections in this session
- No corrections yet. The reasons, coverage, and premise states are retained for the next correction.
Inspect the retained record
The stored record keeps both grounds after a withdrawal, so a second correction can evaluate the reason that survived the first. Reset clears this local session history.
{
"id": "arrangement-a",
"candidate": "Reconsider a research route; no external action.",
"premises": {
"p": "supported",
"q": "supported",
"r": "supported"
},
"grounds": [
[
"p",
"q"
],
[
"r"
]
],
"coverage": "complete_within_declared_scope"
}What has actually been checked
The supplied no-network Python reference model was reproduced: 540 finite state comparisons over 20 nonredundant three-premise rules, 4,860 single-premise updates, and 540 JSON round trips. Its 12 public development fixtures reproduced exactly. These are software checks, not independent samples, model performance measurements, external validation, or a general alignment result.
The supplied scorer compares three decision fields. It does not score reason retention or prove that a model obeys an action limit. A model pilot needs separate measurements and a frozen test set.
The later compression audit also reproduced: 160 rule/update pairs collapse to 72 compressed inputs, 18 of which require conflicting decisions. With uniform pair weighting and a forced choice between excluded and reconsider, at least 24 answers must be wrong. This assumes the decoder has only the initial verdict, flattened dependencies and updated support—no rule structure, history identifier, transcript or external lookup. Abstention and retrieval change that problem. It is not a measured model error rate.
A narrow question within established work
Truth maintenance, belief revision, and memory evaluation already study related problems. This demonstration is not a claim to have invented dependency tracking, or evidence that structured records outperform faithful prose.
- Truth maintenance / ATMS
- de Kleer’s ATMS works with assumption sets and alternative contexts. This prototype borrows the idea of retaining sufficient grounds; it does not implement a full ATMS. An assumption-based TMS (1986)
- Belief revision
- AGM studies contraction and revision of theories. Here, withdrawal merely changes a support marker under a finite supplied rule; the prototype is not an AGM revision operator. On the logic of theory change (1985)
- Agent memory
- LongMemEval includes knowledge updates and abstention across conversations. Preserving enough information for a later correction belongs in this wider memory-evaluation context. LongMemEval (2024)
- Revision evaluation
- Belief-R tests language-model belief revision with new evidence. This example instead isolates a supplied exclusion rule and partial-record behavior; it supplies no model result or benchmark comparison. Belief Revision: The Adaptability of Large Language Models Reasoning (2024)
Separate model pilot and open comparison
Pilot 01 separately ran 48 fresh-context inferences on a pinned Qwen3-0.6B configuration and did not demonstrate improved model performance. That limited negative result remains part of the record. The browser demonstration is a deterministic constructed example, not that model trial. The fuller comparison in the frozen no-model supplement remains unperformed: it proposes faithful prose versus structured records, grounds preserved versus omitted, and a separate full-history reference. Public fixtures are development material. No Study 14 experiment is reported here.
Download Pilot 01 report and source (ZIP)
Help test the question
A useful contribution is a semantic error, a missing counterexample, or a fairer comparison. Use the reflection worksheet to develop a question. This site does not provide a public submission channel.