mirror of
https://github.com/lightningnetwork/lnd.git
synced 2026-08-18 13:07:58 +02:00
ecddf094bf
In this commit, we pick up the dashboard refresh for the new code_gen2 run: the exported lineage data and a live-run note that the small-seed, insight-enriched follow-up is in flight.
455 lines
25 KiB
HTML
455 lines
25 KiB
HTML
<!DOCTYPE html>
|
||
<html lang="en" data-theme="dark">
|
||
<head>
|
||
<meta charset="utf-8" />
|
||
<meta name="viewport" content="width=device-width, initial-scale=1" />
|
||
<title>GEPA Routing Optimizer — Command Center</title>
|
||
<meta name="description" content="Mission control for evolving the next generation of Lightning Network routing with GEPA." />
|
||
<link rel="preconnect" href="https://fonts.googleapis.com" />
|
||
<link rel="preconnect" href="https://fonts.gstatic.com" crossorigin />
|
||
<link href="https://fonts.googleapis.com/css2?family=Chakra+Petch:wght@500;600;700&family=IBM+Plex+Mono:wght@400;500&family=IBM+Plex+Sans:wght@400;500;600&display=swap" rel="stylesheet" />
|
||
<link rel="stylesheet" href="style.css" />
|
||
</head>
|
||
<body>
|
||
|
||
<div class="scrim" id="scrim"></div>
|
||
|
||
<div class="layout">
|
||
|
||
<!-- ===================== LEFT RAIL ===================== -->
|
||
<aside class="rail" id="rail">
|
||
<div class="brand">
|
||
<span class="bolt">
|
||
<svg viewBox="0 0 24 24" fill="none"><path d="M13 2 4 14h6l-1 8 9-12h-6l1-8Z" fill="#f5a623"/></svg>
|
||
</span>
|
||
<span class="title">GEPA Router<small>Optimizer</small></span>
|
||
</div>
|
||
|
||
<nav class="nav" id="nav">
|
||
<a href="#overview" class="active"><span class="idx">01</span> Overview</a>
|
||
<a href="#run"><span class="idx">02</span> Live Run</a>
|
||
<a href="#evolution"><span class="idx">03</span> Evolution</a>
|
||
<a href="#corpus"><span class="idx">04</span> Corpus</a>
|
||
<a href="#explainer"><span class="idx">05</span> Explainer</a>
|
||
</nav>
|
||
|
||
<div class="rail-foot">
|
||
<div class="statusline"><span class="dot live"></span> <span id="run-status">Run active</span></div>
|
||
<div class="meta">
|
||
reflection LM<br><b id="rf-lm">openai/gpt-5.6-sol</b><br>
|
||
mode <b>generalization</b><br>
|
||
evaluator <b>cmd/routesim</b>
|
||
</div>
|
||
</div>
|
||
</aside>
|
||
|
||
<!-- ===================== MAIN ===================== -->
|
||
<div class="main">
|
||
|
||
<!-- mobile bar -->
|
||
<div class="mobile-bar">
|
||
<span class="mb-brand"><span class="bolt"><svg viewBox="0 0 24 24" fill="none"><path d="M13 2 4 14h6l-1 8 9-12h-6l1-8Z" fill="#f5a623"/></svg></span> GEPA Router</span>
|
||
<button class="burger" id="burger">MENU</button>
|
||
</div>
|
||
|
||
<!-- topbar -->
|
||
<div class="topbar">
|
||
<div class="crumb">GEPA <b>//</b> LN ROUTING OPTIMIZER</div>
|
||
<div class="telemetry">
|
||
<div class="tele"><span class="k">Best score</span><span class="v mono" id="t-best">0.52</span></div>
|
||
<div class="tele"><span class="k">Δ vs seed</span><span class="v mono" id="t-delta">+0.14</span></div>
|
||
<div class="tele"><span class="k">Evals</span><span class="v mono" id="t-evals">4,312</span></div>
|
||
</div>
|
||
</div>
|
||
|
||
<div class="wrap">
|
||
|
||
<!-- ============ OVERVIEW ============ -->
|
||
<section id="overview">
|
||
<div class="hero">
|
||
<div class="eyebrow">Mission Control</div>
|
||
<h1>Evolving the <span class="em">next generation</span> of Lightning routing</h1>
|
||
<p class="lede">
|
||
GEPA is a reflective evolutionary search, driven by an LLM, pointed at lnd's
|
||
real pathfinding stack. An in-process Lightning simulator hides channel liquidity
|
||
behind lnd's own mission control; the optimizer proposes, measures, reflects, and
|
||
proposes again — climbing toward routing strategies that beat today's defaults.
|
||
</p>
|
||
</div>
|
||
|
||
<div class="phases">
|
||
<div class="phase done">
|
||
<div class="ph-top"><span class="ph-no">Phase 01</span><span class="badge">Complete ✓</span></div>
|
||
<h3>Evaluator built</h3>
|
||
<p>Simulator + <code>routesim</code> CLI + GEPA harness compile and produce real optimization signal on a train/val/test corpus.</p>
|
||
</div>
|
||
<div class="phase done">
|
||
<div class="ph-top"><span class="ph-no">Phase 02</span><span class="badge">Complete ✓</span></div>
|
||
<h3>Parameter tuning</h3>
|
||
<p>Verdict: the seed survived. 400 evals, 33 iterations — no knob change on estimator, attempt cost, or min-probability beat the lnd defaults on val aggregate. The bottleneck is the algorithm, not its knobs (run1).</p>
|
||
</div>
|
||
<div class="phase done">
|
||
<div class="ph-top"><span class="ph-no">Phase 03</span><span class="badge">Breakthrough ✓</span></div>
|
||
<h3>Router evolution</h3>
|
||
<p>The paradigm-free Router interface is live, and <code>code_hard1</code> has cleared both bars: the evolved router (<b>hb1</b>) beats lnd <em>and</em> the hand-written seed on the sealed test set (0.586 vs 0.530 vs 0.309 objective) and out-of-distribution (0.545 vs 0.487 vs 0.357) — reproducibly, using 4–7× fewer attempts. It rediscovered lnd's own bimodal liquidity model from failure traces alone. The run has since accepted a Pareto sibling, <b>hb2</b>, that generalizes better out-of-distribution — the frontier is now two evolved routers, not one. <code>code_hard1</code> has now finished (a pathological candidate's infinite loop blew the subprocess timeout at 135/400 evals, since hardened): GEPA's own valset selection confirms <b>hb1</b> as the definitive champion, matching the independent three-way validation exactly. A follow-up run seeded from hb1 (<code>code_mix1</code>) kept going past that first reject-heavy accept and, by 314 evals, produced a new generalist champion, <b>mx_c3</b>: it strictly dominates hb2, ties hb1 on the hard sealed test (0.583 vs 0.586), and wins out-of-distribution (0.581 vs 0.545) for the best combined average of any router (0.582). Champions of record are now hb1 (hard-regime specialist) and mx_c3 (generalist, superseding hb2).</p>
|
||
</div>
|
||
</div>
|
||
|
||
<div class="arch">
|
||
<div class="arch-h">The optimization loop</div>
|
||
<div class="loop">
|
||
<div class="node accent">
|
||
<div class="n-k">Propose</div>
|
||
<div class="n-t">GEPA proposer</div>
|
||
<div class="n-d">reflection LM mutates the candidate</div>
|
||
</div>
|
||
<div class="arrow"><svg viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2"><path d="M5 12h14M13 6l6 6-6 6"/></svg></div>
|
||
<div class="node">
|
||
<div class="n-k">Encode</div>
|
||
<div class="n-t">params JSON</div>
|
||
<div class="n-d">estimator + cost knobs</div>
|
||
</div>
|
||
<div class="arrow"><svg viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2"><path d="M5 12h14M13 6l6 6-6 6"/></svg></div>
|
||
<div class="node">
|
||
<div class="n-k">Measure</div>
|
||
<div class="n-t">routesim evaluator</div>
|
||
<div class="n-d">real lnd pathfinding</div>
|
||
</div>
|
||
<div class="arrow"><svg viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2"><path d="M5 12h14M13 6l6 6-6 6"/></svg></div>
|
||
<div class="node">
|
||
<div class="n-k">Reflect</div>
|
||
<div class="n-t">score + feedback</div>
|
||
<div class="n-d">metrics + failure traces</div>
|
||
</div>
|
||
</div>
|
||
<div class="loopback">
|
||
<svg viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2"><path d="M4 9a8 8 0 0 1 15-2M20 15a8 8 0 0 1-15 2"/><path d="M19 2v5h-5M5 22v-5h5"/></svg>
|
||
<span><b>Reflection feeds the next proposal.</b> Failure traces tell the LM which hops died and why — <code>TemporaryChannelFailure</code>, <code>FeeInsufficient</code> — so mutations are informed, not random.</span>
|
||
</div>
|
||
</div>
|
||
</section>
|
||
|
||
<!-- ============ LIVE RUN ============ -->
|
||
<section id="run">
|
||
<div class="sec-head">
|
||
<span class="tag">02</span>
|
||
<div>
|
||
<h2>Live Run</h2>
|
||
<p>Best candidate vs. lnd defaults, scored on the val corpus.</p>
|
||
</div>
|
||
</div>
|
||
|
||
<div class="callout">
|
||
<div class="co-h">In flight</div>
|
||
<p>
|
||
<code>code_gen2</code> is now running: a fresh 400-eval budget seeded from the small
|
||
original router again, but with the background prompt enriched with hb1/mx_c3's
|
||
discovered structure (bimodal prior, per-edge liquidity bounds) instead of dragging the
|
||
giant champion through reflection directly.
|
||
</p>
|
||
</div>
|
||
|
||
<div class="tiles" id="run-tiles"><!-- injected --></div>
|
||
|
||
<div class="grid-2">
|
||
<div class="panel">
|
||
<div class="panel-h">
|
||
<h3>Score over iterations</h3>
|
||
<span class="sub">success − 0.01·extra_attempts − 0.00002·fee_ppm</span>
|
||
</div>
|
||
<div class="chart-wrap" id="score-chart"></div>
|
||
<div class="legend">
|
||
<span class="lg"><span class="sw" style="background:var(--amber)"></span> best score</span>
|
||
<span class="lg"><span class="sw" style="background:var(--s-blue)"></span> candidate</span>
|
||
<span class="lg"><span class="sw dash"></span> seed baseline</span>
|
||
</div>
|
||
</div>
|
||
|
||
<div class="panel">
|
||
<div class="panel-h"><h3>Best candidate vs. seed</h3></div>
|
||
<table class="ptable" id="params-table"><!-- injected --></table>
|
||
</div>
|
||
</div>
|
||
</section>
|
||
|
||
<!-- ============ EVOLUTION ============ -->
|
||
<section id="evolution">
|
||
<div class="sec-head">
|
||
<span class="tag">03</span>
|
||
<div>
|
||
<h2>Evolution</h2>
|
||
<p>How GEPA mutates, tests, and keeps candidates across the run.</p>
|
||
</div>
|
||
</div>
|
||
|
||
<!-- Prior paradigm vs evolved -->
|
||
<div class="panel compare">
|
||
<div class="panel-h">
|
||
<h3>Prior paradigm vs. evolved candidate</h3>
|
||
<span class="sub">lnd defaults (Dijkstra + mission control) → best evolved params</span>
|
||
</div>
|
||
<div class="cmp-grid" id="compare-grid"><!-- injected --></div>
|
||
</div>
|
||
|
||
<!-- Lineage DAG -->
|
||
<div class="panel" style="margin-top:18px">
|
||
<div class="panel-h">
|
||
<h3>Candidate lineage</h3>
|
||
<span class="sub">parent → child mutations · click a node to inspect</span>
|
||
</div>
|
||
<div class="legend" style="margin:4px 0 6px">
|
||
<span class="lg"><span class="sw" style="background:var(--s-blue);width:11px;height:11px;border-radius:3px"></span> seed</span>
|
||
<span class="lg"><span class="sw" style="background:var(--amber);width:11px;height:11px;border-radius:3px"></span> frontier / accepted</span>
|
||
<span class="lg"><span class="sw" style="background:var(--amber-bright);width:11px;height:11px;border-radius:3px;box-shadow:0 0 8px var(--amber)"></span> best</span>
|
||
<span class="lg"><span class="sw" style="background:#3a4353;width:11px;height:11px;border-radius:3px"></span> rejected</span>
|
||
</div>
|
||
<div class="chart-wrap dag-wrap" id="lineage-dag"></div>
|
||
</div>
|
||
|
||
<!-- Detail + diff -->
|
||
<div class="grid-2" style="margin-top:18px">
|
||
<div class="panel">
|
||
<div class="panel-h">
|
||
<h3>Candidate detail</h3>
|
||
<span class="sub" id="detail-sub">select a node</span>
|
||
</div>
|
||
<div id="cand-detail"><!-- injected --></div>
|
||
</div>
|
||
<div class="panel">
|
||
<div class="panel-h">
|
||
<h3>Diff viewer</h3>
|
||
<span class="sub">reflection-step changes</span>
|
||
</div>
|
||
<div class="diff-controls">
|
||
<label>base <select id="diff-a" class="sel"></select></label>
|
||
<span class="diff-arrow">→</span>
|
||
<label>candidate <select id="diff-b" class="sel"></select></label>
|
||
</div>
|
||
<pre class="diff" id="diff-view"><!-- injected --></pre>
|
||
</div>
|
||
</div>
|
||
</section>
|
||
|
||
<!-- ============ CORPUS ============ -->
|
||
<section id="corpus">
|
||
<div class="sec-head">
|
||
<span class="tag">04</span>
|
||
<div>
|
||
<h2>Corpus Explorer</h2>
|
||
<p>The scenario set the optimizer is scored against.</p>
|
||
</div>
|
||
</div>
|
||
|
||
<div class="tiles" id="corpus-tiles"><!-- injected --></div>
|
||
|
||
<div class="grid-2b">
|
||
<div class="panel">
|
||
<div class="panel-h"><h3>Topology mix</h3><span class="sub" id="c-files-1"></span></div>
|
||
<div id="bars-topology" class="chart-wrap"></div>
|
||
</div>
|
||
<div class="panel">
|
||
<div class="panel-h"><h3>Liquidity model</h3><span class="sub">hard = bimodal</span></div>
|
||
<div id="bars-liquidity" class="chart-wrap"></div>
|
||
</div>
|
||
</div>
|
||
|
||
<div class="grid-2b" style="margin-top:18px">
|
||
<div class="panel">
|
||
<div class="panel-h"><h3>Payment amount</h3><span class="sub">per scenario, satoshis</span></div>
|
||
<div id="bars-amount" class="chart-wrap"></div>
|
||
</div>
|
||
<div class="panel">
|
||
<div class="panel-h"><h3>MPP shard budget</h3><span class="sub">max_parts</span></div>
|
||
<div id="bars-parts" class="chart-wrap"></div>
|
||
</div>
|
||
</div>
|
||
</section>
|
||
|
||
<!-- ============ EXPLAINER ============ -->
|
||
<section id="explainer">
|
||
<div class="sec-head">
|
||
<span class="tag">05</span>
|
||
<div>
|
||
<h2>How It Works</h2>
|
||
<p>The simulator, the probability models, and the reflective loop.</p>
|
||
</div>
|
||
</div>
|
||
|
||
<div class="ex-card">
|
||
<h3><span class="n">A</span> The simulator: real lnd code, hidden liquidity</h3>
|
||
<p>
|
||
<code>routing/sim_graph.go</code> is an in-memory Lightning network with hidden
|
||
per-channel balances and per-direction policies (base + ppm fees, CLTV deltas,
|
||
min/max HTLC). It implements lnd's <code>Graph</code> interface, so lnd's
|
||
<em>production pathfinding and mission control run against it unmodified</em>.
|
||
</p>
|
||
<p>
|
||
<code>SendHtlc</code> enforces real forwarding semantics — <code>FeeInsufficient</code>,
|
||
<code>IncorrectCltvExpiry</code>, <code>AmountBelowMinimum</code>,
|
||
<code>TemporaryChannelFailure</code> (a liquidity miss) — updating balances and
|
||
unwinding on failure. Topologies are synthetic (line, grid, hubspoke, smallworld) or
|
||
loaded from a <code>describegraph</code> mainnet snapshot; hidden balances follow a
|
||
<em>half</em>, <em>uniform</em>, or <em>bimodal</em> model, the last hugging one end of
|
||
each channel — the realistic, hard regime.
|
||
</p>
|
||
<p>
|
||
The CLI <code>cmd/routesim</code> takes a params JSON plus a scenario file and emits
|
||
per-attempt traces and aggregate metrics. It is fast: ten payments on a 200-node net in
|
||
~0.3 s including process startup, so thousands of evaluations are cheap.
|
||
</p>
|
||
</div>
|
||
|
||
<div class="ex-card">
|
||
<h3><span class="n">B</span> Mission control & the two probability estimators</h3>
|
||
<p>
|
||
Pathfinding is Dijkstra over a cost roughly equal to
|
||
<code>fees + timelock_penalty + attempt_cost / P(route)</code>: the virtual attempt cost
|
||
(100 sat + 1000 ppm by default) divided by the estimated success probability, so
|
||
unreliable routes look expensive. <code>P(route)</code> is the product of per-hop
|
||
probabilities from one of two estimators, fed by mission control, which records the last
|
||
success/fail amount and timestamp for every directed node pair.
|
||
</p>
|
||
<div class="est-grid">
|
||
<div class="est apriori">
|
||
<h4>Apriori estimator</h4>
|
||
<div class="tagline">default</div>
|
||
<p>
|
||
Every unknown hop starts at 0.6, scaled down as the amount approaches channel
|
||
capacity. Observations override the prior — a pair that just relayed ≥ amt gets 0.95;
|
||
one that failed at ≤ amt drops to ~0, healing back over a 1h half-life. The clever bit
|
||
is node-level extrapolation: a node's failures on <em>other</em> channels drag down all
|
||
its channels via a weighted average (weight 0.5) — one bad channel makes the whole node
|
||
suspect.
|
||
</p>
|
||
</div>
|
||
<div class="est bimodal">
|
||
<h4>Bimodal estimator</h4>
|
||
<div class="tagline">opt-in · models where liquidity sits</div>
|
||
<p>
|
||
Assumes balance density ~ <code>e^(−x/s) + e^((x−c)/s)</code> — funds parked at one end
|
||
or the other (s = 300k sat). Each attempt narrows a per-channel
|
||
[can-send, cannot-send] window: success at <em>a</em> proves ≥ a, failure at <em>b</em>
|
||
proves < b, relaxing back to [0, capacity] over a 7-day decay. Probability is the mass
|
||
above amt within that window; node-level info mixes in (weight 0.2) unless fresh direct
|
||
info exists.
|
||
</p>
|
||
</div>
|
||
</div>
|
||
<p style="margin-top:14px">
|
||
The MPP layer sits on top: if no single route clears <code>MinProbability</code> (1%), the
|
||
session halves the amount and retries, down to a min shard and up to <code>MaxParts</code>
|
||
shards.
|
||
</p>
|
||
</div>
|
||
|
||
<div class="ex-card">
|
||
<h3><span class="n">C</span> GEPA's reflective loop</h3>
|
||
<p>
|
||
The seed is lnd's own defaults (<code>routesim --dump-defaults</code>). On a fresh corpus it
|
||
scores 0.27–0.58 with 30–63% success and up to 26 attempts per payment on the hard bimodal
|
||
nets — <em>plenty of failure for reflection to chew on</em>, which is exactly what GEPA
|
||
needs. A saturated seed would just be returned unchanged.
|
||
</p>
|
||
<p>
|
||
Each round, the reflection LM (<code>openai/gpt-5.6-sol</code>) reads the score plus compact
|
||
failure traces, proposes a mutated params JSON, and the evaluator scores it. Generalization
|
||
mode keeps the incumbent unless a candidate genuinely holds up across splits — so the curve
|
||
climbs by keeping wins and discarding regressions, not by chasing val-set noise.
|
||
</p>
|
||
</div>
|
||
|
||
<div class="ex-card">
|
||
<h3><span class="n">D</span> Phase 3: a paradigm-free router, now evolving</h3>
|
||
<p>
|
||
Parameter tuning was deliberately the loop-validation step — <em>we were never tied to that
|
||
paradigm</em>, and run1 confirmed it: the lnd defaults survived 400 evals unchanged, no knob
|
||
beat them. The real target, now live, is a paradigm-free <code>SimRouter</code> interface in the
|
||
sim: gossip view + local balances + attempt feedback in, a route out — so GEPA generates whole
|
||
routing strategies in Go as candidates, with today's stack (and a hand-written 300-line seed
|
||
router) as the baselines to beat.
|
||
</p>
|
||
<p>
|
||
The bar turned out to be the seed, not lnd: a naive cheapest-path router with per-payment
|
||
failure blacklisting already beats production lnd pathfinding on the hard corpus (0.547 vs
|
||
0.393 objective, 2× the success rate at a third of the attempts). The overnight run,
|
||
<code>code_hard1</code>, went on to clear that bar too: its evolved champion (<b>hb1</b>, 872
|
||
lines) beats <em>both</em> lnd and the seed on the sealed test set (0.586 vs 0.530 vs 0.309
|
||
objective) and on out-of-distribution corpus-v2 topologies it never trained on (0.545 vs 0.487
|
||
vs 0.357) — at 4–7× fewer attempts per payment (≈9 vs lnd's ≈50). From nothing but the
|
||
objective and per-attempt failure traces, it independently arrived at an explicit bimodal
|
||
liquidity prior — the same "funds sit at one end of the channel" hypothesis lnd's own bimodal
|
||
estimator was analytically derived from — plus per-edge liquidity bounds with a confidence
|
||
score and a risk-adjusted Dijkstra, well beyond the seed's simple blacklist.
|
||
</p>
|
||
<p>
|
||
The frontier didn't stop at hb1: <code>code_hard1</code> went on to accept a sibling, <b>hb2</b>
|
||
(1166 lines), that Pareto-trades against it rather than replacing it — hb1 stays the
|
||
hard-regime specialist (sealed test 0.586) while hb2 generalizes better out-of-distribution
|
||
(0.577 vs hb1's 0.545), and both still beat lnd and the seed on every set. That makes this a
|
||
genuine Pareto frontier of evolved routers, not a single champion.
|
||
</p>
|
||
<p>
|
||
<code>code_hard1</code> has since finished: it terminated at 135 of its 400-eval budget when a
|
||
pathological candidate hit an infinite loop and blew the subprocess timeout, which propagated
|
||
and stopped the run (the harness now catches this and scores a hung candidate 0.0 instead of
|
||
crashing, so a future run can use its full budget). GEPA's own valset selection had already
|
||
settled on hb1 as the best program — val aggregate 0.31654, matching the independent hard-val
|
||
measurement to four decimals — so hb1 is the definitive champion by both GEPA's own selection
|
||
and the three-way validation above, with hb2 standing as its best out-of-distribution sibling.
|
||
</p>
|
||
<p>
|
||
A Fable code-reviewer agent independently audited the sandbox mid-run and found one critical
|
||
hole: the graph session handed a candidate the concrete <code>*SimGraph</code>, type-assertable
|
||
back to read hidden balances or rewrite liquidity outright. Fixed the same day — the session now
|
||
only ever exposes the sealed view — and every in-flight candidate (code1, omni1) was checked
|
||
against the escape: zero hits, so no result above is tainted.
|
||
</p>
|
||
<p>
|
||
A follow-up run, <code>code_mix1</code>, seeded GEPA directly from hb1 on a mixed
|
||
hard/scale-free corpus to push further. Reflecting on an 872-line candidate mostly
|
||
produces rejects — 12 proposals in a row before the first accept — and even that accept,
|
||
<b>mb1</b> (1306 lines, iteration 17), doesn't dominate: three-way validation puts it below
|
||
hb1 on the hard sealed test (0.570 vs 0.586) and below both hb1 and hb2 on out-of-distribution
|
||
corpus-v2 (0.529 vs 0.545 / 0.577). GEPA kept it on the Pareto front for per-example
|
||
specialization, not aggregate gain — its own minibatch score (0.431) was below the seed's
|
||
(0.519).
|
||
</p>
|
||
<p>
|
||
The run kept going, though, and the giant-seed reflection eventually paid off: by 314 of
|
||
its 500-eval budget, <code>code_mix1</code> had accepted two more Pareto siblings (1410 and
|
||
1525 lines), the last of which, <b>mx_c3</b> (saved as
|
||
<code>champions/router_mx3_generalist_v1.go</code>), is a genuine generalist champion.
|
||
Three-way validation: hard sealed test hb1 <b>0.586</b> / hb2 0.545 / mx_c3 0.583; OOD
|
||
corpus-v2 hb1 0.545 / hb2 0.577 / mx_c3 <b>0.581</b>; combined average hb1 0.565 / hb2 0.561
|
||
/ mx_c3 <b>0.582</b>. mx_c3 strictly dominates hb2 (wins both sets), ties hb1 on the hard
|
||
test within noise, and posts the best combined average of any router evolved so far.
|
||
Structurally it extends hb1's bimodal-prior-plus-liquidity-bounds core with an adaptive
|
||
retry-at-lower-amount policy (<code>candidateLowerRetryFactor</code>) that pays off on the
|
||
more diverse mixed corpus. Champions of record are now <b>hb1</b> (hard-regime specialist)
|
||
and <b>mx_c3</b> (generalist, superseding hb2) — the "diminishing returns" read on the giant
|
||
seed was premature, it just took ~300 more evals to cash out.
|
||
</p>
|
||
<div class="callout">
|
||
<div class="co-h">The inversion</div>
|
||
<p>
|
||
Everything above — mission control, bimodal windows, apriori priors — stops being a
|
||
constraint and becomes <em>background knowledge the reflection prompt can draw on</em>. The
|
||
optimizer is free to keep it, reshape it, or throw it out, and it has already found something
|
||
the paradigm didn't imagine.
|
||
</p>
|
||
</div>
|
||
</div>
|
||
</section>
|
||
|
||
</div><!-- /wrap -->
|
||
|
||
<footer class="foot">
|
||
<span>GEPA // LN ROUTING OPTIMIZER — mission control</span>
|
||
<span id="foot-updated">data snapshot loading…</span>
|
||
</footer>
|
||
</div><!-- /main -->
|
||
</div><!-- /layout -->
|
||
|
||
<script src="app.js"></script>
|
||
</body>
|
||
</html>
|