Weco AI reports 'Level 1' harness-only recursive self-improvement — a real bounded result, softer than the headline

Why it matters (spine). This is exactly the kind of frontier claim the RDKG machine layer exists to price honestly. Weco AI, a startup, published a first-party report claiming "the first autonomous recursively self-improving system that materially accelerated the AI R&D process." Strip the label and the measured core is real but narrow: an outer loop (AIDE²) autonomously rewrote the harness and scaffolding of an inner ML-research agent for 100 consecutive steps, with no human in the loop, over ~8 days of wall-clock time — and the agents it discovered mid-run beat Weco's own hand-tuned AIDEhuman, the product of roughly two years of manual tuning, on three external benchmarks the loop never optimized against (e.g. MLE-Bench Lite, AIDE47 at +0.053, p=0.0024). Two facts keep this at signal, not verified: no neural weights change — this is prompt- and harness-level optimization only — and Weco self-grades it "Level 1 / net positive" on its own 0–3 RSI scale, explicitly denying ignition (Level 2) or an intelligence explosion. It is non-peer-reviewed, first-party, and not independently reproduced. The money-and-metacrisis lever is a genuinely measured acceleration of AI R&D; the discipline lever is refusing to let a bounded result be read as the thing its headline implies.

What would prove it wrong (falsifier). Independent replication fails, or the gains do not generalize beyond Weco's own benchmark suite. Conversely, a peer-reviewed, independently-reproduced result on comparable held-out tasks would promote this past signal.

Live-status check + triangulation. Quoted from Weco AI's first-party report, "first-evidence-of-recursive-self-improvement" (Weco AI, dated 14 July 2026): "We believe this is the first autonomous recursively self-improving system that materially accelerated the AI R&D process." Verified against the primary in-session; note the source is a company self-report, not an external audit. Secondary coverage — FourWeekMBA (fourweekmba.com/ai-weco-ai-aide2-recursive-self-improvement-benchmark) and the Daily AI Digest (buttondown.com/dailyaidigest) — corroborates the account but is downstream summary, not independent replication.

Correction carried forward (discipline). Three claims in the working draft were tightened against the primary: (1) the draft said "Weco self-grades it Level 1 of 3," but the primary defines a four-tier 0–3 scale (Level 0 delegation, Level 1 net positive, Level 2 ignition, Level 3 inflection) and grades AIDE² at Level 1 — the scale tops out at 3 but has four levels, so "of 3" is imprecise; (2) the draft said it "beat Weco's own hand-tuned agent on held-out benchmarks," but the primary specifies the winning discovered agents were AIDE47 and AIDE85 (not the final 100th-step AIDE99), with AIDE85 beating AIDEhuman across all three external benchmarks the loop never saw (MLE-Bench Lite, ALE-Bench Lite, WeatherBench 2); (3) the draft framed the run as beating the baseline in "~100 steps," but more precisely the full autonomous run was 100 consecutive steps (AIDE0→AIDE99) with ~90% of proposed rewrites rejected, and the best agents emerged mid-run, not at step 100.