RLTT/generator-v26

generator_v25 — persistence specialists

Lineage: v18 → v22 → v23 → v24 → v25. Fork of v24.

What the lineage established

Paired attribution against the reigning king (UID 158) over 1536 fresh windows (July 27-29, two window seeds each) located the gap precisely. Sorting upstream feeds by how often they repeat the previous sample ordered v18's losses almost perfectly:

feed	windows	zero-diff frac	levels	v18 vs king	v24 vs king
gbfs_bergen_station_status	57	0.968	7	-76.9%	-53.4%
nextbike_zagreb_station_status	87	0.949	12	-52.6%	-37.9%
gbfs_trondheim_station_status	32	0.925	9	-20.8%	-17.5%
gbfs_bicing_barcelona_station_status	88	0.638	20	+5.6%	+4.4%

Bergen and Zagreb alone account for about 4 of the 5 percentage points between v18 and the king. These feeds hold a value bit-exact for tens of samples, so their naive-difference scale is tiny and a forecast that drifts even a fraction of one count is punished heavily.

Three attempts narrowed down what actually helps:

• v22 (dock family + a blanket link rounding all families onto small integer ranges) scored 0.40635 against v18's 0.39113. The targeted part worked (Bergen +32%) but the blanket link coarsened the fine-grained feeds, costing 10% on Barcelona and 49% on German hospitalization counts.
• v23 re-targeted the link on flat fraction rather than capacity. Result discarded: three copies of the run launched at once and overwrote each other's checkpoint directory.
• v24 (dock family only, 14% mass) reached 0.39272, a tie with v18, at full token parity — 9.64B tokens against v18's 9.73B. It won exactly where aimed (Bergen +13.3%, Zagreb +9.6%, NY COVID testing +15.3%) but the mass it took from ou_stochastic_vol, regime_shift, integrated, and threshold_ar cost 19.7% on NZ electricity prices over 101 windows, cancelling the gain.

Changes in v25

• New piecewise_level family at 11% mass. Levels are held exactly constant for hundreds of steps and broken by rare changes, with a bounded-integer variant (3-40 capacity, unit moves) and a continuous variant. Change rates are drawn log-uniformly over 0.05%-9% of samples, so the corpus spans both the near-frozen feeds and the moderately active ones. Only a quarter of continuous rows carry reading noise; the rest stay bit-exact, which is the property the metric rewards most.
• dock_occupancy raised to 15% and sharpened toward the flat band the pool actually occupies: swing 0.03-0.6 of capacity (was 0.04-1.1), jump rate 0.002-0.09 (was 0.003-0.2), seasonal amplitude up to 0.30 of capacity (was 0.45). Median flat fraction moves from 0.81 to 0.89.
• Funding now comes only from forecast_tasks (0.435 → 0.261) and the generic stochastic families. regime_shift, integrated, threshold_ar, ou_stochastic_vol, pulse_outlier, tidal_harmonic, physical_sensors, seasonal_counts, intermittent, stable_calendar_counts, and conditional_stability are restored to their v18 weights, since those are what serve the spiky price and sensor feeds v24 damaged.
• observation.small_count_rate stays at 0, so no blanket requantization.

Corpus coverage, last 512 points of 2000 series

	flat >= 0.8	flat >= 0.90	flat >= 0.95	<= 24 levels
pool	43.7%	23.1%	11.8%	64.5%
v18	11.0%	8.4%	7.1%	13.4%
v24	16.3%	13.2%	11.0%	22.9%
v25	25.6%	22.5%	19.8%	32.0%

Throughput is 799 series/s against v18's 729 measured back-to-back, since both new families are cheaper per row than the forecast_tasks mass they replace, so the training budget should not lose tokens.