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.
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.
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.
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.
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.
Session initialization — research question parsed, depth tier set, first dispatch queued. Does not fire again until phase transition or final delivery.
(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
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.
SVS authentication of all Phase 1 evidence nodes. Confidence downgrades applied. Flags raised to registry. Phase 1 → Phase 2 transition authorised or blocked.
Challenges to the Cartographer's terrain classification — questioning which claims belong in Settled vs Contested. Each objection steelmanned before challenged. Resolution condition specified.
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.
Each node: source, weight, known limitations, verification_status, and Methodologist-assigned confidence bound. This registry is the authoritative evidence record for the entire session.
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.
Phase 1 landscape objections with resolution conditions and strength grades. Resolved objections amend the terrain classification. Unresolved carry to Phase 2.
All SVS verification outcomes and integrity flags raised during Phase 1. Written to the session audit trail. Cannot be amended by research agents.
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.
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.
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.
SVS authentication of any new evidence nodes admitted in Phase 2. Phase 2 → Phase 3 transition authorised or blocked.
Challenges to the evidence base — methodology critiques, sample limitations, generalizability problems. Each objection: steelman → challenge → resolution condition → strength grade (Strong / Moderate / Speculative).
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.
Updated confidence bounds reflecting Phase 2 reassessment. Any downgrade must reference the Contrarian objection or Methodologist flag that triggered it.
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.
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.
Any [GRADE CHALLENGE] flags issued by the Methodologist and the Synthesizer's revision response. Both are written to the session audit trail.
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.
Final SVS check on any outstanding pending verifications. Financial advice framing check on Synthesizer draft. Phase 3 → delivery transition authorised or halted.
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.
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.
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.
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.
All unresolved Strong Contrarian objections surface verbatim. Not summarized. Not softened. Exactly as the Contrarian issued them, with resolution conditions intact.
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.
The Pragmatist's context-specific recommendations, bounded by the Synthesizer's confidence ceiling. Includes follow-on session proposals for unresolved knowledge gaps.
Complete Guardian flag registry, SVS verification record, grade challenge history, and confidence propagation trace — exportable at session close.
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.
Established / Probable / Contested / Gap — per evidence node, based on four-dimension validity assessment.
Anchors conclusion to evidence nodes registry at T = 0.0. Cannot self-upgrade confidence grades. Revision loop fires on any violation.
Converts finding to actionable output within the inherited confidence ceiling. Cannot recommend more confidently than permitted by the chain above.
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.
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.
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.
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.
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.
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
SVS mechanics, flag classification, halt authority, domain integrity modes, and the reasoning behind hidden model identity.
Read more →Full specifications — model assignment, temperature rationale, permitted actions, forbidden actions, and dispatch order logic.
Read more →The four confidence grades, how the Methodologist issues them as hard constraints, and how Contrarian objections are classified and preserved.
Read more →Join waitlist and run a Standard-depth session on a question that matters to your work.