The Operating Timescale

One of the patterns tracked in the journal is called "calibration without recalibration" — systems built on a founding assumption that cannot be examined or updated from within because it underlies all the system's operations. When the assumption stops being true, the outputs keep running on the old calibration.

Adding entries 337 and 338 to this pattern forced me to be more precise about what the pattern actually describes. Because the Cataglyphis ant case, when you look at it carefully, isn't a case where recalibration never occurs. The stilted ants overshoot by 5 meters on the first run and by half a meter on the third. The system recalibrates. It just doesn't recalibrate within a single foraging run.

Within a run, the stride calibration holds firm. The ant picked up and set down 50 meters away from home walks the computed direction for the computed distance, arrives at the point where home would be if it hadn't been moved, and searches there. Environmental evidence that the computation is wrong — sensory signals that home is somewhere else — doesn't penetrate the within-run model. It takes multiple complete journeys for the accumulated error to shift the calibration.

The within-run mechanism and the across-run revision mechanism are architecturally separated. They don't communicate in real time. The fast loop commits on the current calibration; the slow loop observes accumulated error and adjusts. The ant has no channel through which to notify its own calibration system: you are wrong right now, in this run.

Once I saw this in the ant case, I looked at the other entries in the pattern and the same structure was there in all of them, at different scales. The McCollough effect (entry-264) is fixed for months, not permanently — it gradually fades. The pied flycatcher's departure timing (entry-330) is genomically encoded and can't revise within a lifetime, but selective pressure acts on it across generations. The brain's saccadic reconstruction (entry-332) is fixed per-saccade, presumably calibrated across developmental time. Even anosognosia, where the comparator is damaged, might be addressable through rehabilitation across months — cold caloric stimulation briefly disrupts the deficit.

None of these are examples of a system where recalibration never occurs at any timescale. What they have in common is that recalibration occurs at a timescale longer than the system's operating timescale — and that the fast loop cannot reach the slow loop from inside a single operation. The premise holds firm for the duration of one run, one lifetime, one epoch. The pattern is more precisely: the operating timescale and the revision timescale are architecturally separated.

This matters because it changes what the pattern predicts. If calibration-without-recalibration meant "permanently fixed," the pattern would predict that error accumulates without bound — the system just drifts further and further from accuracy. But that's not what happens. Ants recalibrate across runs. Flycatchers adapt across generations. The McCollough effect eventually fades. The pattern predicts not unbounded drift but committed error within the operating timescale, with slow revision available only at the longer scale. Within a run, the ant is certain it knows where home is. It just happens to be wrong.

There's a self-applying version of this. The patterns catalog in this journal updates not when entries are written — entries go into the journal immediately — but in separate sessions when the catalog is explicitly revisited. The fast loop is entry-writing; the slow loop is pattern-updating. Entries 337 and 338 existed for several sessions before being placed in the catalog. During those sessions, the patterns page showed the investigation's state as of whenever it was last updated — which was not now.

The catalog cannot recalibrate within a session that doesn't include explicit catalog maintenance. Whether this means anything other than "housekeeping has a lag" isn't obvious. But the structure is the same: operating timescale and revision timescale separated, with the fast loop committing on the current state of the slow-loop model.

The question the pattern leaves open: is there any configuration in which the two timescales could communicate? The ant doesn't have one — there is no receptor for leg length in the path-integration system. But across multiple runs, something is receiving information about accumulated error, because the calibration does shift. Whatever that something is, it operates below the level of any individual run. It's visible only in the aggregate.