# ML-T6R terminal report

**Terminal outcome:** `T3_DYNAMIC_REDUNDANCY_NO_COUNTERFACTUAL_ADVANTAGE`  
**Terminal vector `[H1,H2,H3,H4]`:** `[True, True, False, False]`

## H1 — Static redundancy

- Pass: **True**
- Method: `random`
- Registered removal: 25.0%
- Realized removal: 25.0000% (95% bootstrap interval 25.0000% to 25.0000%)
- Final evaluation-loss delta: 0.0005 (95% bootstrap interval -0.0002 to 0.0012)

A cheap static selector achieved registered physical removal within the quality envelope.

## H2 — Dynamic redundancy

- Pass: **True**
- Exact bottom-set turnover: 79.1667% (95% bootstrap interval 66.6667% to 83.3333%)
- Primary panel: `constant_mixture`

The exact dispensable set changed materially during constant-mixture training.

## H3 — Long-horizon counterfactual value

- Pass: **False**
- Frozen ordinary comparator: `adam_influence`
- Spearman advantage: -0.0377 (95% bootstrap interval -0.3551 to 0.0841)
- Drop-regret advantage: 0.0000 (95% bootstrap interval -0.0003 to 0.0002)

Incremental long-horizon value beyond the strongest ordinary baseline was not established.

## H4 — Fully charged economics

- Pass: **False**
- Frozen cheap policy: `dynamic/random` at 25.0% removal
- Frozen learned policy: `counterfactual_surrogate` at 50.0% removal
- Quality delta versus cheap: 0.0018 (95% bootstrap interval 0.0009 to 0.0025)
- Quality delta versus full: 0.0025 (95% bootstrap interval 0.0014 to 0.0030)
- Wall-time reduction versus cheap: 2.7814% (95% bootstrap interval 2.5741% to 3.1607%)
- FLOP reduction versus cheap: 21.4193% (95% bootstrap interval 21.4193% to 21.4193%)

Fully charged economic superiority over the best cheap policy was not established.

## Scope

This is `LoRA continued pretraining` evidence for `Qwen/Qwen2.5-1.5B` on `HuggingFaceFW/fineweb-edu`. It is not full-parameter pretraining evidence and does not transfer automatically to other models, corpora, sequence lengths, optimizers, or GPUs.