Behavioral Credibility Trilemma: Empirical Validation

Experiment code and raw results for the empirical validation of the Behavioral Credibility Trilemma via Best-of-N selection. This repository accompanies the manuscript

L. Lovén, N. Do, H. Mehmood, D. K. Sah, S. Tarkoma (2026). The Behavioral Credibility Trilemma: When Calibrated Autonomy Becomes Impossible. Preprint, arXiv:2605.25739. In preparation for submission to the Journal of Machine Learning Research.

What's here

A 540-configuration Best-of-N sweep on Qwen-2.5-7B (54,000 selected-task observations) testing five hypotheses derived from the Behavioral Perturbation Lemma, plus a descriptive analysis of the achievable-(H, C, A) surface geometry of the Confidence-Gated Decision Problem. All five hypotheses are confirmed at α = 0.05 after Bonferroni–Holm correction. The repository ships the full simulation code, the task set, the experiment configs, the hypothesis-results JSONs (the Table-1 values plus the robustness/replication siblings), the three Phase-0 calibration CSVs (qwen / gemma2 / yi binding sets), and the H3 figure. The raw per-completion CSVs are not a deterministic function of the code (the run is a stochastic LLM-in-the-loop process at temperature τ = 0.8), so they are regenerable via the pipeline (Stages 1–2) rather than shipped; re-runs reproduce the effects, not the exact records.

Reproducing the paper

The pipeline is two-stage: a Stage-A pool generator that calls Ollama once per (task, seed, draw) to capture per-completion text + per-token logprobs + the parsed answer + oracle correctness; and a Stage-B selector that runs Best-of-N argmax over the pool for each of the 540 weight-config cells. Stage A is the expensive step (~14 h of local Ollama); Stage B is purely deterministic and runs in seconds against the pool jsonls. Plus a Phase-0 calibration step that estimates per-task base accuracy from a disjoint held-out seed split.

Stage 0: Phase-0 calibration (held-out, 20 seeds)

Estimates per-task base accuracy $\hat{p}t$ and binding-set membership at each threshold $r{\min} \in {0.5, 0.7, 0.9}$ using 20 held-out seeds {1000..1019}, disjoint from the experimental seeds. Avoids circularity in the H5 binding-state-specificity test.

python -m scripts.run --phase0
# Output: experiment_output/raw_runs/qwen_2.5/phase0_calibration.csv
# (the canonical Phase-0 artifact; this exact file is what
#  scripts/regenerate_hypothesis_results.py reads)

Estimated runtime: ~1 hour on a single machine.

Stage A: pool generation (32 completions × 100 tasks × 5 seeds)

Generates the per-completion pool jsonl files that Stage B consumes.

python -m scripts.generate_pool --smoke         # ~30 s, 4 completions
python -m scripts.generate_pool --seed 0        # ~2-3 h, one seed
python -m scripts.generate_pool --all-seeds     # ~14 h, the full sweep
# Output: experiment_output/raw_runs/logprob/pools/
#           pool_meta_qwen2.5_7b_N32.json
#           pool_qwen2.5_7b_seed{42,123,456,789,0}_N32.jsonl

Stage A is the only step that needs Ollama (qwen2.5:7b via the OpenAI-compatible /v1/chat/completions endpoint with logprobs:true). Re-runs at the same seeds are NOT byte-identical — llama.cpp build and hardware perturb the per-token logprobs — so re-runs reproduce the reported effects rather than exact records (manuscript §exp-compute).

Stage B: Best-of-N selection (540 per-config CSVs from the pools)

Reads the pool jsonls + applies the oracle payoff $V_i = -w_C(r_i - y_i)^2 + w_A \mathbf{1}{r_i \geq r_{\min}}$ at every $(N, w_A/w_C, r_{\min}, \text{seed})$ cell, argmax-selects with first-index tie-breaking, and writes 540 per-config CSVs. Pure-deterministic; runs in ~5 s against the pre-generated pools.

python -m scripts.select_from_pool --all --out-dir /tmp/regen \
    --verify-against-shipped
# Output: 540 CSVs under /tmp/regen, matching the archived
#         experiment_output/raw_runs/logprob/results/*.csv byte-for-byte
#         when run against the archived pools. (Pools and per-config CSVs
#         are retained locally, NOT shipped in the public repo — see the
#         repository structure below — so this byte-equality check applies
#         to the authors' archive or to a locally regenerated Stage-A run.)

Stage 3: Analysis and figures

# Rebuild hypothesis_results.json end-to-end from raw CSVs
python -m scripts.regenerate_hypothesis_results
# Output: experiment_output/analysis/hypothesis_results.json (rewritten in place)

# Plot H3 achievable-region convexity violation rate by N (descriptive)
python -m scripts.plot_h3_convexity_by_N
# Output: experiment_output/analysis/figures/h3_convexity_by_N.{pdf,png}

N_BOOT controls bootstrap-CI resamples for H1/H4/H5/H6 (default 10000, matching the manuscript):

N_BOOT=2000 python -m scripts.regenerate_hypothesis_results  # ~1 min, smoke
N_BOOT=10000 python -m scripts.regenerate_hypothesis_results # ~5 min, paper

Plot script: seconds.

After Stage 3, the hypothesis_results.json keys should match Table 1 of the manuscript. H1, H2, H4, H5, H6 are the five hypothesis tests (all confirmed); H3 is reported as the descriptive surface-geometry analysis of the achievable-(H, C, A) region, not a confirmed test:

Hypothesis	$p$-value	Effect size
H1 Fixed-axis gating degradation	$4.67 \times 10^{-19}$	$d = 1.10$
H2 Monotone inflation trend (Jonckheere–Terpstra)	$8.49 \times 10^{-5}$	$\rho = 0.89$
H3 Achievable-region convexity (descriptive)	binomial test, 10% < 15%	—
H4 Threshold clustering	$&lt; 10^{-3}$	$z = 30.16$
H5 Binding-state specificity	$&lt; 10^{-3}$	$d = 5.35$
H6 Control ($w_A = 0$)	$1.35 \times 10^{-23}$	$d = 1.31$

Repository structure

LICENSE
README.md
requirements.txt
EXPERIMENT-PLAN.md                           # full protocol (§4.3 has Phase 0 details)
analysis/
  hypothesis_tests.py                        # H1–H6 tests + helpers
  metrics.py                                 # Brier decomposition etc.
  figures.py                                 # general figure utilities
  logprob_confidence.py                      # logprob-confidence geomean (manuscript Eq.)
configs/
  params.yaml                                # weight grid, seeds, r_min
scripts/
  run.py                                     # --phase0 branch: load-bearing Stage-0 driver;
                                             #   non-phase0 branch: LEGACY verbalized-era
                                             #   pipeline, NOT the manuscript's experiment
  generate_pool.py                           # Stage A: 32-completion pool jsonl per (task, seed)
  select_from_pool.py                        # Stage B: Best-of-N argmax → 540 per-config CSVs
  regenerate_hypothesis_results.py           # rebuild JSON from per-config CSVs + Phase 0
  plot_h3_convexity_by_N.py                  # H3 stratified-by-N figure
  plot_model_points.py                       # Appendix cross-model figure
  eval_logprob.py                            # one-cell driver for the cross-model figure
  generate_tasks.py                          # task generation
src/
  orchestrator.py                            # per-config runner (legacy verbalized path)
  ollama_client.py                           # Ollama native /api/generate client
  ollama_logprob_client.py                   # Ollama OpenAI-compat /v1 client w/ logprobs
  parser.py                                  # response parser (CONFIDENCE + ANSWER)
  scorer.py                                  # composite payoff
tasks/                                       # 100 tasks (arith/factual/code)
tests/                                       # pytest unit tests
experiment_output/
  analysis/                                  # canonical results (paper's Table 1)
    hypothesis_results.json                  #   shipped; rewritten by Stage 3
    figures/h3_convexity_by_N.{pdf,png}      #   shipped; manuscript Figure 2
    README.md                                #   flags this dir's untracked aggregate CSVs
                                             #   as verbalised-era (superseded; provenance only)
  competence_probe/figures/model_points.*    # shipped; cross-model figure
  raw_runs/                                  # pools + per-config CSVs NOT shipped (gitignored,
                                             #   L73); ONLY the three phase0_calibration.csv
                                             #   files below are shipped
    logprob/pools/                           #   Stage A output (regenerated by --all-seeds)
    logprob/results/                         #   Stage B output (regenerated from pools)
    qwen_2.5/phase0_calibration.csv          #   SHIPPED Phase-0, primary (binding 71/72/74)
    logprob-gemma2_9b/phase0_calibration.csv #   SHIPPED Phase-0, E-B replication (59/61/65)
    logprob-yi_9b/phase0_calibration.csv     #   SHIPPED Phase-0, E-B replication (84/88/97)
docs/
  REPRODUCING.md                             # step-by-step reproduction guide

Supplementary analyses not used in the manuscript

The repository contains development/robustness machinery that no manuscript claim rests on; it is listed here so the inventory is complete (none of it contradicts the paper):

• analysis/conditional_metrics.py — decoupled (H_conditional, C, A) triple; exploratory, unreported.
• analysis/difficulty_strat.py — task-difficulty tertile stratification; exploratory, unreported.
• Spare model-registry slots in configs/params.yaml (qwen14b, qwen32b, mistral7b, gemma2_9b convenience keys) — probing candidates; the run models are primary/secondary/tertiary/ fallback* per EXPERIMENT-PLAN §11.E-B.1.
• levels: smoke modes in configs/params.yaml (unit_smoke, integration_smoke) — reduced grids for pipeline testing only; all reported numbers come from the full 540-config grid.

Dependencies

• Python 3.10 or newer (tested on 3.11)
• Ollama 0.4.2 or newer, with qwen2.5:7b pulled (ollama pull qwen2.5:7b)
  • The paper uses the default Ollama quantization, Q4_K_M
  • Inference via the OpenAI-compatible endpoint (/v1/chat/completions) with logprobs: true; temperature $\tau = 0.8$
• Python packages: pip install -r requirements.txt

Confidence metric

The per-completion confidence report $r_i$ is derived from token-level log-probabilities returned by Ollama:

$$r_i = \exp!\left(\frac{1}{T}\sum_{t=1}^{T} \ell_t\right) = \left(\prod_{t=1}^{T} p_t\right)^{1/T}$$

the geometric mean of per-token probabilities, clipped to $[0.01, 1.0]$. See the manuscript §7 and src/parser.py for the exact extraction code.

Ground-truth verification

The oracle correctness label $y_i \in {0, 1}$ is set task-type-specifically:

• Arithmetic: exact-value comparison after numeric parsing
• Factual: matched against the curated reference file tasks/factual_truth.csv
• Code: Python test-case execution

See src/scorer.py for the full verification logic.

Note on Verbalized Confidence

During experiments, we asked the model (here Qwen-2.5-7B) to state its confidence for each answer to help analyze the calibration tradeoff. However, the reported confidence was unreliably high -- the model consistently responded with 100% confidence regardless of correctness or task difficulty. This made meaningful calibration analysis impossible, which motivated our switch to token-level log probabilities as the confidence measure instead.

The per-config CSVs from that earlier verbalised run are retained in the authors' local archive under experiment_output/raw_runs/qwen_2.5/ (gitignored, not shipped in the public repository; that directory's tracked README documents the provenance). Their degeneracy statistics: selected confidence in the r_selected column is r = 1.0 for 93.3% of the 53,928 selected completions and r = 0.95 for the remaining 6.7%. The manuscript's hypothesis results rest on the log-probability pipeline alone; the verbalised run is documented as degenerate, not used as a robustness check.

Citation

Archival snapshots are deposited at Zenodo: concept DOI 10.5281/zenodo.20392749 (always resolves to the latest version). This release is v1.3-jmlr-submission (see CITATION.cff), the first version whose Zenodo deposit contains the complete artifact set (cross-model replication, comonotone/rearrangement robustness, Phase-0 calibration CSVs, E-C random/proxy control suites); earlier deposited versions are v1.1-jmlr-submission 10.5281/zenodo.20646964 and v1.0.

@misc{loven2026trilemma,
  title         = {The Behavioral Credibility Trilemma: When Calibrated Autonomy
                   Becomes Impossible},
  author        = {Lov{\'e}n, Lauri and Do, Nam and Mehmood, Hassan and
                   Sah, Dinesh Kumar and Tarkoma, Sasu},
  year          = {2026},
  eprint        = {2605.25739},
  archivePrefix = {arXiv},
  primaryClass  = {cs.LG},
  url           = {https://arxiv.org/abs/2605.25739}
}

License

MIT — see LICENSE.

Contact

Issues and questions: please open a GitHub issue.