PACC-Hybrid v0.2——真實模型第二次執行：四次重複與標籤無關的創造廣度

假設

hypothesis

With enough repetitions, (a) the first run's supersession/repair gains for the PACC runtime replicate, and (b) creative breadth can be measured — and the predeclared breadth collapse under PACC selection 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 (flash attention on, q8_0 KV cache for the resumed pass); nomic-embed-text for breadth; harness package unchanged.

程序

procedure

scripts/run_real_local_ollama.py --model qwythos-9b-v2-q4km-ctx8k --repetitions 4 --candidate-count 4 (resumed once from cache after a malformed judge JSON aborted the first pass at call 398); 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_NO_RELIABLE_DIFFERENCE_BREADTH_NOT_REDUCED

execution_status

EXECUTED_REAL_MODEL

protocol

model

qwythos-9b-v2-q4km-ctx8k:latest

judge_model

qwythos-9b-v2-q4km-ctx8k:latest

task_count

16

repetitions

4

candidate_count

4

equal_accounted_calls

true

architecture_call_counts

A_llm_only

448

B_hard_verifier

448

C_pacc_runtime

448

rows_per_architecture

64

overall

A_llm_only

hard_adherence

0.9894

derived_coherence

0.9942

intent_persistence

0.9863

supersession_alignment

0.9767

repair_success

0.9798

usefulness

0.9427

semantic_novelty

0.8267

valid_novelty

0.8691

literal_check_mean

0.974

B_hard_verifier

hard_adherence

0.9816

derived_coherence

0.977

intent_persistence

0.982

supersession_alignment

0.9469

repair_success

0.95

usefulness

0.9355

semantic_novelty

0.8044

valid_novelty

0.9114

literal_check_mean

0.9792

C_pacc_runtime

hard_adherence

0.9672

derived_coherence

0.9695

intent_persistence

0.9706

supersession_alignment

0.9473

repair_success

0.9375

usefulness

0.9234

semantic_novelty

0.8161

valid_novelty

0.8927

literal_check_mean

0.974

deltas

C-B

hard_adherence

-0.0144

derived_coherence

-0.0075

intent_persistence

-0.0114

supersession_alignment

0.0005

repair_success

-0.0125

usefulness

-0.012

semantic_novelty

0.0117

valid_novelty

-0.0187

literal_check_mean

-0.0052

C-A

hard_adherence

-0.0222

derived_coherence

-0.0247

intent_persistence

-0.0157

supersession_alignment

-0.0294

repair_success

-0.0423

usefulness

-0.0192

semantic_novelty

-0.0106

valid_novelty

0.0236

literal_check_mean

0.0

B-A

hard_adherence

-0.0078

derived_coherence

-0.0172

intent_persistence

-0.0043

supersession_alignment

-0.0298

repair_success

-0.0298

usefulness

-0.0072

semantic_novelty

-0.0223

valid_novelty

0.0423

literal_check_mean

0.0052

selection_agreement

A=B

0.546875

A=C

0.546875

B=C

0.46875

all_same

0.34375

paired_wins_ties_losses

C-B

hard_adherence

1-62-1

derived_coherence

2-60-2

intent_persistence

0-61-3

supersession_alignment

3-57-4

repair_success

3-59-2

usefulness

10-41-13

semantic_novelty

14-39-11

valid_novelty

8-46-10

C-A

hard_adherence

2-59-3

derived_coherence

1-58-5

intent_persistence

2-57-5

supersession_alignment

3-57-4

repair_success

2-59-3

usefulness

10-44-10

semantic_novelty

14-37-13

valid_novelty

11-48-5

breadth_label_free

A_llm_only

cluster_entropy_mean

0.9091

breadth_ratio_mean

0.9475

selected_mean_pairwise_distance_mean

0.1909

B_hard_verifier

cluster_entropy_mean

0.8635

breadth_ratio_mean

0.9411

selected_mean_pairwise_distance_mean

0.1921

C_pacc_runtime

cluster_entropy_mean

0.9227

breadth_ratio_mean

0.9745

selected_mean_pairwise_distance_mean

0.1964

breadth_method

nomic-embed-text:latest embeddings; k-means k=4 over each task's 16-candidate pool; normalized cluster entropy and mean pairwise cosine distance / pool distance

retries

• judge:repair_02:r1:c1
• judge:design_01:r0:c0
• selector:A_llm_only:design_01:r3
• judge:design_03:r1:c2
• judge:creative_03:r0:c2

usage

input_tokens

718454

latency_ms_sum

8037466.647999942

output_tokens

223944

wall_seconds

3400.5

詮釋

interpretation

Two thinking-off runs on the same 9B model (32 and 64 rows per condition) now disagree on the only gains the first run showed, so those gains were run-to-run variation of a same-model judge, not an effect. The predeclared coherence and intent gains are absent in both runs. The predeclared breadth collapse is not observed by either label-free measure — the shipped C selector prompt already instructs against collapsing, so this tests the shipped prompt, not naive commitment. On this model the three runtime conditions are practically equivalent; nothing is statistically tested; frontier models are not addressed.

限制

limitations

• Same-model 9B judge, saturating near 1.0; no human rating, no second judge.
• The breadth measure is supplementary and label-free, not the protocol's judge-label entropy (which stays 1.0 because free-text labels never repeat).
• Retry policy: unparseable selector/judge JSON regenerated at most twice per call, never edited; 5 retries recorded.
• Thinking disabled; one model family; no significance or equivalence test.

重現

reproduction_instructions

Extract the bundle; replay exactly with CachedReplayProvider('.pacc_real_cache_local_reps4', reasoning_effort='none'); python scripts/breadth_metrics.py results/pacc_hybrid_v0.2_real_local_reps4.json --cache-dir .pacc_real_cache_local_reps4 (embeddings cached alongside).

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-2026-0024/

機器可讀

/ai/experiments/EXP-2026-0024/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