sessions: checkpoint after exp-017

.sessions/active/routing-evolution-019f95d9-7ff7-709b-8847-ec904f72acdd.md

@@ -4,8 +4,8 @@ shortname: routing-evolution
 status: active
 task_ids: []
 created_at: 2026-07-24T20:38:07Z
-updated_at: 2026-07-26T21:47:04Z
-compaction_count: 2
+updated_at: 2026-07-27T06:48:53Z
+compaction_count: 3
 git_branch: gepa
 git_last_commit: 127c34067
 ---

.sessions/journal/019f95d9-7ff7-709b-8847-ec904f72acdd/progress.md

@@ -1,11 +1,11 @@
 ---
 session_id: 019f95d9-7ff7-709b-8847-ec904f72acdd
 shortname: routing-evolution
-last_updated: 2026-07-26T21:55:00Z
-compaction_count: 3
-progress_pct: 100
+last_updated: 2026-07-27T00:00:00Z
+compaction_count: 4
+progress_pct: 0
 current_step: 1
-total_steps: 3
+total_steps: 4
 ---
 
 # Quick Resume: routing-evolution
@@ -13,46 +13,51 @@ total_steps: 3
 ## TL;DR
 Execution-state session for the GEPA routing-evolution program; the
 science lives in `simulation/lab/` and orientation in root CLAUDE.md.
-Everything is committed and pushed; the only thing in flight is
-exp-013 (`code_hybrid1`), a continuation evolution seeded from
-atomic1. Monitor it to verdict, then validate.
+Current phase: exp-017 (liquidity generator-family robustness sweep,
+the de-circularization) with three Opus 5 background agents building
+the pieces; omni adjudication (exp-018) prepped in parallel.
 
 ## Checklist
-1. [x] exp-013 `code_hybrid1` completed (352 evals, 30 iterations)
-2. [x] Exploit-grep clean, overlay-compiled, six-tier paired sweep run
-3. [x] Verdict NEGATIVE: below mx_c3 on all six tiers and below its
-       own seed on gepa's held-out test. No champion swap. Written up
-       in exp-013-hybrid-continuation.md, pushed at 8360a3829.
-4. [ ] Next: traffic-engine fix (tree is now free)   <- HERE
+1. [x] exp-017 implementation (agents A+B, committed 817804ea5..388fc426b)
+2. [x] exp-017 sweep + writeup + docs + site v58 (verdict: paradigm
+       survives 13/13; hb1 >= mx_c3 on 12/13; atomic1 flat-liquidity
+       specialist; pushed through 5de6a7cb3)
+3. [x] exp-018 prep (agent C; runner committed d6aafd52d, gepa fix
+       durable at ~/codez/gepa main 7c20d98c, both venvs refreshed)
+4. [ ] Awaiting user: launch exp-018 (locks tree) or degraded
+       attribution (needs tree) first   <- HERE
 
 ## Key Context
-- Scratch (wiped by reboot, regenerable from fixed seeds):
-  /private/tmp/claude-501/-Users-roasbeef-gocode-src-github-com-lightningnetwork-lnd-gepa/8563fa98-1f3e-4c15-8b0f-7223a827b9a2/scratchpad
-- exp-013 closed 17:22 on 2026-07-26. Its lesson: a seed already at
-  the attempt frontier makes continuation evolution walk toward
-  giving up, not toward quality.
-- TREE IS FREE — no code-mode run holds `routing/`. All four
-  pre-upstream tasks in CLAUDE.md are unblocked.
-- The exp-012 multivantage mainnet set is NOT a valid champion tier
-  (identical 0.227 success for every router). Use scen-mainnet.json.
-- `summary.json`'s `best_score` is an inflated per-minibatch metric
-  (currently 0.996 — meaningless). Champions are decided ONLY by
-  held-out paired validation.
-- Closed and written up this session: exp-010b, exp-012, exp-002b.
-  Docs reconciled through commit 201724c80; dashboard at v55.
+- HEAD at session-current start: 18cc7719a (exp-016 retraction).
+  Uncommitted: codex xhigh effort (codex_lm.py, run_gepa_code.py),
+  give_up_rate/bg_settle_rate ASI surfacing (evaluate_code.py), plus
+  whatever agents A/B produce.
+- USER DIRECTIVE: GEPA searcher/reflection agents run codex
+  gpt-5.6-sol at high/xhigh — now the default (per-call -c override
+  beats the stale harness-home config.toml).
+- USER DIRECTIVE: lead orchestrates Opus 5 agents for implementation
+  (simulator, gepa tweaks, notebooks, site); Fable advisor for
+  design checks.
+- Advisor queue: exp-017 family sweep -> distillation patch (FailAmt
+  as pathfinding constraint + amount-adaptive MPP in lnd's own
+  stack) -> exp-018 omni adjudication -> degraded attribution ->
+  offline replay on real payment data.
+- Tasks tracked in the task tool: #19 (agent A), #20 (agent B),
+  #21 (sweep, blocked on 19+20), #22 (agent C).
+- Tree is being EDITED by agents A/B — do not launch any code-mode
+  evolution run until they finish and the sweep completes.
+- Champions unchanged: hb1 + mx_c3. Dashboard at v57.
 
 ## Current Position
-- File: (monitoring, no active edits)
-- Last action: exp-013 closed negative, written up, pushed, digest mailed.
+- Last action: exp-017 closed and published (v58), digest mailed.
+  HEAD 5de6a7cb3, tree clean. Champion adjudication (hb1 vs mx_c3)
+  queued as CLAUDE.md item 3.
 
 ## Open Blockers
-- None. The four next-priority experiments all need `routing/`, so
-  they wait on exp-013 rather than being blocked in the usual sense.
+- None.
 
 ## Resume Commands
 ```bash
 git status
-ps aux | grep "[r]un_gepa"
-ls <scratch>/outputs/code_hybrid1/evals | wc -l
-grep -c "reflection unavailable" <scratch>/runs/code_hybrid1/run_log.txt
+git log --oneline -3
 ```