custom_miner v3 CANDIDATE — de-trended (worker_2)
A calibrated evolution of v2 that fixes v2's largest measured divergence from real data. It is a strong lead, not a confirmed ship — read the caveats.
The insight (data-driven)
A statistical fingerprint (scratchpad/statgap.py) showed v2's corpus has far too much linear
trend, and the excess grows with series length (v2's slope*t): trend-strength median ~0.42
at length 1536 vs real weather ~0.026. v2 was teaching the model to over-extrapolate trends.
The change (vs v2 — a clean 2-family diff)
_trend_seasonal_ar and _multiplicative draw the total trend excursion over the series
directly (exc * t/(L-1)) instead of slope*t. Effect: the trend is ~16x weaker than v2's and
bimodal (~75% low-trend + ~25% trending). (Note: because builders run at max_len then crop, this
is v2-with-smaller-bimodal-slope, not true per-series normalization.) Persistence, noise, and all
family weights are identical to v2. Tunable via config (tr_exc_lo/hi, gr_exc_lo/hi,
tr_hi_frac). Deterministic; cascade verify passes.
Measured results (real Toto2, multi-seed pooled, honest feed-level clustering, real weather)
| regime | v3 vs v2 |
|---|---|
| ctx 1024 / gen-max 2048 | +12.9%, LCB +0.077 — robust WIN |
| ctx 1536 / gen-max 4096 (production length) | +8.4%, LCB +0.056 — robust WIN |
Fingerprint: v3 trend-strength median at length 1536 is 0.008 (near real 0.026) vs v2's 0.42. Unlike a first slope-based variant (which won +14% at ctx1024 but collapsed to a tie at gen-max 4096), this holds across both regimes tested.
Caveats (from adversarial review — DO NOT ship without resolving)
- Not tested at production CONTEXT 4096. The win shrinks as context grows (+12.9% → +8.4% from ctx 1024 → 1536); extrapolated to 4096 it could narrow toward the +2% margin. Production uses ctx 4096; I could only reach ctx 1536.
- It is a LOW-TREND SPECIALIST — it LOSES multi-domain pools. On the synthetic 8-domain pool at ctx 1536 it scores −22%: it wins low-trend domains (energy +23%, traffic +21%, weather +17%) but loses trending ones (web −47%, econ −10%, retail −2%). The real subnet pool is multi-domain; v3 nets positive only if that pool is low-trend-dominated (plausible — it is weather + pageviews heavy — but unconfirmed). Its big wins are all on genuinely low-trend REAL data (weather); its losses are on trending data.
- Decisive test: v3 at generation & eval context 4096 on the full mixed real pool, per-seed win rate. Needs the operator's real private pool / more GPU than this run had.
Recommendation: keep v2 (worker_1) live; treat v3 as the leading candidate and A/B it against the real pool at ctx 4096 before any deploy. Found via a rigorous, adversarially-reviewed search that also rejected TempoPFN+CauKer, CauKer/GP injections, and 6 reweights (none beat v2).
UPDATE — production-geometry re-test (ctx 4096, 2026-07-13)
Re-tested at the operator's new production geometry (context 4096, gen-max 4096, after the max_length 2048→4096 change). The big short-context wins FADE: v3 vs v2 context trajectory = +12.9% (ctx1024) → +8.4% (1536) → +1.9% (2048) → +3.4% (ctx4096, 5/5 seeds positive but LCB +0.005 < the 0.02 margin = TIE) on real weather. Per-domain at ctx4096: still a LOW-TREND SPECIALIST — wins energy +16%/traffic +15%/weather +3%, loses web −47%/finance −6%/retail −6% (synthetic multi-domain OVERALL −30%). VERDICT: v3 does NOT robustly beat v2 at production; it's a consistent hair-better-on-low-trend, worse-on-trending trade-off. v2 (worker_1) remains the multi-domain choice. v3's only unambiguous edge: numpy generation at 3.16M tok/s (17x the 185k wall-clock cap), immune to the new deadline_hit truncation that threatens the GP/CauKer base king.