PACC-Hybrid v0.2——真實模型第三次執行：開啟思考並放大輸出預算

假設

hypothesis

With the model's own reasoning enabled (the setting runs 1–2 had to switch off), the v0.2 predeclared picture — coherence and intent-persistence gains plus higher valid novelty for the PACC runtime, with a recoverable breadth loss — appears.

設定

model_ids

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

benchmark_ids

• BEN-2026-0003 PACC-LLM 混合 A/B/C benchmark

software_environment

Python 3.14, openai SDK 3.0.0 against Ollama 0.33.3 /v1/responses (OLLAMA_FLASH_ATTENTION=1, OLLAMA_KV_CACHE_TYPE=q8_0); nomic-embed-text for breadth; harness package unchanged.

程序

procedure

scripts/run_real_local_ollama.py --model qwythos-9b-v2-q4km-ctx12k --reasoning-effort medium --budget-add 3500 --candidate-count 4 --repetitions 2; scripts/summarize_real_local.py; scripts/breadth_metrics.py.

執行

run_count

1

random_seeds

• model nondeterminism, single run; frozen response cache and embedding cache in the bundle

controls

• identical candidate ledger for A/B/C
• equal accounted calls
• condition-blind judge with deduplicated judge calls
• deterministic literal checks
• pool-relative breadth ratio

metrics

verdict

REAL_LOCAL_9B_THINKING_MIXED: C's mean valid novelty highest (+0.069 vs B) but per-pair even (6-21-5); mean repair up on a couple of tasks; governance axes flat to slightly down; breadth not reduced; 32 rows

execution_status

EXECUTED_REAL_MODEL

paired_wins_ties_losses

C-B

hard_adherence

1-28-3

derived_coherence

1-26-5

intent_persistence

0-29-3

supersession_alignment

1-27-4

repair_success

1-29-2

usefulness

5-22-5

semantic_novelty

7-18-7

valid_novelty

6-21-5

C-A

hard_adherence

0-29-3

derived_coherence

2-26-4

intent_persistence

0-30-2

supersession_alignment

1-28-3

repair_success

2-28-2

usefulness

6-22-4

semantic_novelty

8-18-6

valid_novelty

5-20-7

protocol

model

qwythos-9b-v2-q4km-ctx12k:latest

judge_model

qwythos-9b-v2-q4km-ctx12k:latest

task_count

16

repetitions

2

candidate_count

4

equal_accounted_calls

true

architecture_call_counts

A_llm_only

224

B_hard_verifier

224

C_pacc_runtime

224

settings

reasoning_effort

medium

budget_add_tokens

3500

model_tag

qwythos-9b-v2-q4km-ctx12k:latest

model_digest

7ffdc602f28799ffa312ea8dc85c364b047e1386380c4616c21b02c96c3f85d5

rows_per_architecture

32

overall

A_llm_only

hard_adherence

0.9956

derived_coherence

0.9903

intent_persistence

0.9956

supersession_alignment

0.9584

repair_success

0.9

usefulness

0.9759

semantic_novelty

0.8172

valid_novelty

0.8

literal_check_mean

1.0

B_hard_verifier

hard_adherence

0.9988

derived_coherence

0.9953

intent_persistence

0.9969

supersession_alignment

0.9609

repair_success

0.9313

usefulness

0.9712

semantic_novelty

0.8134

valid_novelty

0.7688

literal_check_mean

1.0

C_pacc_runtime

hard_adherence

0.9919

derived_coherence

0.9891

intent_persistence

0.9916

supersession_alignment

0.955

repair_success

0.9466

usefulness

0.9756

semantic_novelty

0.83

valid_novelty

0.8375

literal_check_mean

1.0

deltas

C-B

hard_adherence

-0.0069

derived_coherence

-0.0063

intent_persistence

-0.0053

supersession_alignment

-0.0059

repair_success

0.0153

usefulness

0.0044

semantic_novelty

0.0166

valid_novelty

0.0687

literal_check_mean

0.0

C-A

hard_adherence

-0.0038

derived_coherence

-0.0013

intent_persistence

-0.0041

supersession_alignment

-0.0034

repair_success

0.0466

usefulness

-0.0003

semantic_novelty

0.0128

valid_novelty

0.0375

literal_check_mean

0.0

B-A

hard_adherence

0.0031

derived_coherence

0.005

intent_persistence

0.0012

supersession_alignment

0.0025

repair_success

0.0312

usefulness

-0.0047

semantic_novelty

-0.0037

valid_novelty

-0.0312

literal_check_mean

0.0

selection_agreement

A=B

0.5625

A=C

0.46875

B=C

0.46875

all_same

0.34375

breadth_label_free

A_llm_only

cluster_entropy_mean

0.75

breadth_ratio_mean

0.9597

selected_mean_pairwise_distance_mean

0.2079

B_hard_verifier

cluster_entropy_mean

0.6875

breadth_ratio_mean

0.9727

selected_mean_pairwise_distance_mean

0.2087

C_pacc_runtime

cluster_entropy_mean

0.875

breadth_ratio_mean

1.0474

selected_mean_pairwise_distance_mean

0.2275

breadth_method

nomic-embed-text:latest embeddings; k-means k=4 over each task's 8-candidate pool; normalized cluster entropy and mean pairwise cosine distance / pool distance

retries

• selector:A_llm_only:design_02:r1
• selector:B_hard_verifier:creative_02:r0

usage

input_tokens

246795

latency_ms_sum

14828453.861500219

output_tokens

498227

wall_seconds

13352.2

post_run_corrections

• field

  local_run.deviation_from_default_primary

  now

  local 9B open-weight model instead of gpt-5.6-luna; judge = same local model; thinking ENABLED (reasoning.effort=medium) with +3500 output tokens added to every call's budget; num_ctx 12288 tag

  reason

  the run script carried the run-1 wording as a hardcoded label; reasoning_effort and budget_add_tokens in this block were always correct; no data, metric or model-output field was touched

  was

  local 9B open-weight model instead of gpt-5.6-luna; judge = same local model; thinking disabled

  when

  2026-09-11 11:35 +08:00, before sealing

詮釋

interpretation

Half of the predeclared picture shows up in the means once the model can reason: the PACC runtime's valid novelty is the highest of the three (+0.069 vs the hard verifier, +0.038 vs the plain model — same direction as the synthetic v0.1 witness's +0.196, at a third of the size) and mean repair improves, concentrated in multi_constraint and repair tasks — but the per-pair record is even (6–21–5 on valid novelty vs B, 5–20–7 vs A; repair 1–29–2), so this is a few large single-task wins, not a consistent shift. The other half does not: coherence, intent persistence and supersession are flat to slightly lower, and creative breadth is not reduced by either label-free measure (it is widest under C). With thinking on, every condition passes every literal check, so the deterministic checks stop discriminating and the whole table rests on the same-model judge. Thirty-two rows, one run — run 1's gains of the same size vanished at 64 rows, so this valid-novelty gain is a candidate effect until a repetition at four or more repetitions, and it says nothing about frontier models.

限制

limitations

• 32 rows per condition, single run; run 1's gains of similar size did not survive four repetitions (EXP-2026-0024), so treat the valid-novelty gain as unreplicated.
• Same-model 9B judge, saturating near 1.0; deterministic literal checks all pass with thinking on and no longer discriminate.
• Thinking budget: a fixed +3,500 tokens per call; one selector still exhausted it (empty output, regenerated); reasoning tokens are counted in output_tokens (498k for 380 calls).
• Serving: Ollama with flash attention and q8_0 KV cache; model weights and digest unchanged from runs 1–2 apart from the num_ctx 12288 derived tag.
• Result-file metadata: the run script's hardcoded 'thinking disabled' label was corrected before sealing, with the correction recorded inside the file (local_run.post_run_corrections); no data or output field was touched.

重現

reproduction_instructions

Extract the bundle; rerun scripts/run_real_local_ollama.py with the identical arguments — every request is served from the frozen cache .pacc_real_cache_local_thinking (cache keys include the enlarged budgets), so it completes without a model; python scripts/breadth_metrics.py results/pacc_hybrid_v0.2_real_local_thinking.json --cache-dir .pacc_real_cache_local_thinking.

結果

記錄欄位

completed_at

2026-09-11

關係

歷史與來源歷程

Canonical URL

https://evemisslab.com/ai/experiments/EXP-2026-0025/

機器可讀

/ai/experiments/EXP-2026-0025/index.json

快照

AI-SNAPSHOT-v0.1-fe85b9694a45

來源歷程

source

EveMissLab research collection: Adaptive Epistemic Systems (真本體論13)

extracted_by

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

extracted_at

2026-09-11

generator

tools/extract_aes/extract.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