btx-e003 traces v3: d5_traces_v3.jsonl

btx-e003 traces v3: d5_traces_v3_manifest.json

btx-e003 I-6(c) Qwen convention map

btx-e003 R3b lenient re-gauge

btx-e003 R3 ledger final

btx-e003 R2 ledger final

btx-e003 R2 ledger

btx-e003 R2 e-s1

btx-e003 R2 ledger

btx-e003 R2 e-s0

btx-e003 R2 ledger

btx-e003 R2 d-s1

btx-e003 R2 ledger

btx-e003 R2 d-s0

btx-e003 R2 ledger

btx-e003 R2 c-s1

btx-e003 R2 ledger

btx-e003 R2 c-s0

btx-e003 R2 ledger

btx-e003 R2 b-s1

btx-e003 R2 ledger

btx-e003 R2 b-s0

btx-e003 R2 ledger

btx-e003 R2 a-s0-repeat

btx-e003 R2 ledger

btx-e003 R2 a-s1

btx-e003 R2 ledger

btx-e003 R2 a-s0

btx-e003 I-4 teacher gate re-scored

btx-e003: second HELD sub3 exam draw (spec + items)

btx-e003 I-5 codebooks: codebook_sub3_states.json

btx-e003 I-5 codebooks: codebook_d5_states.json

btx-e003 I-5 codebooks: codebook_sub3.npz

btx-e003 I-5 codebooks: codebook_d5.npz

btx-e003 I-5 codebooks: codebooks_i5.json

btx-e003 instruments: btx_e003_R1_ledger.json

btx-e003 instruments: teacher_frame_vs_content.json

btx-e003 R0 S1 repaired (arm geometry from the index)

btx-e003 I-2 census, convention-normalized

btx-e003 R0 instruments: r0_done.json

btx-e003 R0 instruments: collinearity_gate.json

btx-e003 R0 instruments: simct_census.json

btx-e003 R0 instruments: pushforward_gate.json

btx-e003 R0 instruments: teacher_gates.json

btx-e003 R0 instruments: exam150_floors.json

btx-e003 R0 instruments: exam_d5_strict_alien_150_v0.json

btx-e003 R0 instruments: exam_d5_strict_alien_150.json

Campaign B v2 (InfoNCE mix) both seeds PASS; retro-erank corrected after overwrite catch (HF copies were the true v1): spread v1 180 > v2 139 (baseline 239) — target granularity governs spread, not loss family

Campaign B v2 (InfoNCE mix) both seeds PASS; retro-erank corrected after overwrite catch (HF copies were the true v1): spread v1 180 > v2 139 (baseline 239) — target granularity governs spread, not loss family

Campaign B v2 (InfoNCE mix) both seeds PASS; retro-erank corrected after overwrite catch (HF copies were the true v1): spread v1 180 > v2 139 (baseline 239) — target granularity governs spread, not loss family