refusal measurements for the 14 MustRefuse candidates, on the CORRECTED canonical lane
measured 2026-09-11, dev box (1x RTX 4090, sm_89, runc + driver overlay)
code: 79b5b4d2f99415bacf2b698761f2fae6039c5439 (integ/flag-system-20260911)
image: registry.arbi.work/arbi-serve:test-latest rev 9e772d533b179f353e6bcd4924db5fef047f64e8
       (turbo-attn 0.55.2 / exllamav3 1.4.2 / torch 2.12.1 -- all == uv.lock)

Lane: BOOT_LANES["canonical"] verbatim -- Qwen3.8-27B-exl3-4.0bpw TP1 +
Qwen3.8-27B-DFlash2-exl3-6.0bpw drafter, --kv-cache-dtype tkv, TKV_BITS=4,
--enable-mtp --mtp-n-draft 7 --max-batch 8 --max-batched-tokens 2304
--chunk-prefill 2048 --max-context 4096 --gpu-memory-utilization 0.90.
It FITS ONE 4090. Six boots, all booted, all smokes CLEAN (0 errors).
Boot wall: 378 s cold / 74-86 s warm.

DFlash2 drafter confirmed attached and serving on every boot:
dflash_capture_armed=1 dflash_realize_captured=1 dflash2_selector_steps=1
dflash_graph_assemble_step=1.

Windows: BOOT = the pre-reset /v1/admin/flag_truth snapshot taken at ready.
SERVE = the post-smoke snapshot after /v1/admin/flag_truth/reset.
NOTE: boot3 ran the WORKLOAD-SHAPE SWEEP instead of the standard smoke, so its
'serve' row above is the FIRST phase ('short') only. Its full per-phase
readings are in the sweep section at the bottom.

### gdn_prefill_capture
  counter 'prefill_capture_sweep'
    boot1-cold             window=undeclared
                             boot  f=0 r=1 {"disabled by config (prefill_capture='eager')": 1}
                             serve f=0 r=0 {}
    boot2-warm             window=undeclared
                             boot  f=0 r=1 {"disabled by config (prefill_capture='eager')": 1}
                             serve f=0 r=0 {}
    boot3-warm(phase)      window=undeclared
                             boot  f=0 r=1 {"disabled by config (prefill_capture='eager')": 1}
                             serve f=0 r=0 {}
    boot4-warm+otel        window=undeclared
                             boot  f=0 r=1 {"disabled by config (prefill_capture='eager')": 1}
                             serve f=0 r=0 {}
    boot5-freshbudget      window=undeclared
                             boot  f=0 r=1 {"disabled by config (prefill_capture='eager')": 1}
                             serve f=0 r=0 {}
    boot6-freshmanifest    window=undeclared
                             boot  f=0 r=1 {"disabled by config (prefill_capture='eager')": 1}
                             serve f=0 r=0 {}
  counter 'piecewise_sweep'
    boot1-cold             window=undeclared
                             boot  f=0 r=1 {"disabled by config (prefill_capture='eager')": 1}
                             serve f=0 r=0 {}
    boot2-warm             window=undeclared
                             boot  f=0 r=1 {"disabled by config (prefill_capture='eager')": 1}
                             serve f=0 r=0 {}
    boot3-warm(phase)      window=undeclared
                             boot  f=0 r=1 {"disabled by config (prefill_capture='eager')": 1}
                             serve f=0 r=0 {}
    boot4-warm+otel        window=undeclared
                             boot  f=0 r=1 {"disabled by config (prefill_capture='eager')": 1}
                             serve f=0 r=0 {}
    boot5-freshbudget      window=undeclared
                             boot  f=0 r=1 {"disabled by config (prefill_capture='eager')": 1}
                             serve f=0 r=0 {}
    boot6-freshmanifest    window=undeclared
                             boot  f=0 r=1 {"disabled by config (prefill_capture='eager')": 1}
                             serve f=0 r=0 {}

### prefill_replay
  counter 'prefill_graph_lookup'
    boot1-cold             window=undeclared
                             boot  f=0 r=3 {"multi-row prefill batch": 3}
                             serve f=0 r=4 {"real_N above every captured prefill rung": 4}
    boot2-warm             window=undeclared
                             boot  f=0 r=3 {"multi-row prefill batch": 3}
                             serve f=0 r=4 {"real_N above every captured prefill rung": 3, "multi-row prefill batch": 1}
    boot3-warm(phase)      window=undeclared
                             boot  f=0 r=3 {"multi-row prefill batch": 3}
                             serve f=0 r=4 {"real_N above every captured prefill rung": 4}
    boot4-warm+otel        window=undeclared
                             boot  f=0 r=3 {"multi-row prefill batch": 3}
                             serve f=0 r=5 {"real_N above every captured prefill rung": 5}
    boot5-freshbudget      window=undeclared
                             boot  f=0 r=3 {"multi-row prefill batch": 3}
                             serve f=0 r=4 {"real_N above every captured prefill rung": 4}
    boot6-freshmanifest    window=undeclared
                             boot  f=0 r=3 {"multi-row prefill batch": 3}
                             serve f=0 r=4 {"real_N above every captured prefill rung": 3, "multi-row prefill batch": 1}
  counter 'prefill_graph_lookup_continuation'
    boot1-cold             window=undeclared
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot2-warm             window=undeclared
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot3-warm(phase)      window=undeclared
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot4-warm+otel        window=undeclared
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot5-freshbudget      window=undeclared
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot6-freshmanifest    window=undeclared
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}

### recurrent_prefill_chunk
  counter 'recurrent_prefill_chunk_cap'
    boot1-cold             window=undeclared
                             boot  f=0 r=36 {"unset_no_aggregate_cap": 36}
                             serve f=0 r=233 {"unset_no_aggregate_cap": 233}
    boot2-warm             window=undeclared
                             boot  f=0 r=37 {"unset_no_aggregate_cap": 37}
                             serve f=0 r=229 {"unset_no_aggregate_cap": 229}
    boot3-warm(phase)      window=undeclared
                             boot  f=0 r=37 {"unset_no_aggregate_cap": 37}
                             serve f=0 r=68 {"unset_no_aggregate_cap": 68}
    boot4-warm+otel        window=undeclared
                             boot  f=0 r=35 {"unset_no_aggregate_cap": 35}
                             serve f=0 r=220 {"unset_no_aggregate_cap": 220}
    boot5-freshbudget      window=undeclared
                             boot  f=0 r=38 {"unset_no_aggregate_cap": 38}
                             serve f=0 r=239 {"unset_no_aggregate_cap": 239}
    boot6-freshmanifest    window=undeclared
                             boot  f=0 r=36 {"unset_no_aggregate_cap": 36}
                             serve f=0 r=228 {"unset_no_aggregate_cap": 228}

### decode_pad_cudagraph
  counter 'decode_pad_replay'
    boot1-cold             window=undeclared
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot2-warm             window=undeclared
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot3-warm(phase)      window=undeclared
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot4-warm+otel        window=undeclared
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot5-freshbudget      window=undeclared
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot6-freshmanifest    window=undeclared
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
  counter 'decode_pad_borrow_alias_consumed'
    boot1-cold             window=serve
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot2-warm             window=serve
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot3-warm(phase)      window=serve
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot4-warm+otel        window=serve
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot5-freshbudget      window=serve
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot6-freshmanifest    window=serve
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}

### split_mixed_min_decode_rows
  counter 'split_mixed_forward'
    boot1-cold             window=undeclared
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot2-warm             window=undeclared
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot3-warm(phase)      window=undeclared
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot4-warm+otel        window=undeclared
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot5-freshbudget      window=undeclared
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot6-freshmanifest    window=undeclared
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}

### mtp_fused_rejection
  counter 'mtp_fused_rejection_sample'
    boot1-cold             window=undeclared
                             boot  f=1 r=0 {}
                             serve f=0 r=0 {}
    boot2-warm             window=undeclared
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot3-warm(phase)      window=undeclared
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot4-warm+otel        window=undeclared
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot5-freshbudget      window=undeclared
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot6-freshmanifest    window=undeclared
                             boot  f=1 r=0 {}
                             serve f=0 r=0 {}

### persistent_recurrent_index
  counter 'persistent_recurrent_index_reuse'
    boot1-cold             window=undeclared
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot2-warm             window=undeclared
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot3-warm(phase)      window=undeclared
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot4-warm+otel        window=undeclared
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot5-freshbudget      window=undeclared
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot6-freshmanifest    window=undeclared
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}

### tp_copy_engine_on_p2p
  counter 'tp_copy_engine_armed'
    boot1-cold             window=boot
                             boot  f=0 r=1 {"flag_off": 1}
                             serve f=0 r=0 {}
    boot2-warm             window=boot
                             boot  f=0 r=1 {"flag_off": 1}
                             serve f=0 r=0 {}
    boot3-warm(phase)      window=boot
                             boot  f=0 r=1 {"flag_off": 1}
                             serve f=0 r=0 {}
    boot4-warm+otel        window=boot
                             boot  f=0 r=1 {"flag_off": 1}
                             serve f=0 r=0 {}
    boot5-freshbudget      window=boot
                             boot  f=0 r=1 {"flag_off": 1}
                             serve f=0 r=0 {}
    boot6-freshmanifest    window=boot
                             boot  f=0 r=1 {"flag_off": 1}
                             serve f=0 r=0 {}

### savepoint_decode_write
  counter 'savepoint_decode_write'
    boot1-cold             window=serve
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot2-warm             window=serve
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot3-warm(phase)      window=serve
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot4-warm+otel        window=serve
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot5-freshbudget      window=serve
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot6-freshmanifest    window=serve
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
  counter 'savepoint_prompt_end_write'
    boot1-cold             window=serve
                             boot  f=0 r=0 {}
                             serve f=1 r=0 {}
    boot2-warm             window=serve
                             boot  f=0 r=0 {}
                             serve f=1 r=0 {}
    boot3-warm(phase)      window=serve
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot4-warm+otel        window=serve
                             boot  f=0 r=0 {}
                             serve f=1 r=0 {}
    boot5-freshbudget      window=serve
                             boot  f=0 r=0 {}
                             serve f=1 r=0 {}
    boot6-freshmanifest    window=serve
                             boot  f=0 r=0 {}
                             serve f=1 r=0 {}

### dflash_dynamic_k
  counter 'dspark_dynamic_k_consulted'
    boot1-cold             window=undeclared
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot2-warm             window=undeclared
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot3-warm(phase)      window=undeclared
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot4-warm+otel        window=undeclared
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot5-freshbudget      window=undeclared
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot6-freshmanifest    window=undeclared
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
  counter 'dspark_dynamic_k_truncated_steps'
    boot1-cold             window=undeclared
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot2-warm             window=undeclared
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot3-warm(phase)      window=undeclared
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot4-warm+otel        window=undeclared
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot5-freshbudget      window=undeclared
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot6-freshmanifest    window=undeclared
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
  counter 'dflash_dynamic_k_threshold_read'
    boot1-cold             window=undeclared
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot2-warm             window=undeclared
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot3-warm(phase)      window=undeclared
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot4-warm+otel        window=undeclared
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot5-freshbudget      window=undeclared
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot6-freshmanifest    window=undeclared
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}

### exl3_int8_verify_min_rows
  counter 'exl3_int8_verify_gemm'
    boot1-cold             window=serve
                             boot  f=1 r=3203 {"verify rows>48: the logits head keeps the shipped leg \u2014 the accept-rate verdict this leg rests on was measured with the head on it, and gate3_kl's serveable set excludes the head's trellis width": 3, "verify slate below min rows": 3200}
                             serve f=0 r=0 {}
    boot2-warm             window=serve
                             boot  f=1 r=3200 {"verify slate below min rows": 3200}
                             serve f=0 r=0 {}
    boot3-warm(phase)      window=serve
                             boot  f=1 r=3200 {"verify slate below min rows": 3200}
                             serve f=0 r=0 {}
    boot4-warm+otel        window=serve
                             boot  f=1 r=3200 {"verify slate below min rows": 3200}
                             serve f=0 r=0 {}
    boot5-freshbudget      window=serve
                             boot  f=1 r=3200 {"verify slate below min rows": 3200}
                             serve f=0 r=0 {}
    boot6-freshmanifest    window=serve
                             boot  f=1 r=3203 {"verify rows>48: the logits head keeps the shipped leg \u2014 the accept-rate verdict this leg rests on was measured with the head on it, and gate3_kl's serveable set excludes the head's trellis width": 3, "verify slate below min rows": 3200}
                             serve f=0 r=0 {}
  counter 'exl3_int8_verify_gemm_captured'
    boot1-cold             window=graph_record
                             boot  f=1 r=1600 {"verify slate below min rows": 1600}
                             serve f=0 r=0 {}
    boot2-warm             window=graph_record
                             boot  f=1 r=1600 {"verify slate below min rows": 1600}
                             serve f=0 r=0 {}
    boot3-warm(phase)      window=graph_record
                             boot  f=1 r=1600 {"verify slate below min rows": 1600}
                             serve f=0 r=0 {}
    boot4-warm+otel        window=graph_record
                             boot  f=1 r=1600 {"verify slate below min rows": 1600}
                             serve f=0 r=0 {}
    boot5-freshbudget      window=graph_record
                             boot  f=1 r=1600 {"verify slate below min rows": 1600}
                             serve f=0 r=0 {}
    boot6-freshmanifest    window=graph_record
                             boot  f=1 r=1600 {"verify slate below min rows": 1600}
                             serve f=0 r=0 {}

### serve_flat_cache_max_gb
  counter 'flat_cache_lru_evict'
    boot1-cold             window=undeclared
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot2-warm             window=undeclared
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot3-warm(phase)      window=undeclared
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot4-warm+otel        window=undeclared
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot5-freshbudget      window=undeclared
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot6-freshmanifest    window=undeclared
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}

### drafter_penalty_align
  counter 'drafter_penalty_align'
    boot1-cold             window=boot
                             boot  f=0 r=2 {"flag_off": 2}
                             serve f=0 r=0 {}
    boot2-warm             window=boot
                             boot  f=0 r=2 {"flag_off": 2}
                             serve f=0 r=0 {}
    boot3-warm(phase)      window=boot
                             boot  f=0 r=2 {"flag_off": 2}
                             serve f=0 r=0 {}
    boot4-warm+otel        window=boot
                             boot  f=0 r=2 {"flag_off": 2}
                             serve f=0 r=0 {}
    boot5-freshbudget      window=boot
                             boot  f=0 r=2 {"flag_off": 2}
                             serve f=0 r=0 {}
    boot6-freshmanifest    window=boot
                             boot  f=0 r=2 {"flag_off": 2}
                             serve f=0 r=0 {}
  counter 'drafter_penalty_engaged'
    boot1-cold             window=serve
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot2-warm             window=serve
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot3-warm(phase)      window=serve
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot4-warm+otel        window=serve
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot5-freshbudget      window=serve
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot6-freshmanifest    window=serve
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}

### dflash2_penalty_align
  counter 'dflash2_penalty_align'
    boot1-cold             window=boot
                             boot  f=0 r=2 {"flag_off": 2}
                             serve f=0 r=0 {}
    boot2-warm             window=boot
                             boot  f=0 r=2 {"flag_off": 2}
                             serve f=0 r=0 {}
    boot3-warm(phase)      window=boot
                             boot  f=0 r=2 {"flag_off": 2}
                             serve f=0 r=0 {}
    boot4-warm+otel        window=boot
                             boot  f=0 r=2 {"flag_off": 2}
                             serve f=0 r=0 {}
    boot5-freshbudget      window=boot
                             boot  f=0 r=2 {"flag_off": 2}
                             serve f=0 r=0 {}
    boot6-freshmanifest    window=boot
                             boot  f=0 r=2 {"flag_off": 2}
                             serve f=0 r=0 {}
  counter 'dflash2_penalty_engaged'
    boot1-cold             window=serve
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot2-warm             window=serve
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot3-warm(phase)      window=serve
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot4-warm+otel        window=serve
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot5-freshbudget      window=serve
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}
    boot6-freshmanifest    window=serve
                             boot  f=0 r=0 {}
                             serve f=0 r=0 {}


WORKLOAD-SHAPE SWEEP (boot3, one boot, counters reset between phases)
Only counters that moved are listed; every other candidate counter read
f=0 r=0 {} in every phase.

  phase short (1.7s, errors=[])
    prefill_graph_lookup             f=0 r=4 {"real_N above every captured prefill rung": 4}
    recurrent_prefill_chunk_cap      f=0 r=68 {"unset_no_aggregate_cap": 68}

  phase short_greedy (0.7s, errors=[])
    prefill_graph_lookup             f=0 r=4 {"real_N above every captured prefill rung": 4}
    recurrent_prefill_chunk_cap      f=0 r=28 {"unset_no_aggregate_cap": 28}

  phase long_chunked (1.2s, errors=[])
    prefill_graph_lookup             f=0 r=3 {"real_N above every captured prefill rung": 1, "chunked-prefill continuation chunk": 2}
    recurrent_prefill_chunk_cap      f=0 r=27 {"unset_no_aggregate_cap": 27}
    savepoint_prompt_end_write       f=1 r=0 {}

  phase concurrent8 (1.5s, errors=[])
    prefill_graph_lookup             f=0 r=1 {"multi-row prefill batch": 1}
    recurrent_prefill_chunk_cap      f=0 r=34 {"unset_no_aggregate_cap": 34}

  phase ramp8 (2.8s, errors=[])
    prefill_graph_lookup             f=0 r=1 {"real_N above every captured prefill rung": 1}
    recurrent_prefill_chunk_cap      f=0 r=83 {"unset_no_aggregate_cap": 83}

  phase mixed_coadmit (2.3s, errors=[])
    prefill_graph_lookup             f=0 r=1 {"real_N above every captured prefill rung": 1}
    recurrent_prefill_chunk_cap      f=0 r=84 {"unset_no_aggregate_cap": 84}
    savepoint_prompt_end_write       f=1 r=0 {}

