ResearchRES-2026-0004v0.1
A PACC-style runtime on language models: reasoning, intent and creative breadth
Does wrapping candidate selection in a canonical hard/derived-constraint commit space change what a generator produces? A synthetic A/B/C witness (generator only / hard verifier / PACC runtime) shows derived coherence and constraint-satisfying novelty up, soft-preference fit slightly down, and a creative-breadth collapse that a post-hoc 'elastic' exploration policy recovers. The real-language-model version of the benchmark has a validated, fail-closed harness but has not been executed.
Research questions
research_questions- Is there a reasoning-up / imagination-down trade-off, or is the observed breadth loss a selection-policy artefact separable from the commit constraints?
- Will a frontier language model show the same effect as the synthetic witness?
Claims
claims- Within the synthetic witness: reasoning coherence ↑, valid novelty ↑, soft-preference fit slightly ↓, creative breadth ↓ under naive selection and recoverable with exploration separated from commitment (8/8 seeds sign-stable).
- Real local 9B model, thinking off, two runs (32 and 64 rows per condition): the three runtime conditions are practically equivalent on every judge axis, and by label-free measures the PACC runtime does not narrow creative breadth. The synthetic v0.1 prediction did not appear on this model.
- Real local 9B model with thinking enabled (32 rows per condition): the PACC runtime has the highest valid novelty (+0.069 vs the hard verifier, +0.038 vs the plain model) and repair, governance axes are flat to slightly lower, and breadth is again not reduced — half of the synthetic prediction, unreplicated.
Limitations
limitations- Zero real LLM calls so far; no claim about real-model reasoning, intent understanding, imagination, hallucination or human-rated usefulness is permitted before a real-model result exists.
- The v0.1 numbers are synthetic data and theoretical reasoning; until a real hybrid model exists, an inference is only an inference — theoretically possible is not actually possible.
Relations
| Source | Relation | Target | Status | ID |
|---|---|---|---|---|
RES-2026-0004 A PACC-style runtime on language models: reasoning, intent and creative breadth | belongs_to | PRG-2026-0001 Adaptive Epistemic Systems | ACTIVE | REL-2026-0004 |
RES-2026-0004 A PACC-style runtime on language models: reasoning, intent and creative breadth | extends | RES-2026-0002 PACC conjecture — do non-probabilistic primitives converge to probability-like structure? | ACTIVE | REL-2026-0006 |
SYS-2026-0003 PACC-LLM Hybrid Lab | supports | RES-2026-0004 A PACC-style runtime on language models: reasoning, intent and creative breadth | ACTIVE | REL-2026-0087 |
DAT-2026-0002 PACC-Hybrid v0.1 shared candidate pools | supports | RES-2026-0004 A PACC-style runtime on language models: reasoning, intent and creative breadth | ACTIVE | REL-2026-0093 |
History and provenance
- Canonical URL
- https://evemisslab.com/ai/research/RES-2026-0004/
- Machine-readable
/ai/research/RES-2026-0004/index.json- Snapshot
AI-SNAPSHOT-v0.1-fe85b9694a45- Provenance
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