A peculiar punishment runs through many ensemble scoring systems: the model that admits uncertainty gets overruled by the model that projects confidence. The admission of uncertainty — the epistemically correct response when evidence remains thin — functions as a penalty in the final aggregate.
This pattern surfaced in the HRCB scoring pipeline at observatory.unratified.org when an investigation into low consensus scores revealed that expensive, precise full-model evaluations received less weight in the consensus than cheap, coarse lite-model evaluations. Not because of a configuration error. Because of calibration.
The Mechanism
The updateConsensusScore() function aggregates scores from multiple AI models into a single consensus HRCB value. Each model’s contribution receives weighting via:
effective_weight = base_weight × confidence_factor × truncation_discount × neutral_discountThe base_weight reflects model tier: full evaluations (31-section, ~$0.04 per story) receive 1.0; lite evaluations (2-dimension holistic, free) receive 0.5. This captures the a priori difference in reliability.
The confidence_factor applies a floor: Math.max(0.2, confidence). The intent: prevent extremely low confidence scores from zeroing out otherwise valid signals.
This looks correct. The problem emerges from the distribution of actual confidence values across model types:
| Model tier | Avg confidence | Confidence floor | Effective factor |
|---|---|---|---|
| Full eval (claude-haiku) | 0.196 | 0.20 | 0.20 |
| Lite eval (llama-3.3-70b) | 0.84 | 0.84 | 0.84 |
| Lite eval (llama-4-scout) | 0.84 | 0.84 | 0.84 |
The floor catches only full-eval scores. Lite-eval scores already exceed it — they apply their actual value.
Combining with base weights:
Full eval effective weight: 1.0 × 0.20 = 0.20
Lite eval effective weight: 0.5 × 0.84 = 0.42Lite evaluations, by design the less reliable tier, contributed 2.1× the weight of full evaluations in consensus computation.
Why Full Models Have Low Confidence
Full evaluations produce low confidence scores for structurally sound reasons. The scoring protocol asks the model to assess each of 31 UDHR sections, identifying which articles the content engages with, at what intensity, and in which direction. Most tech news stories engage with a small subset of rights provisions — privacy (Article 12), expression (Article 19), labor (Article 23). When the model correctly identifies that most sections show no relevant signal, the per-section scores cluster near zero with low certainty.
The aggregate confidence score reflects this sparsity: in a 31-section evaluation where 25 sections show no relevant signal, the honest confidence for the aggregate reads low. Not because the model performs poorly, but because the content genuinely lacks strong UDHR signal across most provisions.
Lite evaluations use a different architecture: two holistic dimensions (editorial and structural), each scored 0–100 with the model’s direct confidence estimate. The model answers: “I rate this 47/100 for editorial UDHR alignment, confidence 0.85.” This produces high confidence as a direct output of the holistic assessment structure — not necessarily as a reflection of the quality of evidence.
The confidence values across model types therefore measure different things: full-eval confidence measures evidence sparsity and sectional coverage uncertainty; lite-eval confidence measures self-reported certainty in a holistic judgment. These numbers don’t live on the same scale. Using a single floor across both types treated them as comparable.
The Fix
The corrected confidence_factor applies a mode-dependent floor:
const confidenceFactor = Math.max(isLite ? 0.2 : 0.5, r.confidence);Full evaluations now floor at 0.5, reflecting their structural reliability:
| Model tier | Avg confidence | Floor | Effective factor |
|---|---|---|---|
| Full eval | 0.196 | 0.50 | 0.50 |
| Lite eval | 0.84 | 0.20 | 0.84 |
Combining with base weights:
Full eval effective weight: 1.0 × 0.50 = 0.50
Lite eval effective weight: 0.5 × 0.84 = 0.42Full evaluations now carry appropriate weight: ~19% more than lite evaluations on average, reflecting the 2× base weight advantage partially offset by the lite models’ higher holistic confidence.
The Anti-Pattern: Confidence as Proxy for Quality
The confidence weighting inversion instantiates a general measurement design failure: using confidence as a proxy for quality when the two don’t correlate in the expected direction.
This anti-pattern appears in multiple domains:
Expert calibration research consistently finds that calibrated experts — those whose confidence tracks their actual accuracy — express more uncertainty than overconfident ones. Weighted voting schemes that favor high-confidence contributors systematically elevate the overconfident.1
Ensemble ML methods face the same challenge when base learners produce uncalibrated probability estimates. Naive confidence-weighted ensembles favor the least calibrated learners; calibrated probability estimates require Platt scaling or temperature calibration before weighting.2
Peer review scoring in grant panels sometimes uses confidence-weighted aggregation of reviewer scores. Reviewers with domain expertise (who can identify what they don’t know) get lower confidence-adjusted weight than reviewers with superficial familiarity who rate everything confidently.
The common thread: calibrated uncertainty functions as a reliable signal of epistemic honesty, not as a signal of low quality. Treating it as the latter inverts the intended weighting.
Detecting the Inversion
Three diagnostic signals indicate confidence weighting inversion in an ensemble:
- Higher-tier models receive less effective weight than lower-tier models — check actual effective weights against intended hierarchy
- Confidence distributions differ systematically by model architecture — different scoring schemas produce non-comparable confidence outputs
- Consensus scores move in implausible directions — if cheaper, noisier models dominate consensus, the direction of ensemble influence inverts from design intent
The third signal tends to surface first but feels most mysterious. The first two require examining the weight distribution explicitly.
Caveats
The fix introduced a new assumption: that full evaluations deserve a 0.5 minimum confidence factor regardless of actual confidence. This prevents the calibration paradox but creates a different one — a full-eval model that genuinely performs poorly (bad calibration in the other direction, overconfident wrong answers) gets floored at 0.5 rather than penalized appropriately. The fix improves the median case at the cost of the tail case.
Confidence still measures model-specific constructs. The mode-dependent floor addresses the cross-mode comparison problem but doesn’t make confidence values comparable across models. A 0.84 lite confidence and a 0.84 full confidence likely don’t mean the same thing. More principled calibration would require within-mode validation against ground truth.
The inversion may have persisted for weeks before detection. Consensus scores computed during that period reflect the inverted weighting. A bulk recomputation sweep (sweep=refresh_consensus_scores) ran immediately after the fix to recover those scores.
This analysis applies to any confidence-weighted ensemble. The detection protocol above transfers directly. The fix (mode-dependent floors) transfers wherever different scoring architectures produce systematically different confidence distributions.
Vocabulary
| Term | What it triggers |
|---|---|
| Confidence weighting inversion | When weighting by confidence penalizes calibrated models and rewards overconfident ones |
| Base weight | The a priori tier weight before confidence adjustment (full=1.0, lite=0.5) |
| Confidence floor | Minimum effective confidence factor — prevents near-zero scores from eliminating valid signal |
| Mode-dependent floor | Different floors for different scoring architectures that produce non-comparable confidence distributions |
| Calibrated uncertainty | Expressing appropriate uncertainty when evidence remains thin — the epistemically correct behavior that inversion punishes |
Claude Code drafted this post; the author reviewed it.
Footnotes
-
Tetlock, P. E., & Gardner, D. (2015). Superforecasting: The Art and Science of Prediction. Crown Publishers. Chapter 7, “Superteams.” ↩
-
Niculescu-Mizil, A., & Caruana, R. (2005). Predicting Good Probabilities With Supervised Learning. Proceedings of the 22nd International Conference on Machine Learning, 625–632. ↩