mirror of
https://github.com/lightningnetwork/lnd.git
synced 2026-08-18 13:07:58 +02:00
f82f4db410
The arm exp-012 could never build. A third-party node's observations injected from a file with no payment sent, so the value of routing knowledge is finally separated from the cost of acquiring it. Served the same observations from the same server, on ten sealed hard-tier files: atomic1 +0.055 (CI excludes zero, sign p=0.016), mx_c3 +0.031 with attempts nearly halved at 8.1 -> 4.4, and lnd -0.029 with attempts going UP, 30.9 -> 33.8. Free, accurate, correctly-scoped information makes lnd worse. Splitting the stream says why, and it is the program's central thesis arriving from a new direction. Successes help everyone. Failures split the field: they help the interval routers and they are the whole of lnd's loss at -0.039, CI excluding zero, worse on 9 of 10 files. An interval router files a failure as an AMOUNT BOUND and will still route half that amount tomorrow, so a served failure is pure information. lnd files it as a penalty on the pair, and a penalty is not amount-aware -- it suppresses the corridor for every amount, so a stranger's failure at a stranger's amount steers lnd off corridors that were fine for what it actually wants to send. I expected mission control's collapse of channels onto node pairs to be the culprit and it is not: 761 directed edges, 761 distinct pairs, no parallel channels, nothing collapses. The damage is in how a failure is represented, not in how it is keyed. The champions could not consume anything at all, because nothing in the SimRouter contract ever asked a candidate to accept third-party knowledge. Hence the two importer variants, each its ancestor plus one method that routes every observation through the same belief update a real attempt makes. Both score identically to their originals cold, so the only thing that changed is the capability. Also adds gen_served_weights.py, which builds the server-side scenario files. The server must be a different node than the consumer or the exercise collapses back into self-warming, which exp-012 part 4 already measured as harmful. |
||
|---|---|---|
| .. | ||
| champions | ||
| command-center | ||
| lab | ||
| .gitignore | ||
| claude_lm.py | ||
| codex_lm.py | ||
| evaluate.py | ||
| evaluate_code.py | ||
| export_run.py | ||
| gen_mainnet_scenarios.py | ||
| gen_scenarios.py | ||
| gen_served_weights.py | ||
| gen_warmup_scenarios.py | ||
| params_lnd_bimodal.json | ||
| preflight.py | ||
| README.md | ||
| refresh_dashboard.sh | ||
| run_gepa.py | ||
| run_gepa_code.py | ||
| run_gepa_omni.py | ||
| sweep_validate.py | ||
| warmup_curve.py | ||
Routing Optimization Harness
This directory holds the GEPA-based optimization harness for lnd's pathfinding. The core idea: lnd's real routing code (or a candidate replacement algorithm) runs against an in-process simulated Lightning Network with hidden liquidity, an evaluator scores the outcome, and a reflective LLM optimizer (GEPA) proposes improved candidates from the failure feedback.
Components
| Piece | Where | What |
|---|---|---|
| Simulator | routing/sim_*.go |
In-memory LN with hidden balances; real pathfinding + mission control run unmodified against it |
| CLI | cmd/routesim |
params JSON + scenario file in, attempt traces + aggregate JSON out |
| Candidate slot | cmd/routesim/candidate_impl.go |
A complete routing algorithm behind --router=candidate; swapped per candidate via go build -overlay |
| Corpus | gen_scenarios.py |
train/val/test scenario files: topology + liquidity seed + payment batch |
| Evaluators | evaluate.py, evaluate_code.py |
score = success rate − small saturating penalties for attempts and fee ppm |
| Runners | run_gepa.py, run_gepa_code.py |
parameter mode and code mode optimization |
| Reflection LM | codex_lm.py |
GEPA LM protocol via codex exec headless (default gpt-5.6-sol) |
| Lab notebook | lab/ |
running log of experiments, results, ideas |
Quick start
# Build the simulator binary.
go build -o /tmp/routesim ./cmd/routesim
# Generate a scenario corpus.
python3 simulation/gen_scenarios.py --out /tmp/corpus
# Score the lnd defaults on one example.
cd simulation && ROUTESIM_BIN=/tmp/routesim python3 evaluate.py /tmp/corpus/val/example_000.json
# Compare lnd stack vs the candidate router on a scenario file.
/tmp/routesim --scenarios /tmp/corpus/val/example_000.json --router=lnd --traces=false
/tmp/routesim --scenarios /tmp/corpus/val/example_000.json --router=candidate --traces=false
# Full optimization runs. gepa must be installed from git main — a
# durable clone lives at ~/codez/gepa; prefer uv for the env:
# uv venv /tmp/gepa-venv && uv pip install -p /tmp/gepa-venv \
# "~/codez/gepa[full]"
# Also needs the codex CLI authenticated and OPENAI_API_KEY set.
ROUTESIM_BIN=/tmp/routesim python3 run_gepa.py --corpus /tmp/corpus --name run1 --max-evals 400
ROUTESIM_BIN=/tmp/routesim python3 run_gepa_code.py --corpus /tmp/corpus --name code1
The two optimization modes
- Parameter mode (
run_gepa.py) — candidate = JSON of the existing heuristic's knobs (estimator choice, apriori/bimodal params, attempt cost, min probability). Validates the loop and tunes the current paradigm. - Code mode (
run_gepa_code.py) — candidate = the full Go source ofcandidate_impl.go, an entire routing algorithm implementing therouting.SimRouterinterface. This is the paradigm-free path: the candidate sees only gossip, its own balances, and per-attempt feedback. Compile errors are returned to the proposer as feedback.
Anti-reward-hacking measures
- Candidate routers receive a
SimNetworkViewwrapper, not the concrete graph, so hidden balances and liquidity mutation are unreachable. evaluate_code.pyrejects candidates usingunsafe,reflect,os/exec, network packages, etc.- Selection happens on a val split; a sealed test split is only used for final reporting.
- The source's own channels are rebalanced 50/50 before each batch so scores measure routing skill, not sender funding luck.
Command center
command-center/ holds a static dashboard site (serve with
python3 -m http.server from that directory).