Augle Deliberation Index

Where does AI
reach consensus
and where does it
break down?

The Deliberation Index tracks consensus rates, dissent levels, and confidence grade distributions across domains — as the corpus grows. Not what the models think. What the structured deliberation produces.

Recent sessions · illustrative
PolicyDo apprenticeship programmes improve long-term participant earnings?Probable
TechnologyDoes the evidence support commercially relevant photonic-chip throughput?Contested
PolicyDo indoor masking mandates reduce COVID-19 community transmission?Contested
Life sciDoes current GLP-1 evidence support long-term maintenance without dosing?Contested
EconomicsDoes the incumbent credit-scoring model retain predictive validity today?Contested
PolicyWhat is the realistic SEC exposure on the material non-public information question?Probable
Consensus by domain

Consensus rates across
all deliberation domains.

DomainSessionsConsensusDissent
Life sciences34
84%
12%
Economics51
78%
18%
Finance29
74%
22%
Policy38
72%
26%
Climate22
68%
28%
Geopolitics41
62%
35%
Technology47
58%
38%
AI governance18
49%
46%

How consensus is measured.

Consensus rate is the proportion of sessions in a domain that reached a Probable or Established confidence grade without an Unresolved Strong Contrarian objection. High consensus means the ensemble consistently converged. Low consensus means the questions in that domain consistently produce unresolved adversarial pressure — which is information, not a failure.

High consensus

Cartographer, Methodologist, and Synthesizer converge. Contrarian objections are resolved or graded Speculative. Final grade: Probable or Established.

High dissent

Contrarian raises Unresolved Strong objections that carry to Phase 3. The finding reflects the dispute — Contested grade or multiple unresolved objections in the output.

AI governance has the lowest consensus rate in the corpus — 49%. This is structurally expected: the questions are genuinely contested, the evidence base is thin, and the Contrarian has strong material to work with. The index is accurate because it reflects the actual state of the evidence.
Confidence + dissent heatmap

How confidence and dissent
interact across domains.

Each cell shows the average metric for that domain × confidence grade combination. High consensus doesn't always mean low dissent — a domain can produce many Probable findings while still generating significant Contrarian pressure on individual claims.

Contested %
Consensus %
Dissent flags
Avg grade
Economics
22%
78%
1.4
Probable
Life sciences
16%
84%
0.9
Probable
Policy
34%
72%
2.1
Contested
Geopolitics
46%
62%
2.8
Contested
Technology
52%
58%
3.2
Contested
AI governance
63%
49%
4.1
Contested
Strong (high consensus, low dissent)ModerateWeak (low consensus, high dissent)
Confidence grade distribution

How findings are distributed
across the four grades.

All sessions · illustrative data
Probable48%
The most common grade. Best available evidence supports the claim but replication is limited.
Contested31%
Active dispute or unresolved Strong objection. The finding names the dispute, not a direction.
Established14%
Multiple independent replications across distinct methodologies. Requires formal objection to downgrade.
Gap7%
Insufficient evidence to evaluate the claim. A first-class finding — not a failure state.

What the distribution reveals.

The preponderance of Probable grades reflects an accurate picture of where most research questions sit: supported by evidence, but not at the level of independent replication across methodological contexts. Contested at 31% is structurally honest — the ensemble isn't softening findings. Gap at 7% is small but important: these are the questions the evidence genuinely cannot answer.

Key signal

AI governance questions produce a Contested or Gap grade in 71% of sessions — the highest contested rate of any domain. Technology is second at 58%. Life sciences and economics are the most resolvable domains in the corpus, with Probable or Established grades in over 80% of sessions.

What Contested means in practice

A Contested finding is not a failure to reach a conclusion. It is a structurally complete finding that names the dispute, preserves competing positions, and surfaces the resolution condition. In 31% of all sessions, the honest answer is "the evidence is genuinely split." The Index records that accurately.

Corpus composition

What the corpus looks like
as it grows.

The Index is a descriptive record of how the ensemble behaves across domains — how often it converges, how often the Contrarian's objections go unresolved, and how findings distribute across the four confidence grades. It measures the epistemic behaviour of the system, not a claim of predictive accuracy.

The most useful signal is where deliberation breaks down. A high Contested-or-Gap rate in a domain is not a weakness in the system — it is an accurate reading that the evidence there is genuinely unsettled. The ensemble is built to say so rather than manufacture confidence.

33
Sessions · illustrative
11
Domains covered
38%
Contested or Gap findings
93%
Avg Guardian integrity
Example sessions, shown to demonstrate the Index structure. Live sessions publish here as they run.

Add your sessions
to the Index.

Every session you run contributes to the corpus.
Join the waitlist — one Standard session free.