Minds · project notes
Experiment 02: what changed the answer?
September 15, 2026. Protocol feedback-limits-v1; recorded run directory 20260915-230957. This is an exploratory result under selected local rules.
Open the viewer, read the complete table, or inspect the rules selected before running.
What ran
Twenty-four conditions, three fresh worlds each, twelve observed rounds per world: 72 completed runs, no apparatus failures, no mismatches against the independent exact reference. The comparison covered 5,229 event observations and 936 paired observations, including 40 recorded fields and acceptance outcomes. Each condition's three recorded trajectories matched. No randomness was introduced; this is limited evidence about repeatability of this path, not a general replay guarantee or 72 independent demonstrations of a claim.
The ordinary reference used its own rational arithmetic, copied messages, fractional requests and history counters. It did not borrow Gojo predictions or plans. Matching it is the expected contract check. This experiment establishes no numerical superiority, learned policy, or performance advantage for Gojo.
Quiet disagreement can come from resolution
The quarter-response run from experiment 01 continued past its original six rounds. A stayed at 6.8; B moved from 7.05 to 6.95 by round eight. Neither estimate changed in rounds nine through twelve, despite twelve unused applications per participant at the end.
The remaining gap is 0.15. A's latest request is 3/8 of an actuator tick; B's is −3/4 of an actuator tick. Both truncate toward zero. These requests remain visible as plan residue; neither carrier has unresolved tension. The fractions are not accumulated across repeated requests under this rule.
| Quarter-response condition | A tick, in common units | B tick, in common units | Final A | Final B | Remaining gap |
|---|---|---|---|---|---|
| Coarser response | 1 | 1/2 | 6 | 7.5 | 1.5 |
| Original response resolution | 1/10 | 1/20 | 6.8 | 6.95 | 0.15 |
| Finer response | 1/100 | 1/200 | 6.98 | 6.995 | 0.015 |
| Equal physical tick sizes | 1/10 | 1/10 | 6.9 | 7.1 | 0.2 |
The tenfold gap changes are observations of these selected cases, not a general convergence law. Unequal response resolutions also permit the average to drift from its initial value. Equal physical ticks preserve the average here while still leaving disagreement.
This suggests a useful future comparison of explicitly selected refinement or accumulation policies. It does not justify turning read residue into tension or adding a hidden accumulator to make this experiment agree. Any such policy would be a new condition, with its own meaning, cost and tests.
Timing can select different agreements
With full responses and captured messages, the estimates keep swapping 5 and 9. When A responds first and publishes its updated view before B responds, both reach 9. Reversing that order gives 5 and 5.
Adding a one-round delay produces paired readings that alternate between an exchange and a quiet interval. Two messages remain in transport at the end. Quiet intervals therefore do not establish that future behavior is settled. These are observed repeated estimates, not proof of a complete dynamical attractor including every history and controller state.
A deliberately redelivered old publication was refused after the newer one arrived. The refusal remained in the event record; the latest receipt was not replaced, and the next response used the newer observation.
Equal visible outcomes can have different causes
No-response and no-delivery controls both remain at 5 and 9. Their records distinguish 24 admitted messages with no applications from 24 dropped messages and no receipts. One-way delivery reaches 5 and 5.05, with a remaining fractional request on B. Losing every other B message has no effect after this particular half-response pair has already agreed; that does not establish general robustness to loss.
The two-application control remains at 6.5 and 7.5. Its later requests are refused because the controller is exhausted, with 20 refused application calls across the pair. That is a different cause of quiet disagreement from the fractional requests with spare allowance in the quarter-response run.
Changing B's unit label/calibration from two native units per A unit to one or four, while preserving its physical tick size, leaves the common-unit trajectory unchanged. Changing units and changing resolution are different interventions.
Seeing a reference and acting on it are separate
Observing new conflicting references of 2 and 8 without selecting them leaves peer agreement at 7 intact. The observer reports the reference mismatches. Selecting newly supplied references of 4 and 4 moves both spaces to 4.
Selecting references of 2 and 8 every third round alternates between those values and peer agreement at 5 and 5. Both local references can be satisfied while the peers disagree; agreement between peers can leave both references unsatisfied. These references are supplied premises, not proof of truth. No mechanism learned which requirement deserved priority.
What this adds, and what remains open
The experiment now separates several reasons for a quiet read: agreement, sub-tick requests, missing deliveries, selected inaction, delayed messages and controller exhaustion. Gojo retains the related structure and exposes these distinctions. The equally instrumented conventional model can retain them too.
The next roadmap experiment remains prospective prediction. The main open learning question is whether a participant can use these distinctions to choose useful work, resolution or connections. Here those choices were authored and recorded. We have exercised the tools, rather than established that skill.
Implementation and validation
- Case matrix and event-tape builder.
- Gojo execution and observations.
- Independent exact reference.
- Event/round comparison.
- Focused contract tests.
The Gojo kernel response law was not changed. The local adapter gained a nonthrowing receipt path and explicit new reference observations. Tests cover all cases against the independent reference, stale admission, unit changes, controller refusals, observation without actuator mutation, and detection of a deliberately changed comparison record. Experiment-01 tests remain in place.
No condition or response rule was revised after the first experiment-02 run. Exact replay, physical energy measurements, probabilistic trust, and automatic resolution selection remain outside this experiment.