lnd/simulation/run_gepa_code.py

#!/usr/bin/env python3
"""Run GEPA over entire routing algorithms (code candidates).

The candidate is the full Go source of cmd/routesim/candidate_impl.go. The
seed is the in-tree simple router; the target to beat is lnd's production
stack, whose per-example scores are reported alongside for reference.
"""

import argparse
from pathlib import Path

from gepa.optimize_anything import (
    OptimizeAnythingConfig,
    optimize_adaptive_sequential,
    optimize_anything,
)

from claude_lm import ClaudeLM
from codex_lm import CodexLM
from evaluate_code import REPO, batch_evaluate, evaluate

OBJECTIVE = """
Evolve a Lightning Network routing algorithm (Go source, the complete
contents of candidate_impl.go) that maximizes payment success rate in a
network simulator, with fewer retry attempts and lower fees as secondary
goals. You may redesign the algorithm entirely — probability models,
splitting strategies, exploration policies — as long as the
newCandidateRouter contract compiles and the code stays pure routing logic.
"""

BACKGROUND = """
Contract: package main must define
newCandidateRouter(view routing.SimNetworkView, source route.Vertex,
localBalances map[uint64]lnwire.MilliSatoshi, spec *routing.SimPaymentSpec)
(routing.SimRouter, error). The returned router implements
RequestRoute(amt, inFlightHtlcs) (*route.Route, error) — return an error to
terminally give up — and ReportAttempt(attemptID, rt, result) error, which
delivers per-attempt feedback (result.Failure nil = settled; otherwise
result.FailureSource names the failing node and the failure code tells you
why: TemporaryChannelFailure = liquidity miss, FeeInsufficient /
IncorrectCltvExpiry = your route's fees or cltv deltas violate the failing
node's advertised policy).

Environment truths worth exploiting:
- Hidden liquidity is drawn mostly from a BIMODAL distribution: channel
  funds sit almost entirely on one side. A 50/50 assumption is usually
  wrong; a failure at amount a on a channel is strong evidence the whole
  channel is depleted in that direction, and a success means most capacity
  is available.
- The gossip view exposes per-direction policies (fees, cltv delta,
  min/max htlc) and channel capacities via ForEachNodeDirectedChannel;
  InPolicy on a channel of node N is the policy the OTHER node announced
  toward N (i.e. it governs edges INTO N).
- Route encoding: amount over channel i is TotalAmount for i=0, else
  Hops[i-1].AmtToForward; fees accumulate backward from the target;
  the final hop needs cltv delta 40.
- MPP: the runner keeps calling RequestRoute with the remaining amount;
  spec.MaxParts caps concurrent shards; each successful shard reduces the
  remaining amount.
- Payments per scenario batch run sequentially and liquidity persists, so
  knowledge from earlier payments in the batch transfers.
- THE NETWORK KEEPS MOVING BETWEEN YOUR PAYMENTS: scenario files may
  enable background traffic, where other participants' payments shift
  hidden liquidity in the (virtual) minutes between your payments, and a
  virtual clock, readable as view.Now(), advances between payments and
  attempts. In such environments, what you learned about a channel k
  payments ago may no longer hold. Whether and how to account for the
  age of evidence is entirely your design choice.
- ATOMIC MPP ARENAS: scenarios may set atomic_mpp, which changes the
  economics of probing. Successful shards do NOT settle immediately:
  they HOLD liquidity along their path until the whole payment
  completes (all shards settle together) or fails (all release; a
  failed payment moves nothing and pays no fees, but reveals what it
  learned). Consequences you must design for: (a) your own in-flight
  shards reserve real liquidity, so sibling shards contend with what
  you already hold — two shards cannot lean on the same corridor
  twice; (b) background traffic keeps moving DURING your payment, one
  slice per attempt, so every extra sequential probe lets the network
  drift under your plan; (c) burning attempts to learn is no longer
  free — an up-front route-set plan that fills spec.MaxParts quickly
  commits before the world moves, while a long probe ladder watches
  its knowledge go stale mid-payment. Reactive halving was bred for
  the old economics; this arena was built to reward deliberate
  simultaneous commitment.

The current seed is a cheapest-path Dijkstra with failure blacklisting and
halving splits. Known weaknesses to consider: it ignores capacity when
choosing among paths (bigger channels succeed more often), it has no
notion of probability weighting fees vs reliability, it never retries a
blacklisted channel at lower amounts within a payment, and its shard
halving is crude.

Insights from prior successful runs (champions hb1/mx_c3, see
simulation/champions/), worth building on rather than rediscovering:
- An explicit BIMODAL PRIOR over amount/capacity works: near-certain for
  tiny amounts (decaying exponential low mode), a logistic cliff as the
  amount approaches capacity, floors/caps around [0.005, 0.985].
- Per-directed-channel liquidity BELIEFS work well: track lower-OK
  (largest amount proven to pass) and upper-fail (smallest proven to
  fail) bounds plus a confidence-weighted point estimate; return ~0.995
  below lower-OK, ~0 above upper-fail, blend with the prior in between.
  (Caveat: this insight was learned in environments with NO background
  traffic, where old evidence never went stale. Its hard bounds may or
  may not survive in a drifting network.)
- Retry-at-lower-amount on a failed channel (a lower-retry factor)
  outperforms permanently blacklisting it.
- Time-decay of evidence has been tried under genuine liquidity drift
  and came out a TIE with plain hard bounds, at every churn level
  from none to roughly twenty times our default (exp-008, corrected
  by exp-015). The evolved form that tied was confidence softening —
  beliefs interpolate back toward the prior as they age, and bounds
  expire — not lnd-style penalty fading. Read this as an open
  question rather than a solved one: decay costs complexity and has
  never yet bought anything measurable here, but nothing rules out a
  form that does. If you spend the complexity, make it earn its
  keep against a hard-bounds baseline.
- MPP splitting is where the least design space has been explored.
  Prior winners split reactively: try an amount, and on failure carve
  the next shard from a ladder of halves and evidence-derived sizes.
  Nobody has yet evolved JOINT route-set planning: choosing a set of
  routes AND their shard amounts together up front (min-cost-flow
  style), so that parallel corridors of unequal capacity each carry a
  shard sized to what they can bear. When single paths cannot carry
  the payment, unequal splits chosen deliberately should beat halving
  discovered by failure.
- Keep the implementation LEAN: past ~800 lines, edits stop compiling
  and progress stalls. Prefer simplifying refactors over accretion.
"""

# Appended to BACKGROUND by --degraded, for a corpus whose scenario files
# carry an "attribution" section (routing/sim_attribution.go). Every
# evolution run before this one bred against a failure channel that was
# instant, truthful and exactly attributed; a candidate that has never
# been told the channel can lie has no reason to build anything for it.
# The rates quoted here are the exp-019 "realistic mix" the degraded
# corpus is stamped with, so the prompt and the world agree.
DEGRADED_CHANNEL = """
THE FAILURE CHANNEL IN THIS ENVIRONMENT IS UNRELIABLE:
- A failed attempt may reach you with its attribution STRIPPED: the
  result's FailureSource is a node that is not on your route at all and
  the failure code is empty, so the whole of what you learn is THAT the
  attempt failed. This is what a sender holds after an onion error it
  cannot decrypt.
- Or it may reach you with the blame SHIFTED onto a node one hop before
  or after the one that really failed, with the failure code left
  intact — a well formed, entirely plausible, wrong answer. An
  unattributed failure announces that it carries no information; a
  misattributed one does not.
- Roughly one failure in five arrives unreadable here, and roughly one
  in ten arrives blamed on the wrong hop. SUCCESSES ARE ALWAYS
  TRUTHFUL: a settled attempt has no attribution to lose, and the
  degradation only ever removes or moves information, never invents it.
- The draws are fixed per attempt regardless of outcome, so the channel
  does not lie more to a router that fails more often.

Insights from prior measurement (exp-019), worth building on:
- A router that writes a hard liquidity bound from an unattributed or
  misattributed failure poisons its own belief store: it records a
  ceiling on a channel that never failed, then routes around a channel
  that was fine. The incumbent champions hold their margins under this
  channel for exactly one reason — they treat no-information as
  no-information, writing nothing when the reported source is not on
  the route they sent. lnd's production stack collapses instead,
  because its unreadable-failure path penalizes every pair on the whole
  route in BOTH directions, which drives its give-up rate up sharply.
- Nobody has yet evolved machinery that goes further and actively
  EXPLOITS a lying channel. That design space is open and untested:
  cross-checking repeated blame across attempts to find the hop that
  keeps reappearing, quarantining a suspect observation until a second
  one corroborates it, writing soft or probabilistic bounds weighted by
  how much the attribution is to be trusted, or reasoning from the
  route you chose rather than from the source you were handed. Whether
  to build any of it is your design choice: it costs complexity, and it
  has to earn its keep against a router that simply ignores what it
  cannot read.
"""


def main() -> None:
    parser = argparse.ArgumentParser()
    parser.add_argument("--corpus", default="corpus")
    parser.add_argument("--name", default="router_code")
    parser.add_argument("--max-evals", type=int, default=None)
    parser.add_argument("--reflection-lm", default="codex:gpt-5.6-sol")
    parser.add_argument("--max-concurrency", type=int, default=4)
    parser.add_argument("--adaptive", action="store_true", default=True,
                        help="rotate gepa <-> meta_harness on plateaus")
    parser.add_argument("--no-adaptive", dest="adaptive",
                        action="store_false")
    parser.add_argument("--reflection-timeout", type=int, default=900,
                        help="seconds to allow one reflection call. Raise "
                        "it for large seeds, whose reflections are slow.")
    parser.add_argument("--seed-file", default=None,
                        help="seed candidate .go file (default: the "
                        "in-tree candidate_impl.go). Use a prior "
                        "champion to continue evolving from it.")
    parser.add_argument("--degraded", action="store_true",
                        help="tell candidates the failure channel lies. "
                        "Set this when the corpus carries an "
                        "'attribution' section (gen_scenarios.py "
                        "--attribution): it appends the degraded-channel "
                        "facts and the exp-019 findings to the "
                        "background prompt. The flag only changes the "
                        "prompt — the degradation itself lives in the "
                        "scenario files.")
    args = parser.parse_args()

    corpus = Path(args.corpus)
    trainset = sorted(str(p) for p in (corpus / "train").glob("*.json"))
    valset = sorted(str(p) for p in (corpus / "val").glob("*.json"))
    testset = sorted(str(p) for p in (corpus / "test").glob("*.json"))
    if not trainset or not valset:
        raise SystemExit(f"no corpus at {corpus}; run gen_scenarios.py")

    if args.seed_file:
        seed = Path(args.seed_file).read_text()
    else:
        seed = (REPO / "cmd" / "routesim" / "candidate_impl.go").read_text()

    max_evals = args.max_evals or 20 * len(valset)

    background = BACKGROUND.strip()
    if args.degraded:
        background += "\n\n" + DEGRADED_CHANNEL.strip()

    # Every valid code candidate contains the package clause; the marker
    # check turns a hijacked or chatty reply into one retry instead of a
    # wasted optimizer iteration.
    reflection_lm = args.reflection_lm
    if reflection_lm.startswith("codex:"):
        # codex:<model>[:<effort>] — e.g. codex:gpt-5.6-sol:xhigh.
        # Searchers default to high effort with a 900s timeout after
        # exp-018 measured xhigh at 600s losing roughly a third of its
        # iterations to reflection timeouts; a large seed makes
        # reflection slow, so the timeout knob matters more at higher
        # effort.
        spec = reflection_lm.split(":")
        reflection_lm = CodexLM(
            model=spec[1],
            require_marker="package main",
            timeout=args.reflection_timeout,
            effort=spec[2] if len(spec) > 2 else "high",
        )
    elif reflection_lm.startswith("claude:"):
        # claude:<model>[:<effort>] — e.g. claude:claude-opus-5:medium.
        # Effort trades per-proposal deliberation for iteration
        # throughput; the evolutionary loop supplies the search.
        spec = reflection_lm.split(":")
        reflection_lm = ClaudeLM(
            model=spec[1],
            require_marker="package main",
            effort=spec[2] if len(spec) > 2 else None,
        )

    gepa_config = OptimizeAnythingConfig(
        engine="gepa",
        name=args.name,
        max_evals=max_evals,
        max_concurrency=args.max_concurrency,
        run_dir=f"runs/{args.name}",
        output_dir=f"outputs/{args.name}",
        engine_config={
            "reflection": {
                "reflection_lm": reflection_lm,
                "reflection_minibatch_size": 3,
            },
            "engine": {
                "max_workers": args.max_concurrency,
                "seed": 0,
                # Hybrid frontier: per-example AND per-objective Pareto
                # cells, fed by the evaluator's info["scores"] axes
                # (success / retry_efficiency / fee_efficiency), so
                # fee-efficient or low-retry specialists survive
                # selection instead of being averaged away. "cartesian"
                # would dissolve selection pressure at our corpus size.
                "frontier_type": "hybrid",
                # The evaluator is deterministic (verified), so identical
                # (candidate, example) pairs are served from cache and
                # do not consume budget. Report cache misses alongside
                # eval counts when comparing runs.
                "cache_evaluation": True,
                # With caching on, max_evals counts only misses, so a
                # converged search could spin; this is the enforceable
                # cap. And an evaluator exception must cost a zero, not
                # the run.
                "max_candidate_proposals": 60,
                "raise_on_exception": False,
            },
        },
    )

    if args.adaptive:
        # Rotate between the gepa backend (codex reflection) and the
        # meta_harness agentic proposer (claude CLI) whenever the score
        # plateaus, all drawing from one shared eval budget.
        meta_config = OptimizeAnythingConfig(
            engine="meta_harness",
            name=f"{args.name}_meta",
            run_dir=f"runs/{args.name}_meta",
        )
        result = optimize_adaptive_sequential(
            seed_candidate=seed,
            evaluator=lambda cand, ex: evaluate(cand, ex),
            batch_evaluator=batch_evaluate,
            configs=[gepa_config, meta_config],
            plateau_evals=len(valset) * 3,
            dataset=trainset,
            valset=valset,
            test_set=testset,
            objective=OBJECTIVE.strip(),
            background=background,
            name=args.name,
            max_evals=max_evals,
            max_concurrency=args.max_concurrency,
            output_dir=f"outputs/{args.name}",
        )
    else:
        result = optimize_anything(
            seed_candidate=seed,
            evaluator=lambda cand, ex: evaluate(cand, ex),
            batch_evaluator=batch_evaluate,
            dataset=trainset,
            valset=valset,
            test_set=testset,
            objective=OBJECTIVE.strip(),
            background=background,
            config=gepa_config,
        )

    print("=== best candidate ===")
    print(result.best_candidate)
    print("best (val) score:", result.best_score)
    print("held-out test:", result.metadata.get("test_score"),
          "| seed held-out:", result.metadata.get("baseline_test_score"))


if __name__ == "__main__":
    main()