# READING · the quantization ladder (generated by bin/reading.py; every figure from its file)

Internal. Counts and scores only; the doorman's items never appear. Plan: PREREG-QUANT.md (frozen 2026-10-14T20:52:44Z) + AMENDMENT-1.md (stamped 2026-10-14T22:15:53Z and 2026-10-15T02:18:44Z); gates: its section 6.

## Speed (the speed ladder; 300 W; a void window carries no figure)

| level | store GiB | largest whole window | decode @4,096 tok/s (spread) | @32,768 | TTFT @4,096 ms | board W | J/1k | resident | void windows | file |
|---|---:|---:|---:|---:|---:|---:|---:|---|---|---|
| q8_0 | 12.13 | 98304 | 60.545 (0.0005) | 60.41 | 548.22 | 259.9 | 4293.87 | True | 131072: not whole: not tested | `gpu-ql3090-24g-nemo-q8-0-control-3090-q8_0-ladder.json` |
| q6_K | 9.37 | 131072 (topped out) | 65.724 (0.0067) | 64.83 | 518.25 | 270.81 | 4102.0 | True | none | `gpu-ql3090-24g-nemo-q6-k-control-3090-q6_K-ladder.json` |
| q5_K_M | 8.13 | 131072 (topped out) | 78.942 (0.0009) | 78.158 | 516.45 | 264.11 | 3345.62 | True | none | `gpu-ql3090-24g-nemo-q5-k-m-control-3090-q5_K_M-ladder.json` |
| q4_K_M | 6.96 | 131072 (topped out) | 88.203 (0.0029) | 87.234 | 481.8 | 259.65 | 2948.31 | True | none | `gpu-ql3090-24g-nemo-q4-k-m-control-3090-q4_K_M-ladder.json` |
| q3_K_M | 5.67 | 131072 (topped out) | 74.894 (0.0053) | 74.135 | 510.56 | 257.59 | 3439.41 | True | none | `gpu-ql3090-24g-nemo-q3-k-m-control-3090-q3_K_M-ladder.json` |

## Answers (six-way 108, doorman 48; strict parser v2-benchbox-1; paired against Q8_0)

| level | six-way strict (Wilson 95 %) | same choice as Q8_0 | McNemar p vs Q8_0 | doorman strict | same as Q8_0 | median s/decision | board J per correct | decode probe tok/s | growth MiB (store MiB) | not tested / failed |
|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---|
| q8_0 | 54/108 ([0.4073, 0.5927]) | — | — | 34/48 | — | 0.632 | 239.28 | 60.15 | 12795.0 (12419) | — |
| q6_K | 55/108 ([0.4163, 0.6016]) | 103 | 1.0 | 34/48 | 48 | 0.6375 | 230.47 | 65.8 | 10121.0 (9590) | — |
| q5_K_M | 57/108 ([0.4343, 0.6194]) | 103 | 0.25 | 33/48 | 46 | 0.6298 | 223.82 | 78.59 | 8939.0 (8323) | — |
| q4_K_M | 47/108 ([0.3455, 0.5293]) | 87 | 0.0654 | 33/48 | 45 | 0.6148 | 262.99 | 87.22 | 7827.0 (7131) | — |
| q3_K_M | 38/108 ([0.2683, 0.4456]) | 80 | 0.0 | 34/48 | 41 | 0.6209 | 337.58 | 74.18 | 6583.0 (5801) | — |

## Predictions (frozen bars)

- **P1 REFUTED** — {"decode_4k": {"q8_0": 60.545, "q6_K": 65.724, "q5_K_M": 78.942, "q4_K_M": 88.203}, "run_spread_4k": {"q8_0": 0.033, "q6_K": 0.441, "q5_K_M": 0.068, "q4_K_M": 0.252}, "gaps": {"q8_0->q6_K": {"gap": 5.179, "larger_spread": 0.441, "separated": true}, "q6_K->q5_K_M": {"gap": 13.218, "larger_spread": 0.441, "separated": true}, "q5_K_M->q4_K_M": {"gap": 9.261, "larger_spread": 0.252, "separated": true}}, "in_bar": {"q8_0": false, "q6_K": true, "q5_K_M": true, "q4_K_M": true}, "rising_and_separated": true}
- **P2 REFUTED** — {"relative_to_median": {"q8_0": 0.1929, "q6_K": 0.0, "q5_K_M": 0.0424, "q4_K_M": -0.0022, "q3_K_M": -0.3107}, "five_level_median_tok_s_bytes": 660935178565.9081}
- **P3 HOLDS** — {"largest_whole": {"q8_0": 98304, "q6_K": 131072, "q5_K_M": 131072, "q4_K_M": 131072, "q3_K_M": 131072}}
- **P4 HOLDS** — {"six_way_same": {"n": 108, "same_choice": 108, "anchor_strict": 47}, "doorman_same": {"n": 48, "same_choice": 48, "anchor_strict": 33}, "strict_now": 47}
- **P5 HOLDS** — {"strict": {"q8_0": 54, "q6_K": 55, "q5_K_M": 57, "q4_K_M": 47, "q3_K_M": 38}, "q3_strictly_lowest": true, "mcnemar_p_vs_q8": {"q6_K": 1.0, "q5_K_M": 0.25, "q4_K_M": 0.0654}}
- **P6 REFUTED** — {"doorman_strict": {"q8_0": 34, "q6_K": 34, "q5_K_M": 33, "q4_K_M": 33, "q3_K_M": 34}, "q3_strictly_lowest": false}
- **P7 HOLDS** — {"median_s": {"q8_0": 0.632, "q6_K": 0.6375, "q5_K_M": 0.6298, "q4_K_M": 0.6148, "q3_K_M": 0.6209}, "max_over_min": 1.037}
- **P8 HOLDS** — {"growth_over_store_mib": {"q8_0": 376, "q6_K": 531, "q5_K_M": 616, "q4_K_M": 696, "q3_K_M": 782}, "same_order": true}

## Stops and refusals

- tasks-q6_K-20261015T061017Z: {'stop': 'qwatch', 'why': 'foreign compute process(es) on the bench card: pid 25570', 'stamp': '2026-10-15T06:10:40Z', 'reads': 12}