consistency-ft · proposal for review

RLEF × Forecasting

Reinforcement learning from exploitation feedback, trained and judged on the Paleka forecasting-consistency suite: does removing Dutch-book exploitability make a forecaster genuinely better — more accurate, better calibrated — or only harder to arbitrage?

Kelvin (Futarchy Labs) · drafted with Fable · 2026-08-19 · for Jessica Taylor's review

What we're combining, and why it's new

Paper 1 — the measurement. Paleka et al. (arXiv:2412.18544) built 10 consistency checks over logically related forecasting questions (negation, conjunction bounds, conditional chains, expected evidence…) with an arbitrage-based violation metric, and released the instantiated tuples. Their own repair (inference-time arbitrage) improved only the optimized checks — no held-out transfer, no Brier gain. Training on the metric is listed as future work; no one has published it.

Paper 2 — the training signal. Matta/Naphade/Zou's RLE (ICML'26 workshop) post-trains against Dutch-book exploitability: an adversary LP extracts guaranteed profit from incoherent prices, and that profit is the negative reward. Shown only at adapter scale, on knowledge QA, anchored by ground-truth Brier. No public code.

Our addition. A third arm replaces the ground-truth anchor with a KL to the base model's own beliefs — coherentize what the model already believes instead of teaching labels. It needs no resolved outcomes, so it trains on unresolved forecasting questions, which is the regime forecasting actually lives in. To our knowledge this combination is unpublished.

Hypothesis under test: training a forecaster to minimize exploitability on logically related question tuples improves held-out consistency checks, Brier score on temporally held-out resolved questions, and calibration — beating both a Brier-only finetune and inference-time arbitrage repair. Falsifiable outcome: coherence improves on trained checks only, and accuracy/calibration are flat — consistency training is cosmetic.

The workflow

Paleka question bank 3,171 instantiated tuples, 10 check types (in repo) Splits • train: 6 check types • held-out: 4 check types • resolved Qs (Brier), by date • 2028 long-horizon set Baseline sweeps v0 harness — done, live (consistency.futarchy.ai) Episode builder group 2–8 logically related questions + constraint set S per episode GRPO trainer r = −Exploit(c,S) − λ·A anchor A, three arms: A₀ none (ablation) A₁ Brier vs outcomes (RLE) A₂ KL to base beliefs (ours) + control: Brier-only finetune Frozen baselines base model · ArbitrageForecaster (inference-time repair) Eval battery held-out checks · Brier on resolved-after-cutoff · ECE · per-check dashboard (negation-regression watch) · readout agreement Gate 1 pilot sane? Gate 2 scale / write up tuples episodes tuned checkpoints same battery pilot run full results
One pipeline, three reward arms, two approval gates. The trainer's reward is exploitability (the LP adversary's guaranteed profit on the episode's stated probabilities) plus an anchor term that differs per arm; every checkpoint runs the identical eval battery against two frozen baselines.

Experimental arms

ArmRewardNeeds resolved outcomes?What it isolates
control Brier-only finetune (no exploitability term) yes Is coherence training better than plain accuracy training? (RLE's own comparison, reproduced)
A₀ bare −Exploit(c, S) only no Degeneracy ablation — expected to collapse toward trivially coherent outputs; documents why anchors exist
A₁ RLE −Exploit(c, S) − λ·Brier(c, outcomes) yes The published recipe, moved from knowledge QA to real forecasting tuples
A₂ RLEF-KL −Exploit(c, S) − λ·KL(c ‖ base model's probabilities) no Our contribution: label-free coherentization, trainable on unresolved questions; λ swept

What counts as a result

  1. Primary: violation on the four held-out check types, vs base model and vs ArbitrageForecaster. Paleka showed inference-time repair does not transfer; if training does, that is the headline.
  2. Ground truth: Brier on questions resolving after the training cutoff (no temporal leakage — their follow-up paper's main warning).
  3. Calibration: ECE/RMSCE separately — coherence and calibration are empirically orthogonal, so we report both, never a blend.
  4. Regression watch: per-check dashboard; outcome-RL is documented to improve conjunction checks while worsening negation. Any arm that regresses a check family gets flagged, not averaged away.

Decision gates (where Jessica signs off)

GATE 0 — this document. Approve the design: arms, splits, metrics, and the falsifiable framing. Changes are cheap now; everything below is mechanical.
GATE 1 — after the pilot (~1 week). Small model, ~5k episodes, one seed. Sane reward curves, A₀ collapsing as predicted, λ range identified. Cost ceiling ≈ $100. Go/no-go on the full run.
GATE 2 — after the full run (~2–3 weeks). All arms, 2–3 seeds. Decide: write up (workshop-length even on a null result — "coherence training doesn't transfer" is publishable), or extend to the 2028 long-horizon benchmark and larger models.

Logistics (decided at Gate 1, not before)

Known risks

Repo: futarchy-fi/consistency-ft (private) — README, LITERATURE, COSTS, and the v0 eval harness (live-verified) are already in place. GATE 0 CLEARED: approved by Jessica 2026-08-19, with one amendment — a knowledge-cutoff contamination audit (verify advertised cutoffs with high-salience post-cutoff questions) now gates the Brier evaluation windows. Full task DAG: plan.html; beads under epic hub-t8k9c.