From 22cfcbbcdce5fcc5634204468e2f10d88ce53efe Mon Sep 17 00:00:00 2001 From: Olaoluwa Osuntokun Date: Fri, 24 Jul 2026 21:17:47 -0700 Subject: [PATCH] simulation/lab: open exp-010 with the splitting-pressure baseline In this commit, we record the pre-evolution baseline on the corridors corpus, and it already contains a finding: this is the first environment where lnd's production stack beats members of the evolved lineage. Its divide-and-conquer MPP completes 96% of the held-out payments -- second-best objective despite burning 23 attempts per payment -- while the naive seed manages 75% and even gen2 trails it. mx_c3's evidence-derived shard ladder still leads at 0.876, but at ten attempts per payment there is clear headroom for a router that plans its route set up front instead of discovering shard sizes by failure. The code_split1 run now chases exactly that, with the reflection prompt naming joint route-set planning as the unexplored axis. --- .../experiments/exp-010-splitting-pressure.md | 66 +++++++++++++++++++ 1 file changed, 66 insertions(+) create mode 100644 simulation/lab/experiments/exp-010-splitting-pressure.md diff --git a/simulation/lab/experiments/exp-010-splitting-pressure.md b/simulation/lab/experiments/exp-010-splitting-pressure.md new file mode 100644 index 000000000..86e2c62d6 --- /dev/null +++ b/simulation/lab/experiments/exp-010-splitting-pressure.md @@ -0,0 +1,66 @@ +# EXP-010 — Splitting pressure: does joint route-set planning emerge? + +**Date:** 2026-07-25 (started) +**Status:** in flight — baseline done, evolution run `code_split1` live + +## Question +Every winner so far splits reactively: try an amount, and when it +fails, carve the next shard from a ladder of halves and +evidence-derived sizes. Nobody has evolved joint route-set planning — +choosing a set of routes AND their shard amounts together, +min-cost-flow style. Does it emerge when the environment makes +deliberate unequal splits the difference between success and failure? + +## Environment +`corpus-split` (seed 4041), built on the new corridors topology +(commit 11f4ccc65): K = 8–16 parallel corridors of deliberately +unequal capacity tiers (one fat corridor, then rungs each at most half +its size) between one source and one target, with the tier enforced +structurally by the target-inbound channel capacity — the fattest tier +is a hard ceiling on any single shard, the tier sum a hard ceiling on +the payment. Bimodal liquidity, no drift (one variable at a time). +Each file: two cheap probes that seed corridor knowledge, then one +ambitious payment above the fattest tier. A forced max_parts=1 control +fails 40/40 files: splitting is mandatory by construction, and the +uneven ladder makes the right split unequal — halving an above-tier +payment yields shards only the fat corridor can carry. + +## Baseline (before evolution) + +| router | split-val obj | split-test obj | test succ | test att | +|---|---|---|---|---| +| lnd stack | 0.782 | 0.837 | 0.958 | 23.4 | +| seed | 0.594 | 0.644 | 0.750 | 20.1 | +| hb1 | 0.814 | 0.814 | 0.917 | 12.1 | +| **mx_c3** | **0.835** | **0.876** | 0.958 | 10.2 | +| gen2 | 0.801 | 0.770 | 0.875 | 10.7 | +| drift1 | 0.826 | 0.829 | 0.917 | 9.3 | + +Findings before evolution starts: +- **This corpus reverses the usual ordering for lnd.** Its production + divide-and-conquer MPP is genuinely good at completing these + payments (0.958 success on test, second-best objective) — it just + pays 23.4 attempts/payment for it. On corpus-building verification + it beat the naive seed outright (0.79 vs 0.67 mean objective), the + first environment where lnd tops any evolved-lineage member. +- **mx_c3's halving-plus leads**, consistent with its + evidence-derived shard ladder, but at ~10 attempts/payment there is + clear headroom: an efficient joint planner should complete these + payments in roughly half the attempts (the probes reveal corridor + tiers; sizing shards to tiers up front should rarely miss). +- The seed's 0.594/0.644 shows the gradient the run gets to climb. + +## Evolution run +`code_split1`: pure gepa, codex/gpt-5.6-sol reflection, small seed + +insights prompt, 400 evals, corpus-split. The prompt names joint +route-set planning as unexplored design space (commit 12276e6cf) and +carries the exp-008 lesson so budget is not wasted rediscovering +decay. Success criterion: beat mx_c3 on held-out split-test, then +check the winner's structure — does it plan a route SET up front +(min-cost-flow shape) or refine the reactive ladder further? Either +answer is informative; a win by ladder refinement would suggest +sequential adaptivity beats up-front planning even under maximal +splitting pressure. + +## Verdict +(pending run completion)