Add resume-on-new-instance guide

Refresh code/ with latest BLT-Reasoner sources (post-campaign)

HANDOFF v3: 77.5% via bottleneck-as-regularizer at inference

BREAKTHROUGH: same GRPO ckpt eval without bottleneck = 77.5% GSM8K AR (vs 52.5% with bottleneck)

Per-CoT-step multi-objective experiment

GRPO Phase C from Options 1+3 ckpt

GRPO Phase C from Options 1+3 ckpt

EXP 1+3: AR n=200 K=16 ablation result (51% normal, delta_zero=50.5pp)

EXP 1+3: TF n=200 K=16 ablation result (delta_zero=23pp)

EXP: Options 1+3 (full-y InfoNCE + MLP projector)

Add ckpt-step3000

Add ckpt-step2000

Add ckpt-step1000

EXP: block_z_to_x=True (leak-closure principled test)

Add teacher-forced ablation on pilot7b final (baseline for block_z_to_x exp)

EXP: 7B with block_z_to_x=True (leak-closure principled test)

7B pilot: trained policy + final ablation

Add ckpt-step3000

Add ckpt-step2000

Add ckpt-step1000

HANDOFF: add GRPO Phase C null result + revised next steps

GRPO Phase C: trained policy + final ablation

Add publication-grade HANDOFF: 4-way ablation, mechanism, lineage

Add control: no_bottleneck (final ckpt + n=100 ablation)

Add control: no_infonce (final ckpt + n=100 ablation)

Regenerate README with final ablation results

Refresh logs/metrics.jsonl at end of pilot

Refresh logs/run.log at end of pilot

Refresh logs/auto_eval.log at end of pilot

Add final ckpt + n=200 pre-registered z-ablation

Add ckpt-step12000

Add ckpt-step10000

BLT-Reasoner pilot 1: ckpts + code + logs + ablations

initial commit