ExperimentEXP-2026-0010v0.1

PACC-Lab v0.3 — learned source reliability without an oracle

Both sides lose the source-quality oracle: the Bayesian reference learns Beta-Bernoulli source quality; the non-probabilistic systems learn a qualitative reputation (trust, friction, streak, familiarity). Fitted on training worlds and evaluated on unseen source-quality permutations, the reputation state maps to the Beta-Bernoulli state with JS ≈ 0.00714 versus 0.017–0.018 for shuffled/constant controls, and feedback-update commutation ≈ 0.00130 — stable 6/6 across seeds. Task-state convergence is architecture- and seed-sensitive: at exact primary scale N0/N1 4/4, N2 3/4, N3 2/4.

Research status: STABLE the current conclusions are relatively stable
Evidence level: E3 Repeated experiment
Result: MIXED
Data basis: SYNTHETIC Synthetic data and theoretical reasoning. Many now treat synthetic data as if it were real; this laboratory says the opposite deliberately — until a real hybrid model exists, an inference is only an inference, and theoretically possible is not actually possible.
Version: 0.1
Updated: 2026-09-09
Created: 2026-09-09
Domain: Model Representation, Evaluation, Formal AI
Program: PRG-2026-0001 Adaptive Epistemic Systems
Authors: Neo.K (EveMissLab)
AI collaborators: Sol (GPT-5.6, OpenAI ChatGPT)

Hypothesis

hypothesis: Convergence survives when source reliability must be learned from delayed feedback rather than given.

Setup

model_ids: MOD-2026-0001 N0 — signed support
MOD-2026-0002 N1 — ordinal tournament
MOD-2026-0003 N2 — signed graph
MOD-2026-0004 N3 — constraint competition
MOD-2026-0005 Bayesian reference (exact posterior / Beta-Bernoulli / joint common-cause / HMM)
dataset_ids: DAT-2026-0001 PACC synthetic evidence worlds
benchmark_ids: BEN-2026-0001 PACC micro-lab protocol v0.1 (frozen gates)
software_environment: Python; deterministic seeded generators; no network, no LLM.

Runs

run_count

random_seeds

20260909
6 secondary seeds (smaller)
4 seeds at exact primary scale

controls

shuffled-target mapping
constant prediction
broken composition (v0.6+)
target-local refit (v0.7+)

metrics

verdict

RELIABILITY-STATE CONVERGENCE ROBUST; TASK CONVERGENCE PARTIAL / BASIN-SENSITIVE

primary_scale

20 worlds / 280 episodes / 28 observations

reliability_mapping_js

0.007136

control_js

0.017–0.018

feedback_commutation_js

0.001298

reliability_independent_families

task_pass_at_primary_scale

N0: 4/4
N1: 4/4
N2: 3/4
N3: 2/4

Interpretation

interpretation: Calibration-state convergence can be robust while task-state convergence has architecture-dependent basins. A first fixed-quality diagnostic was rejected because Beta means became near-constant and shuffled targets fit almost as well — that redesign is part of the evidence.

Limitations

limitations: All four wrappers share one NonProbReputationLedger, so reliability convergence counts as one family, not four.
Does not prove Beta-Bernoulli learning and qualitative reputation universally equivalent.

Reproduction

reproduction_instructions: Extract the version's FINAL bundle; python -m pytest -q; run the version's primary script with the recorded seed; docs/PACC_LAB_v0.N_RESULTS.md and docs/EXPERIMENT_PROTOCOL_v0.N.md are inside the bundle.

Recorded fields

completed_at: 2026-09-09

Relations

Source	Relation	Target	Status	ID
`EXP-2026-0010` PACC-Lab v0.3 — learned source reliability without an oracle	`runs_on`	`SYS-2026-0002` PACC-Lab — micro-lab harness for the convergence conjecture	`ACTIVE`	`REL-2026-0182`
`EXP-2026-0010` PACC-Lab v0.3 — learned source reliability without an oracle	`uses_benchmark`	`BEN-2026-0001` PACC micro-lab protocol v0.1 (frozen gates)	`ACTIVE`	`REL-2026-0183`
`EXP-2026-0010` PACC-Lab v0.3 — learned source reliability without an oracle	`uses_dataset`	`DAT-2026-0001` PACC synthetic evidence worlds	`ACTIVE`	`REL-2026-0184`
`EXP-2026-0010` PACC-Lab v0.3 — learned source reliability without an oracle	`uses_model`	`MOD-2026-0001` N0 — signed support	`ACTIVE`	`REL-2026-0185`
`EXP-2026-0010` PACC-Lab v0.3 — learned source reliability without an oracle	`uses_model`	`MOD-2026-0002` N1 — ordinal tournament	`ACTIVE`	`REL-2026-0186`
`EXP-2026-0010` PACC-Lab v0.3 — learned source reliability without an oracle	`uses_model`	`MOD-2026-0003` N2 — signed graph	`ACTIVE`	`REL-2026-0187`
`EXP-2026-0010` PACC-Lab v0.3 — learned source reliability without an oracle	`uses_model`	`MOD-2026-0004` N3 — constraint competition	`ACTIVE`	`REL-2026-0188`
`EXP-2026-0010` PACC-Lab v0.3 — learned source reliability without an oracle	`uses_model`	`MOD-2026-0005` Bayesian reference (exact posterior / Beta-Bernoulli / joint common-cause / HMM)	`ACTIVE`	`REL-2026-0189`
`EXP-2026-0010` PACC-Lab v0.3 — learned source reliability without an oracle	`tests`	`THY-2026-0005` PACC conjecture — the four-level convergence ladder	`ACTIVE`	`REL-2026-0190`
`EXP-2026-0010` PACC-Lab v0.3 — learned source reliability without an oracle	`extends`	`EXP-2026-0009` PACC-Lab v0.2 — third family (N3) and adversarial evidence geometry	`ACTIVE`	`REL-2026-0191`
`EXP-2026-0010` PACC-Lab v0.3 — learned source reliability without an oracle	`produced`	`ART-2026-0023` PACC-Lab v0.3 Learned Reliability FINAL `artifact://evemisslab/adaptive-epistemic-systems/PACC-Lab_v0.3_Learned_Reliability_FINAL.zip`	`ACTIVE`	`REL-2026-0192`
`EXP-2026-0011` PACC-Lab v0.4 — correlated sources and dependence geometry	`extends`	`EXP-2026-0010` PACC-Lab v0.3 — learned source reliability without an oracle	`ACTIVE`	`REL-2026-0202`

History and provenance

Canonical URL

https://evemisslab.com/ai/experiments/EXP-2026-0010/

Machine-readable

/ai/experiments/EXP-2026-0010/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