Depletion Adaptation

A previously safe food source becomes toxic through overuse. The tardigrade unlearns its positive association through the same channel that learned it — no separate forgetting mechanism required.

---

The Question

How does an animal unlearn a previously reliable food source when that source runs out — without resetting everything else it has learned?

The Answer, Short Version

The same valence-accumulation channel that wrote the positive association can write a negative one. When an over-depleted food source causes pain, the pain event slides the channel's valence negative, and the size of the slide reflects how much weight the feature was carrying. Unlearning is structurally symmetric with learning. No separate forget mechanism, no catastrophic interference, no special case.

The Setup

The test environment has plants of various colour-shape combinations. Each plant type has a depletion counter. Eating a plant increases its counter. When the counter passes a threshold, the diminished returns of that resource flip over into toxicity — continued consumption causes harm.

The protocol runs in three phases:

1. Phase 1 — Baseline learning. The tardigrade encounters fresh plants and learns which are worth eating. Positive associations form through the feature channels.
2. Phase 2 — Continued consumption. The tardigrade keeps eating its preferred food. Depletion accumulates. Eventually, the preferred food reaches toxicity and the next bite causes pain.
3. Phase 3 — Read the associations. Has the tardigrade unlearned the preferred food while preserving its knowledge of unrelated options?

What Emerged

Phase 1 (baseline, 6 plants eaten, 0 hurt):
  green_triangle: +0.91    ← the favourite
  green:          +0.74
  green_circle:   +0.55
  triangle:       +0.45
  circle:         +0.17

Phase 2 (continued feeding, depletion builds):
  green_triangle depletion: 1.02 (crosses TOXIC threshold)
  Times hurt: 1

Phase 3 (after the pain event):
  green_circle:   +0.55   (unchanged)
  green:          +0.26   (reduced)
  triangle:       +0.21   (reduced)
  circle:         +0.17   (unchanged)
  green_triangle: -0.14   ← flipped negative

Three features of the update are worth looking at.

The conjunction flipped. green_triangle swung from +0.91 to −0.14 — a total shift of about −1.05, driven by a single pain event. The system now avoids the thing it used to prefer.

Component features shifted partially. green alone moved from +0.74 to +0.26, and triangle moved from +0.45 to +0.21. These are the sub-features that participated in the pain event, and they inherited some of the negative valence — but less than the conjunction. The specificity of the association held; the generalisation to component features is real but graded.

Unrelated features stayed put. green_circle and circle did not move. The tardigrade did not overreact and avoid everything green, nor everything remotely similar. The update was surgical.

What This Proves

Unlearning requires no separate mechanism. The same valence-accumulation channel that wrote the positive association writes the negative one. A single pain event is enough to flip a conjunction that took six positive encounters to build, because the valence channel integrates both positive and negative evidence without separate bookkeeping.

The graded update across the feature hierarchy — large shift on the specific conjunction, partial shift on component features, no shift on unrelated features — is a direct consequence of how the features participated in the event. Features that were active during the pain inherit some of the negative valence proportionally. Features that were not, stay put. No explicit "what should update and by how much" rule is needed; the structure of feature participation delivers the right update shape.

This architecture avoids the failure mode conventionally called catastrophic forgetting — see conductor decoupling series for the full analysis. Here, the earlier learning of green_circle and circle is preserved because those features were not part of the update. Selectivity is structural.

What This Does Not Prove

The test uses one pain event, on one conjunction, at one depletion threshold. A fuller characterisation would explore dose-response (how many pain events does it take to flip a well-established positive?), graduated depletion response (does the system reduce consumption before toxicity, in anticipation?), and interference across multiple simultaneously-depleting resources.

Nor does this claim the tardigrade has a model of depletion as a process. Its internal state has a depletion counter per plant type; its behaviour is driven by the resulting valence. Whether this counts as "understanding" depletion is a philosophical question this experiment cannot answer. What it demonstrates is the behavioural adaptation: previously safe becomes avoided, specifically, without collateral damage.

---

Raychell Langan · NEXICOG Ltd · Hampshire, UK