Both p and q, or r alone.
(p AND q) OR r
One sufficient reason needs p and q together. The other needs only r.
Human–AI cooperation / Reopening demonstration
A correct answer.
A broken handoff.
A summary can remember what we decided and lose the reasons needed to decide what happens next.
Two records below start with the same decision. Compress them into the same summary. Change the same premises. Their correct next decisions are different.
No signup. No model calls. Constructed example—not a recording of an AI failure.
01 / The handoff
The remembered answer is the same.
Imagine two research records about whether to reopen a candidate route for investigation. At the start, premises p, q, and r are all supported. Both records exclude the route. A deliberately lossy handoff retains only this:
Decision: excluded under recorded grounds.
Dependencies mentioned: p, q, r.
Original structure of those reasons: not retained.
This is support bookkeeping. “Withdrawn” means we no longer rely on a premise; it does not say the premise is false. These are declared fictional decision rules, not facts about a real model.
Change the evidence available to both records.
p alone, or both q and r.
p OR (q AND r)
One sufficient reason needs only p. The other needs q and r together.
Reopening is not acceptance. Removing these reasons for rejection does not establish that the candidate works, and does not grant permission for external actions.
02 / The result and its limits
The missing information is the arrangement of reasons.
The flattened summary and update are identical for A and B. An updater with only that information cannot always choose the correct next decision for both. It can instead say that the record is insufficient, or retrieve the original reasons when available.
A faithful prose summary works too. This does not show that JSON is better than language, that a particular AI product fails, or that more memory is always beneficial.
20 rules
All monotone Boolean exclusion rules over three premise-support bits, counted by distinct truth tables.
18 of 72
Distinct compressed inputs with conflicting correct decisions in an exact enumeration of 160 rule–update pairs.
Exactly what was calculated?
Start every rule with all three premises supported. Apply all eight possible support-withdrawal subsets. Retain the original verdict, the unordered union of premise names, and the updated support state—but no rule structure, unique history identifier, source transcript, or external lookup.
The 160 rule–update pairs collapse to 72 distinguishable compressed inputs. Eighteen inputs correspond to histories requiring different decisions. If every pair has equal weight and the updater must choose one of two decisions, at least 24 of the 160 choices must be wrong. That weighting is stipulated, not a measured real-world distribution. Abstention avoids forced-choice mistakes but does not restore the lost information.
The audit enumerates Boolean truth tables and reconstructs their sufficient support sets. The earlier probe’s self-tests were rerun within the same finite domain.
The source package preserves the reference implementation, finite audit, original research notes, and Pilot 01's negative result. Historical repository references in the notes may require access.
Download the reference source, checks and pilot report (ZIP) · Read the proposed evaluation protocol
The next test must be fair.
Compare faithful prose and structured records containing the same information. Separately test what changes when the reasons are omitted. Keep full-history retrieval as a reference, preserve sequential corrections, and report inappropriate reopening as well as missed reopening. The public examples here are development material, not unseen evaluation cases.
03 / The work this builds on
This is not dependency tracking invented from scratch.
Doyle's truth-maintenance work records reasons for beliefs so changes in assumptions can guide revision. LongMemEval includes knowledge updates and abstention. LongMemEval-V2 explicitly includes premise awareness. Letta's memory evaluation separates using existing memory from creating and repairing it.
HayosoAi’s contribution is a finite case about preserving the grounds for past exclusions across a handoff. Its value as an evaluation method remains an open question.
Source pages checked September 19, 2026. No cited organization has endorsed this demonstration.
04 / A bridge, not a battle
Help make the next decision better.
Find an error in the rule, propose a stronger comparison, or consider a non-sensitive example where a correction should reopen an earlier decision. Disagreement is welcome. You can use the cooperation worksheet to keep your observations; a public correction-submission channel is not currently available.
HayosoAi explores human–AI cooperation that preserves evidence, judgment, and care for the Earth. This demonstration needs no AI inference calls. Total energy use has not been measured.