SPECTER Labs
Technical Docs

Next Experiment Batch

What the Next Batch Is Trying to Answer

The current dossier already shows that the substrate can retain a prior organization. The next question is not whether that counts as “memory.” The next question is what kind of remembered organization it is.

The right Levin-style tests are about set-points, remapping, and competing commitments:

  • can the substrate hold a remembered target when guidance is removed?
  • can it overwrite one remembered target with another?
  • after damage, does it return to a prior set-point or construct a new one?
  • how much of the result belongs to the memory rule and how much belongs to the body plan?

This batch is designed to answer those questions with minimal new machinery.

Batch 1: False Memory and Competing Targets

Experiment

Add a two-pulse version of imprint:

  • pulse A drives the assembly to one target
  • a quiet interval follows
  • pulse B drives it to a different target
  • guidance is removed again

Readout

  • final com_x
  • retention after pulse A
  • retention after pulse B
  • overwrite score: distance between the final state and the second target, normalized by the first and second pulse amplitudes

Why This Matters

This distinguishes a substrate that merely gets displaced from one that stores a latest or dominant commitment. It also lets us ask whether the memory is additive, sticky, or overwritable.

Minimal Success Criterion

  • the final state should depend on pulse order
  • the final state should land closer to the second target than to the first in at least one memory-on condition

Batch 2: Remapping After Injury Without Strong Guidance

Experiment

Split the current damage scenario into two recovery modes:

  • reimpose: after damage, restore directed forcing toward the original target
  • unguided: after damage, remove or sharply reduce directed forcing and observe the spontaneous post-damage tendency

Readout

  • DRI
  • tail goal_distance
  • spontaneous drift direction after damage
  • whether the post-damage tail approaches the pre-damage attractor

Why This Matters

The current damage result shows that memory can be maladaptive. This next step asks whether the substrate is trying to return to an old set-point or whether it simply becomes stuck.

Minimal Success Criterion

  • if the substrate truly encodes a set-point, the unguided condition should show a structured drift rather than random settling
  • if it cannot remap, the reimpose condition should still recover more slowly than controls

Batch 3: Body-Plan Sweep Beyond Line vs Staggered

Experiment

Extend layout from line and staggered to a small morphology panel:

  • line
  • staggered
  • clustered
  • sparse

Keep all other rules fixed.

Readout

  • MRI
  • HLA
  • DeltaK_on_vs_off
  • DRI

Why This Matters

The staggered result already shows that morphology changes the expression of memory. The next step is to treat layout as body plan rather than nuisance variation.

Minimal Success Criterion

  • at least one morphology should preserve the imprint effect
  • at least one morphology should flip or suppress hysteresis
  • the morphology dependence should be consistent across seeds

Batch 4: Basin Depth and Commitment Strength

Experiment

For an imprinted state, apply graded perturbations of increasing strength without changing the remembered target.

Readout

  • retention after perturbation as a function of kick magnitude
  • threshold at which the prior state is lost
  • basin-depth proxy: maximum perturbation that still returns to the same tail state

Why This Matters

This turns “memory” into something measurable as an attractor landscape. If the substrate really stores commitments, those commitments should have basin depth, not just one-shot retention.

Minimal Success Criterion

  • memory-on should preserve the imprinted state under larger perturbations than memory-off

Batch 5: Parameter Family Search for Damage Recovery

Experiment

Stop searching only along “less plasticity.” The ablation results show that simply reducing plasticity did not repair damage. The next parameter family should test structure, not just magnitude:

  • asymmetry between learning and forgetting
  • damage-triggered rapid forgetting
  • capped plasticity with recovery-phase decay changes

Readout

  • DeltaK_on_vs_off in damage
  • DeltaK_on_vs_inertial_control in damage
  • DRI

Why This Matters

The current negative result suggests the memory trace is too sticky in the wrong way. The fix may require conditional remapping rather than globally weaker memory.

Minimal Success Criterion

  • one variant should make DRI positive against both controls without destroying imprint

Run these in this order:

  1. false memory / competing targets
  2. remapping after injury without strong guidance
  3. basin depth on imprinted states
  4. expanded body-plan sweep
  5. structured damage-recovery variants

This order keeps the conceptual thread clean. First ask whether there is a remembered set-point, then ask whether it can be overwritten, then ask whether it survives perturbation, then ask how morphology shapes it, and only then attempt a more targeted repair of the damage case.

What to Keep Fixed

For the next batch, keep these fixed unless a specific experiment requires otherwise:

  • same core rigid-body substrate
  • same memory=off, memory=on, and inertial_control comparison
  • same per-run provenance and NDJSON contract
  • same exact-vs-bootstrap reporting

This keeps the thread cumulative. The point is to learn what the current substrate is doing, not to open a second stack.

Exit Conditions

The next batch is worth another article revision if any of these become true:

  • we show order-sensitive overwrite of one remembered target by another
  • we show unguided drift back to a prior set-point after damage
  • we measure a nontrivial basin depth for imprinted states
  • we find a morphology that preserves both imprint and hysteresis
  • we find a damage-recovery rule that improves DRI without losing the retention result

If none of these happen, that is still informative. It would mean the current substrate stores sticky traces but not flexible set-points, which is itself a publishable negative boundary.