Phase architecture

Three phases.
Each one constrains
the next.

Augle's three deliberation phases — Exploration, Deliberation, and Synthesis — don't run in parallel. Each phase produces a structured output that the next phase is required to respect. The Methodologist's confidence bounds from Phase 1 become hard limits on the Synthesizer in Phase 3. The Contrarian's unresolved objections from Phase 2 surface verbatim in the final finding. Nothing is softened in transit.

Fixed dispatch order

Agent dispatch order is fixed within every phase. Parallel execution would allow agents to miss each other's outputs — eliminating the structured constraint propagation that makes the architecture reliable.

Typed output contracts

Every agent operates under a typed output contract — a defined structure for what it produces and what it is forbidden from producing. Contracts are architectural constraints, not guidelines.

Hard constraints, not suggestions

Confidence bounds established by the Methodologist are hard upper limits. The Synthesizer cannot produce a conclusion at higher confidence than its supporting evidence — not even if instructed to.

Fixed dispatch order

Every phase. The same sequence.

Topic Architect
Init only
Session params · phase transitions · final delivery
Cartographer
Phases 1–3
Five-component landscape map · evidence nodes · knowledge gaps
Methodologist
Phases 1–3
Per-node confidence bounds · grade challenges
Guardian
Phase boundaries
SVS authentication · flag classification · halt authority
Contrarian
Phases 1–3
Steelmanned objections · resolution conditions · strength grades
Synthesizer
Phases 1–3
Evidence-anchored conclusion · T = 0.0 deterministic
Pragmatist
Phase 3 only
Actionable output · inherits Synthesizer confidence ceiling
Phase
Exploration

Map the terrain.

Phase 1 establishes the evidence landscape — what's settled, what's contested, what's genuinely unknown. The Cartographer produces the five-component terrain map. The Methodologist issues preliminary confidence bounds on every evidence node. The Guardian authenticates every source. The Contrarian challenges the landscape itself.

Topic Architect
T = 0.2 · init only
Output

Session initialization — research question parsed, depth tier set, first dispatch queued. Does not fire again until phase transition or final delivery.

Cartographer
T = 0.8 · dispatched first
Output

(1) Restated research question · (2) Scope boundaries · (3) Evidence terrain map — Settled / Contested / Unknown · (4) Evidence nodes with source and limitations · (5) Knowledge gap register

Methodologist
T = 0.5
Output

Four-dimension validity assessment per evidence node: internal validity · external validity · construct validity · methodology-claim match. Issues Established / Probable / Contested / Gap as hard constraints on Synthesizer.

Guardian
T = 0.1 · phase boundary
Output

SVS authentication of all Phase 1 evidence nodes. Confidence downgrades applied. Flags raised to registry. Phase 1 → Phase 2 transition authorised or blocked.

Contrarian
T = 1.0 · max variation
Output

Challenges to the Cartographer's terrain classification — questioning which claims belong in Settled vs Contested. Each objection steelmanned before challenged. Resolution condition specified.

Synthesizer
T = 0.0 · deterministic
Output

Preliminary evidence integration. Not a finding — a structured summary of Phase 1 evidence nodes and their confidence grades. Cannot exceed Methodologist bounds. Carries forward to Phase 2.

Phase 1 structured outputs
Evidence nodes registry

Each node: source, weight, known limitations, verification_status, and Methodologist-assigned confidence bound. This registry is the authoritative evidence record for the entire session.

Knowledge gap register

Named gaps where evidence is insufficient. Gap is a first-class output — not a failure state. It tells you what isn't knowable from the current evidence base.

Contrarian objection set

Phase 1 landscape objections with resolution conditions and strength grades. Resolved objections amend the terrain classification. Unresolved carry to Phase 2.

Guardian flag registry

All SVS verification outcomes and integrity flags raised during Phase 1. Written to the session audit trail. Cannot be amended by research agents.

Carries forward
Evidence nodes registry with confidence bounds as Methodologist-issued constraints
Unresolved Contrarian objections from the landscape challenge
Knowledge gap register — gaps may be expanded but not silently removed
Synthesizer's preliminary evidence integration — Phase 2 builds from this, not around it
Phase
Deliberation

Apply adversarial pressure.

Phase 2 challenges the evidence base directly. The Cartographer updates the landscape with new evidence or scope refinements. The Methodologist reassesses challenged nodes and grades any new evidence. The Contrarian re-engages against the evidence base itself — not the landscape — and must steelman every claim before challenging it. Each objection must specify a resolution condition and strength grade.

Cartographer
T = 0.8
Output

Updated evidence terrain incorporating Phase 1 Contrarian resolutions and any new evidence admitted under REAP. No evidence can be silently removed — only reclassified with reason.

Methodologist
T = 0.5
Output

Reassessment of any evidence nodes that were challenged in Phase 1. Issues [GRADE CHALLENGE] flags where a preliminary Synthesizer claim exceeds its evidentiary warrant. Synthesizer must revise before proceeding.

Guardian
T = 0.1 · phase boundary
Output

SVS authentication of any new evidence nodes admitted in Phase 2. Phase 2 → Phase 3 transition authorised or blocked.

Contrarian
T = 1.0 · max variation
Output

Challenges to the evidence base — methodology critiques, sample limitations, generalizability problems. Each objection: steelman → challenge → resolution condition → strength grade (Strong / Moderate / Speculative).

Synthesizer
T = 0.0 · deterministic
Output

Updated evidence integration incorporating Phase 2 revisions. Produces an Evidence Landscape summary with per-claim confidence grades. If issued a [GRADE CHALLENGE], must revise before Phase 3 can proceed.

Phase 2 structured outputs
Revised evidence nodes registry

Updated confidence bounds reflecting Phase 2 reassessment. Any downgrade must reference the Contrarian objection or Methodologist flag that triggered it.

Objection registry — Phase 2

All Phase 2 Contrarian objections with resolution status. Strong objections unresolved at this stage carry forward to Phase 3 verbatim — they cannot be softened in transit.

Synthesizer evidence landscape

Per-claim confidence grades after Phase 2 deliberation. This is the evidentiary ceiling the Synthesizer must respect in Phase 3 — it cannot self-upgrade grades in the final synthesis.

Grade challenge record

Any [GRADE CHALLENGE] flags issued by the Methodologist and the Synthesizer's revision response. Both are written to the session audit trail.

Carries forward
Final evidence nodes registry — Synthesizer must anchor exclusively to this, not the discourse thread
All unresolved Strong objections — surface verbatim in final output, no softening permitted
Synthesizer evidence landscape — per-claim grades as hard ceiling on final synthesis
Phase
Synthesis

Produce the finding.

Phase 3 locks the deliberation into a final finding. The Synthesizer anchors exclusively to the evidence nodes registry — not the discourse thread. Temperature is locked at 0.0 for deterministic output. The Contrarian challenges the Synthesizer's draft directly. The Pragmatist converts the finding into actionable output, inheriting the Synthesizer's confidence ceiling as an absolute constraint.

Guardian
T = 0.1 · final integrity
Output

Final SVS check on any outstanding pending verifications. Financial advice framing check on Synthesizer draft. Phase 3 → delivery transition authorised or halted.

Contrarian
T = 1.0
Output

Challenges to the Synthesizer's draft finding — scope overreach, claim-evidence mismatch, anticipated objections from external reviewers. Critical pre-submission objections force a revision loop before Pragmatist dispatch.

Synthesizer
T = 0.0 · locked
Output

Evidence-anchored finding at temperature zero. Three inviolable constraints: (1) conclusion cannot exceed evidence confidence · (2) Insufficient Evidence cannot become directional · (3) financial advice framing forbidden. Plus 2–5 structured reopen conditions.

Pragmatist
T = 0.3 · Phase 3 only
Output

Converts the finding into context-specific actionable output. Inherits the Synthesizer's confidence ceiling as an absolute constraint — cannot recommend more confidently than the synthesis permits. Does not fire in Phase 1 or 2.

Phase 3 final delivery package
Evidence-anchored finding

The calibrated conclusion with confidence grade, anchored to the evidence nodes registry. The same question with the same evidence base must produce the same finding — the Verdict Invariance Requirement.

Unresolved objections · verbatim

All unresolved Strong Contrarian objections surface verbatim. Not summarized. Not softened. Exactly as the Contrarian issued them, with resolution conditions intact.

Reopen conditions · 2–5 required

Structured verdict fragility outputs. Each specifies: a specific observable trigger event · likelihood assessment · the claim whose grade would change · direction of change. General uncertainty not permitted.

Actionable next steps

The Pragmatist's context-specific recommendations, bounded by the Synthesizer's confidence ceiling. Includes follow-on session proposals for unresolved knowledge gaps.

Full audit trail

Complete Guardian flag registry, SVS verification record, grade challenge history, and confidence propagation trace — exportable at session close.

Unidirectional confidence propagation

Confidence flows
one direction only.

The formal constraint: the Synthesizer's conclusion grade cannot exceed the minimum confidence bound of the evidence nodes supporting it. The Pragmatist's recommendation confidence cannot exceed the Synthesizer's. These are hard architectural constraints — not guidelines. Violations trigger a mandatory revision loop.

Methodologist
Issues confidence bounds

Established / Probable / Contested / Gap — per evidence node, based on four-dimension validity assessment.

Hard upper limit — Synthesizer cannot exceed the minimum bound across all supporting nodes for any single claim
Synthesizer
Inherits as hard ceiling

Anchors conclusion to evidence nodes registry at T = 0.0. Cannot self-upgrade confidence grades. Revision loop fires on any violation.

Hard upper limit — Pragmatist cannot produce recommendations more confident than the Synthesizer's conclusion grade
Pragmatist
Inherits as absolute constraint

Converts finding to actionable output within the inherited confidence ceiling. Cannot recommend more confidently than permitted by the chain above.

Gap-graded findings cannot be converted to any directional recommendation — the Pragmatist must surface the gap, not work around it
The Grade Challenge mechanism

When the Methodologist detects that a preliminary Synthesizer claim exceeds its evidentiary warrant — where the Synthesizer has graded a claim higher than the minimum confidence bound of the supporting evidence nodes — it issues a [GRADE CHALLENGE] flag. The Synthesizer must revise the claim downward before Phase 3 can proceed. The challenge and the revision are both written to the audit trail. This loop repeats until no Grade Challenge violations remain or a maximum iteration count is reached.

[GRADE CHALLENGE]
Session depth tiers

Three depths.
One architecture.

The phase architecture is constant across all depth tiers. What changes is the number of rounds within each phase and whether the Guardian is active. The Rapid tier is a single-round preview — useful for quick orientation. Standard and Deep are for substantive research.

Rapid
Single round
Preliminary confidence-bounded finding

Cartographer and Synthesizer only. No Guardian. No Contrarian adversarial pressure. Produces a preliminary confidence-bounded finding without the full deliberative structure. Useful for initial question orientation or time-constrained situations.

AgentsTopic Architect · Cartographer · Synthesizer
GuardianNot active
ContrarianNot active
PhasesSingle round
Output: Preliminary finding with confidence bound. No unresolved objection record. No audit trail.
Standard · recommended
Three full rounds
Complete multi-agent deliberation

Complete multi-agent ensemble across all three phases. Guardian active at every phase boundary. Full Contrarian objection register. Recommended for any substantive research question where the cost of a wrong answer is material.

AgentsAll agents
GuardianActive at all phase boundaries
ContrarianActive all three phases
PhasesExploration · Deliberation · Synthesis
Output: Evidence-anchored finding · confidence grade · unresolved objections verbatim · reopen conditions · actionable next steps · full audit trail.
Deep
Three rounds + expert
Asynchronous domain expert interjection

Standard deliberation plus an asynchronous domain expert review at the Phase 1 / Phase 2 boundary. The expert's contribution is evaluated by the Guardian before integration — it enters the deliberation as authenticated evidence, not unverified context.

AgentsAll agents
ExpertDomain expert at Phase 1/2 boundary
GuardianEvaluates expert contribution before integration
FlagshipContrarian on Opus tier for maximum adversarial pressure
Output: Same as Standard plus authenticated expert contribution in evidence record. Deepest adversarial pressure available.
Worked example

What the architecture
looks like in practice.

The following is a condensed record from an illustrative Standard-depth Letters & Science session.

The session demonstrates all five architectural properties: the Contrarian's Strong objection successfully amended a Settled Ground assignment; the Methodologist's evidence ceiling propagated correctly across all phases; the Grade Challenge mechanism functioned as specified; the Pragmatist correctly declined to produce a directional recommendation for the Gap-graded claim; and the unresolved Moderate objection surfaced verbatim in final delivery.

Research question: Does the current evidence base support long-term weight maintenance without continued GLP-1 dosing?

Depth: Standard · Mode: Letters & Science · Guardian: Active

Illustrative session · representative of session structure
Session log · condensed
Phase 1 · Exploration
Cartographer4 items Settled Ground · 4 items Contested. "Weight regain follows discontinuation" → Settled.
MethodologistSession evidence ceiling issued: no claim may be graded Established. All nodes Moderate or below.
Contrarian"Weight regain follows discontinuation" assignment challenged — regain is well-documented, but its completeness varies and ≤2yr evidence cannot foreclose the maintenance question.Strong Resolved: moved to Contested.
Phase 2 · Deliberation
MethodologistReplication cap on lifestyle-alone maintenance claim.[GRADE CHALLENGE]
SynthesizerEvidence landscape: partial regain within 1yr (Probable) · universal full regain (Contested) · lifestyle-alone maintenance (Contested) · maintenance dosing preserves loss (Probable) · tapering mitigates regain (Contested).
ContrarianDurability of off-drug maintenance beyond 2 years.Moderate Unresolved → carries to Phase 3.
Phase 3 · Synthesis
ContrarianFollow-up gap — all primary nodes limited to ≤104 weeks.Moderate
MethodologistSplit required: Short-term regain ≤1yr (Probable) · Long-term >2yr maintenance (Gap).[GRADE CHALLENGE]
PragmatistDirectional recommendation declined for Gap-graded long-term finding. Gap surfaced as knowledge gap output.

Three phases.
One calibrated finding.

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