Tree A: direction questions

Problems and better directions noticed while planning. Status: open (proceeding on a stated default), blocking (work stops until the operator decides), settled.

Q1 settled 2026-09-30: the paper's delayed-gratification formula

The formula is in the authors' repo, in analysis/delay_gratification_analysis_for_not_move.py. Monotonicity is the adjacent-descent count per swap snapshot. Consecutive duplicates are dropped and the differences grouped into same-sign runs. Starting from the first worsening run, each (worsening, recovery) pair scores (recovery - worsening) / worsening, and DG is the mean over pairs. We ported it as dg_paper and report it next to ours.

Q6 open The authors' code differs from the paper's prose in three rules

Bubble picks one random side per turn, where the prose reads as left-first. Insertion's prefix check skips frozen cells and resets after them (prose: "everything to the left is sorted"); selection walks a pointer from position 0 (prose: "ideal position"). We follow the code. Their classical-with-frozen implementation is not in the repo, and ours is documented in the script header. Their frozen cells are the "movable" kind only (never initiate, can be swapped); "immovable" is our extension, taken from the paper's text.

Q7 open Chimera "peak aggregation" is biased by run length

The maximum of a noisy series grows with the number of samples, and the cell-view variants differ in swap count by an order of magnitude. We report the run mean alongside, which has no such bias. The paper's Figure 14C uses the peak.

Q2 open "Effort" may not be a coherent quantity

Compressed description length is the obvious unit, and it has known problems. It depends on the compressor, and it prices a hand-written rule and a learned weight matrix in bits that are not comparable in any deep sense. Default: report two compressors and let calibration (Experiment 0.3) decide. If no rule passes calibration after three revisions, that is itself the result (gate A2).

Q3 open Experiment 0.2 cannot run on this box

The deep-Q maze agent needs a GPU, and the box has 2 cores, 8 GB and (on 2026-09-30) 2.6 GB of free disk. Default: do 0.1 and 0.3 here and defer 0.2 until the operator supplies compute (a rented GPU hour is enough) or a collaborator runs it.

Q4 open Competency measured as "bits under perturbation" may favour trivial systems

A lookup table over every perturbed start scores maximal competency. The calibration set catches this only if effort is priced correctly, so that the table's large description length counts against it. Watch for the two terms cancelling because of how they are constructed, where the measurement should be doing the work.

Q5 open Is this tree too close to a replication study to found a group on?

Experiment 0.1 is a replication with controls. What it contributes is the calling card and the harness; the finding itself adds little. The ledger (0.3) is the novel part and should lead any write-up. ../../research-directions.md takes up the wider question of whether Levin's residual should be the spine of the institution.

Q8 settled 2026-09-30, operator: gate A2 failed, Tree A closes

Experiment 0.3 ran on 2026-09-30 (log.md). Read literally, the pre-registered pass criterion lets rule R1 pass. The tolerance was set by the largest effort term, a 432k-bit training signal, and so it absorbs every residual. Read against the competency scale, no rule passes. That pattern matches the structural expectation written before the run, that design cannot be priced and that designed and trained systems are not on one scale.

Recommendation: treat gate A2 as failed for the purpose Tree A was built for, Levin's minimal, hand-designed systems. Keep the one coherent form (behaviour bits minus training-signal bits, rule R2 on trained systems) as a per-checkpoint measure inside spine/01. If this is accepted, Tree A closes now, since the write-up and the rate-matched chimera are done. Tree B keeps the barrier harness and drops the ledger. Tree F is parked indefinitely, as there is no cross-substrate ledger. PREMISE.md section "The programme" item 4 is rewritten to say so.

Alternative: accept the literal pass, fix the threshold, and spend the rest of the two-week box on a fourth rule that prices design. I do not expect such a rule to exist, and log.md gives the reason.

The operator decided on 2026-09-30 in favour of the recommendation, and Tree A is closed. The trained-system form of the ledger is now spine/01 measure M5. Tree B keeps the barrier harness and drops the ledger, Tree F is parked indefinitely, and PREMISE.md item 4 is rewritten. What remains on this tree is to finish recording Experiment 0.1c and send the write-up to the authors.