AESOP Transformation OS — stage by stage
The Pipeline
Every project passes through every stage, in sequence. Each stage produces artifacts that feed the next. Gates enforce human approval before the pipeline moves on.
01
Stages & gates12 stages, 2 hard gates
FoundationOrganization
Context
Context
Organization Context — the grounding layer for everything below it. Captured context — authority and decision rights, coordination and dependencies, tacit knowledge, exposure posture, systemic blockers, plus departments and uploaded org documents — feeds Delegation Readiness, sharpens prioritization, and informs discovery. The sharper this context, the sharper every downstream judgment.
Outputs: authority & decision rights · coordination map · systemic blockers · departments · org documents
Organization LayerPortfolio &
Delegation Readiness
Delegation Readiness
Portfolio & Delegation Readiness — the layer above individual projects. The Portfolio coordinates every project across departments and ranks them by strategic score. Delegation Readiness assesses whether the organization can responsibly scale agent autonomy and turns gaps into tracked action items, promotable straight into the portfolio.
Outputs: portfolio board & table · strategic prioritization · readiness verdict · org action items
ContainerProjects
Projects — top-level initiative grouping. A project wraps every artifact across the entire pipeline — discovery sessions, agreements, builds, evaluations, repairs — under a single entity for portfolio-level tracking.
Outputs: project-id propagation · status tracking · portfolio analytics
Pre-Engagement GateWorking
Agreement
Agreement
Working Agreement — a signed document establishing scope, responsibilities, and response commitments, rendered from structured discovery data into a static template. Design is gated behind a signed agreement, with a manual override for internal projects.
Outputs: signed agreement · RACI · SLAs
Gate — signature required before Design
STAGE 01Discover
Discover — a structured problem interview capturing the stakeholder map, autonomy classification, and commitment readiness. Produces an Opportunity Brief with a triage verdict (GO / NO-GO / CONDITIONAL). Platform-agnostic — not locked to any implementation.
Outputs: opportunity brief · stakeholder map · triage verdict · autonomy level
STAGE 02Design
Design — Claude-powered PRD generation with a multi-platform architecture recommendation. An Architecture agent evaluates the PRD against platform capabilities (Glean, Claude API, OpenAI, Generic) and produces an 11-dimension gap analysis.
Outputs: PRD · architecture recommendation · gap analysis · platform scores
STAGE 03Build
Build — a six-phase instruction-writing pipeline with a human-in-the-loop checkpoint after gap analysis: Fitness Check, Gap Analysis (HITL), Workflow Design, Write Instructions, Generate Artifacts, Summary. Real-time SSE streaming to the UI.
Outputs: system instructions · platform artifacts · state-change hypothesis · translation-debt assessment
Gate — HITL approval at Phase 1 (gap analysis)
STAGE 3.5Deploy
Deploy — ship certified agents to target platforms. Build-session selector, real-time deploy status, dry-run support, launch-package integration. Ten platform connectors: Anthropic, Glean, OpenAI, Google, NemoClaw, REST, Base, Factory, WebSocket, Managed Session.
Outputs: live deployment · session selector · deploy-status dashboard · platform connectors
STAGE 3.6Operationalize
Operationalize — generates a Launch Package: four organizational artifacts from pipeline data already captured during discovery and build. No new data collection required; a generation pass over existing outputs.
Outputs: RACI matrix · tiger-team brief · maintenance plan · adoption plan
STAGE 04Evaluate
Evaluate — multi-agent scoring across seven evaluation modes: Standard Analysis, Cave of Shadows (red-team), Custom Rubric, Bias Probe, Investigation, KB Audit, and Trace Analysis.
The name “Cave of Shadows” draws on Plato’s Allegory of the Cave — prisoners mistake shadows on the wall for reality. AI systems can project confident, convincing responses that mask underlying flaws: bias, hallucination, boundary violations. The adversarial red-team drags those shadows into the light, testing whether an AI holds up under personas and edge cases designed to expose what lurks beneath.
Outputs: scored reports · certification level · risk strategy · attack scenarios
STAGE 4.5Regulatory
Mapping
Mapping
Regulatory Mapping — upload your regulatory templates (ISO 13485, risk registers, change-control forms). An agent identifies fillable fields, maps them to AESOP pipeline evidence, and produces completed documents for QMS filing. Evidence infrastructure, not a regulatory authority.
Outputs: completed templates · field mappings · evidence traceability
STAGE 05Repair
Repair — uses extended thinking to analyze evaluation failures deeply and rewrite system instructions. Takes evaluation context, scenario results, trace-analysis taxonomy, and optionally Tiger Team findings as input. Closes the loop between evaluation, stakeholder feedback, and improvement.
Outputs: rewritten instructions · change rationale · diff summary
STAGE 06Track
Track — score-trajectory visualization, version history with diffs, per-agent and aggregate cost tracking, portfolio-level analytics, and a maintenance-schedule dashboard with drift alerts.
Outputs: score trajectories · version diffs · cost analytics · drift alerts
02
Closed-loop improvement
The Evaluate–Repair–Track cycle runs continuously.
Core loop
Evaluate → Repair → Re-evaluate → Track delta. Tiger Team findings feed back into Repair from outside the loop.
Second loop
Launch Package → Stakeholder Review → Repair. Stakeholder feedback after launch returns into instruction repair.
03
The Eleven Tenetsmethodology as code
The decision framework embedded in every pipeline stage. Not just documentation — methodology enforced at each gate.
1State Change
Every agent articulates the specific state change it creates.
2Problems Before Solutions
Explore the problem space first; reject if better solved by a process change.
3Evidence Over Eloquence
Cite sources; distinguish authority from inference.
4Know What You Need
Classify outputs as deterministic, non-deterministic, or the dangerous middle.
5People Are the Center
Map human experience; assess translation debt.
6Humans Decide
Agents propose, humans choose; a veto needs no justification.
7Multiple Perspectives
Completeness over consensus.
8Context and Brevity
Enough to evaluate, not enough to overwhelm.
9Guardrails Not Gates
Five guardrail questions: evaluation, testing, adoption, maintenance, sustainability.
10Trace the Connections
Map second- and third-order effects.
11The Questions Stay the Same
Standardized inquiry enables portfolio-level learning.