Minds · project notes
Experiment 03: what the forecasts showed
Twelve selected sequences ran under predictions-v1: three development checks and nine evaluation cases. There were no apparatus failures or field differences from the independent exact implementation of the same line rule. No kernel law or prediction policy changed after seeing the evaluation results.
Open the prediction viewer. Uncheck “Reveal this target's observation” to inspect the forecast before its outcome arrives. The protocol, complete result table and all input tapes are retained.
The useful result is that it can be wrong visibly
At the reversal, the past samples are 8 and 11, so the line predicts 14. The next observation is 8. Its mismatch is -6, and that old prediction stays
- Using the now-available 11 and 8, the next line predicts 5 correctly. That is a new fit from new observations, not a correction of history and not a learned decision to use a different kind of model.
The line can do better over a whole sequence while making a worse mistake at its turning point. In this case its mean absolute error is 6/7, versus 3 for hold-last, but its largest miss is 6 versus 3. Both readings matter.
Where each rule helped
Errors below use exactly the same targets with observed outcomes.
| Evaluation case | Shared targets | Line mean absolute error | Hold-last mean absolute error | What happened |
|---|---|---|---|---|
| Constant below zero | 6 | 0 | 0 | Both continue the constant; Gojo retains the flat frame's time residue. |
| Unequal-time ramp | 5 | 0 | 39/5 | The line uses elapsed source time correctly. |
| Reversal | 7 | 6/7 | 3 | One large miss, then the newly fitted line follows the fall. |
| Alternating values | 8 | 12 | 6 | The line extends the wrong direction every time. |
| Sudden step | 7 | 16/7 | 8/7 | Both miss the jump; the line also overshoots the following plateau. |
| Gaps across a change | 5 | 26/15 | 11/5 | The older line makes a large miss when measurements return. |
| Missing initial observations | 3 | 1/2 | 3 | Prediction waits for enough known samples; elapsed time includes the gaps. |
| Final blackout | 1 | 0 | 3 | Only one shared target can be scored; five later predictions remain unassessed. |
| Entirely missing stream | 0 | unavailable | unavailable | Neither method has a value to continue. |
The alternating and step sequences are clear losses for the line rule. The blackout is especially easy to misread: zero error on its one scored target does not validate any prediction made afterward. These selected examples do not establish a population-wide winner, and their sizes should not be treated as independent statistical evidence for a general claim.
Missing values remain different from wrong predictions
In the gaps case, the last known samples at times 1 and 2 are 3 and 5. The line predicts 7 and 9 for times 3 and 4; both outcomes are missing, so neither is assigned an error. At time 5 it predicts 11, but the observation is 4. That is an assessable miss of -7, with the newest training sample three time units old. The next line spans the actual source-time gap and predicts 11/3.
The runner skips open samples when choosing its two training receipts. This is an authored rule, not Gojo discovering which observations deserve attention. There is no age cutoff or confidence adjustment. The old line keeps being extended during a blackout; the records expose that limitation.
Hold-last can issue after one known observation, whereas the line needs two. The full records show each method's coverage and hold-last's error on its additional targets. A control that issues nothing has no score, rather than a perfect score. All averaging uses absolute error so opposite misses cannot cancel.
What is implemented, and what is still missing
Implemented here: a sequential experiment adapter over existing Gojo forecast structure, exact issue/arrival/assessment records, fixed selection of known receipts, independent calculations, missing-outcome accounting, a connected timeline, a before/after viewer and preserved input tapes. Source code:
- Participant adapter
- Input and outcome driver
- External reference and measurements
- Exact case definitions
- Report and journal writer
- Boundary tests
The participant prediction and signed mismatch come from Gojo's existing connected frames. External rational arithmetic is used for the independent reference and summary metrics; browser numbers draw the plot. None feeds back into prediction. Matching ordinary exact arithmetic validates this selected calculation; it establishes no numerical advantage over that arithmetic.
Work is recorded as API calls and retained receipts/issues/assessments. Those counts are not timings, allocation measurements or energy. The current implementation retains histories and repeatedly reads them; it has no bounded memory or throughput claim. The JSON files are diagnostic records, not saved worlds with an exact replay promise.
This makes several next questions concrete without requiring a fix now:
- Keep line and hold-last as distinct connected candidates, with a genuine Gojo hold-last construction, before letting past performance change their contribution. Two equal training samples already allow a flat fitted line, but that is not the same policy as copying the most recent value.
- Specify how later misses can request a closer read or change model use. Here mismatch is a retained observation, not automatically a goal, reward or material tension. Experiment 04 can test a broader observer's usefulness.
- Decide how age and missingness should affect a prediction only when a task supplies a reason. Do not turn them into an unexplained confidence number.
- Test any new selection or mixture rule on new sequences. The evaluation cases published here have now been inspected and cannot serve as untouched confirmation for changes chosen because of these results.
This is a small test of prospective expectation and its failure. It is not yet multi-source interpretation, consensus, autonomous model discovery, or evidence for the stronger claims discussed in Ingressing Minds.