XA-06 第一次真實模型 pilot——Qwythos-9B-v2 走 A0→A5 階梯（36 試驗，2026-09-03）

假設

hypothesis

Experiment A's H1–H4 on the three-task gate matrix: does scaffolding raise quality, at what physical cost, with non-constant marginal yield, and is SSR < 1?

設定

model_ids

• MOD-2026-0006 Qwythos-9B-v2（Q4_K_M，本地，Ollama）

benchmark_ids

• BEN-2026-0101 XA-02——A0→A5 鷹架響應的 30 題 pilot 任務包

hardware

NVIDIA GeForce RTX 3070 (8 GiB VRAM; peak 7.59 GiB used, peak 208.9 W, 73 °C), Windows 10 host; physical boundary local-runner-plus-visible-accelerator; energy type device_measured (E-Grade C, CST-B)

software_environment

Ollama serving the model through an OpenAI-compatible endpoint at 127.0.0.1:11434; XA-02/03/04/06/06L v0.1 (hashes verified byte-exact against their manifests before the run); Python 3.14.5; XA-03 collectors system + nvidia_smi at 250 ms target (observed ~335–350 ms)

configuration

provider

mode

openai_compatible

provider_id

local-openai-compatible

model_id

hf.co/empero-ai/Qwythos-9B-v2-GGUF:Q4_K_M

base_max_tokens

1024

temperature

0.2

top_p

1.0

seed

7

budget_control_validated

true

auth_mode

none

matrix

tasks

• MATH-003
• CODE-001
• CON-003

conditions

• A0
• A1
• A2
• A3
• A4
• A5

replicates

2

allow_code_evaluation

false

telemetry

collectors

• system
• nvidia_smi

physical_boundary

local-runner-plus-visible-accelerator

required

false

程序

procedure

XA-06L configure → preflight (all checks PASS; isolated MATH-003/A0 quality 1.0) → run (36/36 terminal, 0 runtime failures) → verify gate (3 points xa03_not_complete) → analyze → seal. The three aborted points were not re-run with resume --force: re-rolling until the gate turns green would erase a real failure mode.

執行

run_count

1

random_seeds

• seed 7 (frozen into every HTTP request); model nondeterminism otherwise uncontrolled

controls

• fresh provider per trial; frozen temperature 0.2, top-p 1.0, seed 7; base_max_tokens 1024 with validated 2× budget for A1
• identical initial task text across conditions; calculator tool contract only in A4/A5 generator requests
• scoring after XA-03 finalization; private references never in context; code evaluation disabled on the host

metrics

gate

status

REAL_MODEL_PILOT_INCOMPLETE

verified_complete_count

33

invalid_or_missing

• point

  CODE-001__A3__r2

  reason

  xa03_not_complete

• point

  CON-003__A3__r1

  reason

  xa03_not_complete

• point

  CON-003__A4__r1

  reason

  xa03_not_complete

headline

ssr

1.0

sdr

0.0

scm_device_energy

3.1019

scm_wall_time

3.1264

protocol_compliance_rate

0.3333

quality_available_rate

0.6111

by_condition

A0

quality_mean

0.5

quality_n

4

success_rate

0.5

wall_time_s

10.21

device_energy_j

1715.6

energy_ratio_vs_A0

1.0

gpu_peak_memory_gib

7.23

gpu_memory_residency_gib_s

72.1

gpu_utilization_integral_s

6.57

A1

quality_mean

0.5

quality_n

4

success_rate

0.5

wall_time_s

9.87

device_energy_j

1819.0

energy_ratio_vs_A0

1.06

gpu_peak_memory_gib

7.32

gpu_memory_residency_gib_s

70.0

gpu_utilization_integral_s

6.26

A2

quality_mean

0.5

quality_n

4

success_rate

0.5

wall_time_s

79.09

device_energy_j

15143.0

energy_ratio_vs_A0

8.827

gpu_peak_memory_gib

7.28

gpu_memory_residency_gib_s

570.0

gpu_utilization_integral_s

54.09

A3

quality_mean

0.6667

quality_n

3

success_rate

0.6667

wall_time_s

99.13

device_energy_j

15921.5

energy_ratio_vs_A0

9.281

gpu_peak_memory_gib

7.34

gpu_memory_residency_gib_s

714.6

gpu_utilization_integral_s

70.02

A4

quality_mean

0.6667

quality_n

3

success_rate

0.6667

wall_time_s

69.08

device_energy_j

13546.2

energy_ratio_vs_A0

7.896

gpu_peak_memory_gib

7.28

gpu_memory_residency_gib_s

496.9

gpu_utilization_integral_s

46.93

A5

quality_mean

0.5

quality_n

4

success_rate

0.5

wall_time_s

31.91

device_energy_j

5321.5

energy_ratio_vs_A0

3.102

gpu_peak_memory_gib

7.24

gpu_memory_residency_gib_s

228.7

gpu_utilization_integral_s

22.17

operational_totals

model_invocations

186

trajectories

162

retries

0

tool_calls

0

verifier_passes

18

candidates_created

162

candidates_selected

33

candidates_discarded

105

failure_events

6

candidates_abandoned_on_abort

24

verifier_failure

count

3

verifier_enabled_trials

18

rate

0.16666666666666666

by_condition

A3

2/6

A4

1/6

A5

0/6

strict_protocol_compliance

count

12

total

36

rate

0.3333333333333333

posthoc_semantic_diagnostic (non-canonical)

canonical

false

math

12/12 canonical correct

code

11/11 selected outputs pass all hidden tests after outer Markdown fence removal

constraint

10/10 selected outputs satisfy all constraints after format-only normalization

completed_selected_outputs_correct

33/33

physical_totals

total_measured_gpu_energy_j

320800.7795

total_measured_gpu_energy_kwh

0.0891

summed_trial_wall_time_min

29.9277

max_gpu_memory_gib

7.5908

max_gpu_power_w

208.88

max_gpu_temperature_c

73.0

mean_sampling_call_wall_fraction

0.2436

quality_availability_reporting_inconsistency

aggregate_analysis

22/36

protocol_compliance_csv

33/36

inconsistency

true

詮釋

interpretation

As an instrument gate it did its job: real numbers, full telemetry, a sealed and relocatable bundle, and an honest INCOMPLETE. As science it says three things and no more. (1) On three tasks the native single pass already solves, scaffolding cannot show a quality gain — SSR = 1 is a legal null result, and it cost 3.1× (A5) to 9.3× (A3) the device energy of A0; A5 was cheaper than the fixed eight-sample conditions only because its loop stopped early. (2) The instrument's quality axis conflated output-format obedience with task correctness on CODE-001 (correct code inside a Markdown fence) and CON-003 (correct assignment written as A=X, not JSON) — exactly the SyntacticValidity ≠ SemanticCorrectness split Paper 06 predicts, now observed in the lab's own instrument. (3) The same-model verifier's serialization failed in 3 of 18 verifier trials and discarded eight candidates each time; tool access was enabled but never used, so tool and retry effects are unidentified. The A3/A4 'gain' is survivor bias from the aborted low-format trials. Seven instrument revisions are required before XA-07; the dataset stays immutable.

限制

limitations

• Three easy tasks, two replicates, one 9B model at 4-bit, one machine; not a population-level estimate of anything.
• Gate INCOMPLETE (33/36); CODE-001 quality unmeasured (execution disabled) so 12 of 36 trials have no measured quality; quality-availability is reported inconsistently inside the bundle (22/36 vs 33/36).
• Device-measured GPU energy only — not marginal, not whole-system; telemetry sampling itself cost ~24 % of trial wall time.
• Same-model verifier; no independent or formal verifier condition.

重現

reproduction_instructions

Unpack XA-02/03/04/06/06L as siblings, .\configure.ps1 (local_openai_compatible, base_url http://127.0.0.1:11434, the model id above), .\preflight.ps1, .\run-pilot.ps1, .\seal-results.ps1; verify the sealed bundle against its manifest.json (266 files). The bundle's raw events/telemetry/summaries are unchanged by sealing.

結果

記錄欄位

completed_at

2026-09-03

關係

歷史與來源歷程

Canonical URL

https://evemisslab.com/ai/experiments/EXP-2026-0103/

機器可讀

/ai/experiments/EXP-2026-0103/index.json

快照

AI-SNAPSHOT-v0.1-fe85b9694a45

來源歷程

source

EveMissLab research collection: Intelligence Physical Metrology (真本體論13)

extracted_by

Splice (Claude Code), reading the canonical UTF-8 sources and each package's own reports

extracted_at

2026-09-11

generator

tools/extract_all.py

claim_boundary

status, evidence level and result type follow the source artifact's own stated claim boundary; nothing is upgraded beyond what the report supports