The ensemble explained

Four models.
Seven agents.
Three phases.
One structured deliberation.

Augle doesn't ask a single model for an answer. It routes your question through a multi-agent ensemble — each agent with a defined role, fixed dispatch order, and typed output contract — across three deliberation phases before any finding is produced.

Topic Architectorchestration
Cartographerlandscape
Methodologistvalidity
Guardianphase boundaries
Contrarianadversarial
Synthesizerconclusion
Pragmatistphase 3 only
Phase architecture

Three phases. Each one builds on the last.

Agents don't run in parallel. Each phase produces a structured output that constrains the next. The Methodologist's confidence bounds become hard limits on the Synthesizer. The Contrarian's unresolved objections surface verbatim in the final finding.

Phase 1 · Exploration
Map the terrain

The Cartographer decomposes the question into settled ground, contested terrain, and unknown territory — producing evidence nodes with preliminary confidence bounds. The Methodologist evaluates validity of each node across four dimensions. The Guardian authenticates every source in real time.

Topic Architectinit
Cartographerlandscape map
Methodologistconfidence bounds
GuardianSVS · phase boundary
Contrarianlandscape objections
Synthesizerdraft integration
Phase 2 · Deliberation
Apply adversarial pressure

The Contrarian re-engages with a steelmanned version of every claim before challenging it. Each objection must specify a resolution condition and a strength classification — Strong, Moderate, or Speculative. Unresolved Strong objections carry forward to Phase 3 verbatim.

Cartographerevidence update
Methodologistgrade challenges
GuardianSVS · phase boundary
Contrarianevidence objections
Synthesizerrevised draft
Phase 3 · Synthesis
Produce the finding

The Synthesizer anchors its conclusion exclusively to the structured evidence nodes registry — not the discourse thread. Temperature is locked at 0.0 for deterministic output. The Pragmatist converts the finding into actionable next steps within the Synthesizer's confidence ceiling.

Guardianfinal integrity check
Contrariansynthesis objections
Synthesizerfinal finding · T=0.0
Pragmatistactionable output
The agents

Each agent has one job.
None of them can exceed it.

Every agent operates under strict typed output contracts — permitted actions and forbidden actions are architectural constraints, not guidelines. No agent can produce conclusions outside its role. No agent can exceed the confidence ceiling set by the Methodologist.

Independent integrity layer
Guardian
Anthropic + Augle· identity hidden
T = 0.1

The Guardian operates exclusively at phase boundaries. It does not produce research findings or participate in deliberation. Its model identity is hidden from all user-facing surfaces — including the other agents — to prevent anchoring effects. It holds halt authority over the entire session.

Source Verification Service — authenticates every citation across all three phases in real time
Issues flags with halt authority — can pause or terminate a session based on integrity violations
Domain-specific integrity modes: academic, legal, clinical, financial, editorial
Cannot produce research findings, conclusions, or directional recommendations
Cannot participate in deliberation discourse between agents
Model identity never surfaced to users or research agents — prevents anchoring
Session orchestration
Topic Architect
Anthropic
T = 0.2

The only agent with a direct user-facing interface. Fires once at session initialization. Manages dispatch sequencing, phase transitions, and final delivery. Explicitly forbidden from editorializing or softening any agent's output.

Cannot editorialize, soften, or reframe any agent output
Does not participate in research discourse
Landscape mapping
Cartographer
Google
T = 0.8

Dispatched first in every phase. Produces a five-component landscape: restated question, scope boundaries, evidence terrain map (settled / contested / unknown), evidence nodes with source and limitations, and a knowledge gap register.

Cannot assess evidence validity
Cannot produce conclusions of any kind
Validity assessment
Methodologist
OpenAI
T = 0.5

Evaluates each evidence node across four dimensions: internal validity, external validity, construct validity, and methodology-claim match. Issues confidence bounds — Established, Probable, Contested, or Gap — as hard constraints on the Synthesizer. Cannot be overridden.

Cannot produce conclusions or directional recommendations
Confidence bounds are architectural constraints — not suggestions
Adversarial pressure
Contrarian
Anthropic
T = 1.0

Runs at maximum temperature to maximize output variation and reduce sycophantic convergence. Must steelman every claim before challenging it. Each objection specifies a resolution condition and a strength grade. Unresolved Strong objections appear verbatim in the final output.

Must steelman every claim before challenging it — adversarial, not contrarian for its own sake
Addresses other agents by @mention — a structured discourse record
Conclusion + finding
Synthesizer
OpenAI
T = 0.0

Temperature is locked at zero — a hard architectural requirement for deterministic verdict production. Anchors its conclusion exclusively to the structured evidence nodes registry, not the discourse thread. Subject to three inviolable constraints that cannot be overridden by any other agent.

Conclusion cannot exceed the confidence grade of its supporting evidence
Insufficient Evidence findings cannot be converted to any directional recommendation
Financial advice framing (buy / sell / long / short) is forbidden in all modes
Application notes
Pragmatist
xAI
T = 0.3

Fires in Phase 3 only. Converts the Synthesizer's finding into actionable output tailored to the user's specific context. Inherits the Synthesizer's confidence ceiling as an absolute constraint — cannot produce recommendations more confident than the synthesis permits.

Cannot exceed the Synthesizer's confidence ceiling
Does not fire in Phase 1 or Phase 2
Confidence + dissent scoring

Every finding is graded.
Every objection is preserved.

Augle doesn't produce a single confidence score — it produces a structured evidence record. Each node carries a grade issued by the Methodologist. Each unresolved objection from the Contrarian appears verbatim in the output alongside the finding that triggered it.

Evidence confidence grades

EstablishedReplicated across multiple independent studies using sound methodology. The Synthesizer can anchor high-confidence findings to Established nodes.
ProbableWell-supported by available evidence but with material limitations — sample size, methodology constraints, or limited replication. The ceiling for most real-world findings.
ContestedEvidence exists on multiple sides without clear resolution. The Synthesizer cannot produce a directional finding from Contested nodes — the finding reflects the contestation.
GapThe relevant evidence does not exist, has not been studied, or cannot be sourced. Gap findings are a product output — not a failure state. They tell you what isn't knowable yet.

Contrarian objections · live session

Unresolved objections · Phase 3
@Contrarian → @SynthesizerStrong · Unresolved

"The notification-triggered self-report measures post-interruption attentional recovery, not naturalistic attention — causally distinct constructs. The thesis conflates them throughout."

Resolution condition: Direct comparison study of notification-triggered vs. researcher-initiated ESM on attentional outcomes

@Contrarian → @CartographerModerate · Resolved

"The innovation claim may not be differentiated from the 2024 Friederici lab publication with functionally equivalent architecture."

Resolved: Innovation reframed to structural differentiation in Phase 2

@Contrarian → @MethodologistSpeculative · Noted

"The 3-item Likert operationalisation of attentional state has no published reliability data for smartphone delivery contexts."

Resolution condition: Psychometric validation study for smartphone-delivered attentional measures

The engine

One engine. Built for research questions.

The multi-agent ensemble runs the same way every session. What adapts is the Guardian's domain-specific integrity configuration, tuned to the field the question sits in.

Research
For open research questions

Designed for the research questions that don't resolve to a binary outcome — grant applications, dissertation defences, policy analysis, clinical evidence reviews, investment thesis validation. Document ingestion supported. Guardian operates in domain-specific integrity mode.

OutputEvidence-anchored finding with confidence grade and actionable next steps
DocumentsUploaded papers, methodology chapters, grant drafts, case documents · PDF or DOCX
GuardianAcademic, legal, clinical, financial, or editorial integrity mode

See it in action.

Join waitlist and run a session on a question that matters to your work.