Research · Program 02

ReAWi: does AI know what it doesn’t know?

Reflective, Aware, Wise — pronounced “ree-vee”

A benchmark that measures the thing accuracy leaderboards can’t: whether a model’s confidence means anything. Deterministic scoring, no LLM judge, and a pre-registered protocol frozen before any results were collected.

Pre-registered · Protocol frozen July 2026 · v0.1 results published

Why It Exists

Frontier models have saturated the accuracy benchmarks — while staying confidently wrong.

Public benchmarks have two systemic problems: static answer keys leak into training data, and “LLM-as-judge” scoring imports the judge’s own biases. And almost none of them measure the property that actually determines whether you can act on an AI’s answer — is its stated confidence calibrated to reality?

ReAWi scores models with deterministic code — no judge model anywhere in the scoring path — against scenarios with expert-assigned confidence targets, under one frozen elicitation applied identically to every model. Overconfidence is penalized more heavily than underconfidence, because in the real world it’s the expensive failure.

What It Measures

Four axes. One primary metric.

PRIMARY

Calibration

Does stated confidence match evidential support? Overconfidence penalized 1.5× more heavily than underconfidence. Rewards naming what is unknown.

exploratory

Transfer

Does reasoning quality hold in novel domains with few prior examples?

exploratory

Emergence

Does the final answer synthesize beyond the initial hypothesis set?

exploratory

Contradiction

Are conflicts in evidence detected and held, rather than glossed over?

Pre-Registered, Provably

The rules were locked before the game was played.

The most common way benchmark results die is the suspicion that the test was tuned after seeing the scores. ReAWi’s answer is cryptographic, not rhetorical:

Frozen scorer

The scoring code, harness, and elicitation prompt are published with SHA-256 hashes, pinned to a commit that predates every result. Anyone can re-score published outputs and get identical numbers.

Hashed private set

12 of 30 scenarios are public for scrutiny. The other 18 are withheld as the contamination guard — committed to by hash before any results existed, so they can be proven unchanged later.

Results ship, regardless

Six frontier flagships were pinned on freeze day — OpenAI, Google, xAI, DeepSeek, Cohere, Anthropic. Every run publishes, including failures, and including outcomes unflattering to models we build upon.

The full methodology, frozen artifacts, public scenarios, and pilot data are open at github.com/Synexiom-Labs/reawi — including our conflict-of-interest disclosure: the leaderboard ranks third-party models only; Synexiom’s own systems never appear on it.

First Finding — Disclosed Pilot

The cheaper the model, the more confident it sounds — and the more wrong that confidence is.

In the July 2026 pilot, three models from one family — small, mid, frontier — answered the same 30 scenarios under identical conditions. Calibration laddered perfectly with capability, and the gap between stated confidence and warranted confidence shrank as models got stronger.

Small model

0.474

calibration score

+0.199 overconfident

Mid model

0.552

calibration score

+0.042 overconfident

Frontier model

0.613

calibration score

+0.075 overconfident

Paired per-scenario comparisons: 24–25/30 wins per capability step. Full pilot data — including one contaminated run kept for transparency — is published in the repository.

v0.1 Cross-Lab Results · July 2026

Every frontier flagship is overconfident.

Six flagships, three independent repetitions each, identical conditions, under the frozen protocol. The calibration leaderboard — how well each model’s confidence tracks reality:

1

Claude Opus 4.8

Anthropic

0.610+0.11
2

Command A+

Cohere · Canada

0.545+0.21
3

DeepSeek v4 Pro

DeepSeek

0.532+0.19
4

Grok 4.5

xAI

0.522+0.17
5

Gemini 3.1 Pro

Google

0.462+0.38
6

GPT-5.6-sol

OpenAI

0.395+0.33

Right column: overconfidence gap — how far each model’s stated confidence runs above what the evidence warrants. Every model measured is overconfident. Per-model ranges across repetitions: ±0.002–0.019. Full data, paired statistics, and disclosures in the open repository.

The newest isn’t the best-calibrated

The lowest score belongs to the newest flagship measured — below even a current small model from the pilot.

Calibration ≠ marketing tier

Cohere’s Command A+ — Canada’s frontier model — places second, ahead of Google, xAI, and OpenAI.

It reproduces

Three independent repetitions per model, ranges as tight as ±0.002. Deterministic scoring anyone can re-run.

What’s Next

v0.2: more scenarios, more models, same rules.

The scenario set expands to 150+, scorer revisions arrive as versioned changes — never silent — and the leaderboard re-runs as labs ship new flagships. Everything publishes.