The Living Corpus

Build the thing that builds the thing. Then grade it like you mean it.

2.5 years. Eight MVPs. Over a million tests exercised across nine hundred builds. A method for building AI that doesn't lie to you about whether it's working.

1.4M+
Tests
4.5K+
Evals & Iterations
30+
Workflows (50+ agents)
15
MVPs & POCs
Section I

The Method

How the work gets made. A stack of habits. Compound interest.

Start from the artifact
Before changing anything, understand why it exists. The system as built, not the system as described. The answer is in what was made, not what was intended.
Plan before building
A half-page plan costs almost nothing. Three thousand lines of the wrong thing is expensive. Get alignment on the approach before touching the code.
Every change gets a feedback loop
Test. Hit the endpoint. Screenshot the UI. Iterate until it's right. Output goes from acceptable to near-perfect in one loop. Two loops. Three. Self-checking is the highest-leverage habit.
Observability before logic
You can't troubleshoot what you can't see. Every system ships with structured telemetry. Every decision traced. Every error captured. If it runs silent, it runs blind.
Proof of correctness ships with the product
A test harness isn't documentation. It's the executable proof that the thing works. Every path. Every edge case. One command. Built alongside the feature, never retrofitted.
No single point of provider failure
The system chooses the right model for the task. Cost optimization and fallback are architecture decisions, not configuration. Provider A goes down, the work continues on Provider B.
Evaluation is infrastructure
Don't build agents you can't grade. The scoring engine ships with the agent. Governance is enforced by code that refuses to deploy, not a committee that meets quarterly.
Artifacts that survive the platform
Everything ships as a single file. No build step. No framework dependency. Open it anywhere, send it to anyone. The medium enforces portability. Nothing rots because nothing depends on infrastructure that won't exist in three years.
Human above the loop
Auto-generate everything. Auto-send nothing. The human reviews every output before it touches the world. Beyond review: the human makes authored decisions about what the organization should become. Machines map what exists. Humans decide what should exist instead.
Section II

The Insights

What the work taught. Some obvious in retrospect. None obvious at the time.

Harvey can't grade its own homework.
When an enterprise deploys agents through a vendor platform, the vendor's analytics can't be the independent evaluator. You need an external evaluation layer. Same dimensions, different incentives. True for Harvey. True for every enterprise AI deployment. The evaluator must be structurally independent from the evaluated.
Agent evaluation is governance.
Governance is a scoring engine with veto power. Four dimensions. Weighted. Tier 1 at 1.5×. Below 75 doesn't ship. Safety enforced by code that refuses to deploy. The AESOP scoring engine, now ported to Legal OS. The pattern ports anywhere.
The missing layer is operations.
Everyone builds AI functions. Nobody builds the operating model around them. The ROI framework. The adoption tracking. The POC pipeline. The enablement kit. The help system. The layer that makes the features actually get used. The 5-layer architecture is a reusable pattern for any enterprise AI program. Knowledge. Governance. Functions. Operations. Organization. Stack them. Ship them.
Visual explainers are source code for understanding.
A thinking tool. Commit to an aesthetic. Pick the right rendering approach for the question. Make every typography and color choice carry meaning. The squint test. The swap test. Self-contained HTML so the artifact survives the platform. The portfolio of diagrams is the portfolio of thought.
Intuition is not a decision process.
Four agents. One output. The literal first word: GO, NO-GO, or CONDITIONAL. A 900-word decision document with system diagrams, metrics baselines, and a 30-second stakeholder test. Language Discipline enforcement flags vague aspirational language and rejects it. Prompts versioned. Rubrics attached. Anti-patterns documented. Prompt engineering treated as software engineering. A decision factory.
Automation is the easy part. Authored change is the work.
Most AI transformation maps what exists and automates it. Phase 1. Phase 2 questions whether what exists is the right answer. Stuart Winter-Tear: "Organizations are compensating mechanisms for the characteristics and limitations of human intelligence." Departments exist because people specialise. Hierarchies because managers cannot coordinate unlimited work. Meetings because context has to move between human beings. Approvals because judgement is scarce. If AI makes context cheaper to retrieve, knowledge persistent rather than personal, routine judgement delegable, coordination costs lower — then the structures built around those constraints deserve to be questioned. The AESOP authored discovery pipeline does exactly this. Phase 1 maps the inherited process. Phase 2 asks 7 categories of decision-first questions about structural assumptions. The decision surface captures what changes, who owns it, what boundaries apply, and what happens next. Rethinking whether the org chart is the right answer.
Help systems should be agentic.
A chat interface that knows your documentation, retrieves relevant context, and answers in plain language. Source references so you can verify. Offline fallback when connectivity fails. Built once, ported across platforms in an afternoon. Not a docs site. A thinking partner for anyone using the system.
Section III

The Style

The fingerprints. The things that show up across every project, every repo, every artifact, whether anyone asked for them or not.

Build Signatures

Structured telemetry on every system
Every project ships with full observability. Every decision traced. Every error captured. Every timing measured. You can't improve what you can't see.
Tests as proof of correctness
Every project carries its own verification. 990 tests in AESOP alone. Collectively exercised 1.4 million times. Proof, not documentation.
Provider independence
No lock-in to any single AI provider. The system picks the right model for the task. Failures degrade gracefully. Cost optimizes without code changes.
Data isolation at the foundation
Tenant boundaries enforced at the database level, not in application logic. Client data cannot leak. The architecture guarantees it.
Development without friction
Every system has a demo mode. Full functionality without full infrastructure. Fast iteration in development. Hardened in production.
Deployment as commodity
Every project deploys the same way. The stack ports. The commands don't change. Infrastructure is a solved problem so the focus stays on what's being built.
Section IV

The Vision

What all of this is in service of. The thesis. The operating model. The thing being assembled.

The AI Transformation Operating Model
AESOP OS identifies what to build, builds it, evaluates it, and certifies it. Legal OS applies the pattern to a specific domain. Together they form a complete operating model for AI transformation: Identify → Build → Evaluate → Certify → Deploy → Monitor → Repair. A running platform with 1.4 million tests.
Agent governance is the missing layer in every enterprise AI strategy.
Companies are deploying agents at scale with no independent evaluation. The vendor provides the dashboard. The vendor grades the output. The vendor declares success. This is not governance. Governance is an external evaluation engine with veto power, drift detection, and a certification system. AESOP and Legal OS are the same pattern applied to different domains. The pattern is the IP.
AI literacy is the signal path.
You can build the best agents in the world. If lawyers and staff don't trust them, don't know when to override them, don't understand what they're good at and what they hallucinate — adoption collapses. The enablement kit, the help system, the workshop materials, the prompt engineering guides: these aren't documentation. They're the human interface to the AI operating model. Build them first. Ship them with the agents.
Honest measurement or nothing.
No tortured ROI numbers. No vanity adoption metrics. Cost impact = time saved × the right rate, not the biggest rate. Adoption = active users / eligible users, with a denominator. Quality = override rate, not satisfaction surveys. If you can't measure it honestly, you don't know if it's working. If you don't know if it's working, you're cosplaying transformation.
Section V

Authored Change

The work that distinguishes transformation from automation. Not mapping what exists and speeding it up. Questioning whether what exists is the right answer.

The Framework

Stuart Winter-Tear's organizing insight: "Organizations are compensating mechanisms for the characteristics and limitations of human intelligence." Departments exist because people specialise. Hierarchies exist because managers cannot coordinate unlimited work. Meetings exist because context has to move between human beings. Approvals exist because judgement is scarce.

If AI changes the economics of those constraints — context becomes cheaper to retrieve, knowledge becomes persistent rather than personal, routine judgement becomes delegable, coordination costs fall — then the organizational structures built around those constraints deserve to be questioned. Not optimized. Questioned.

Two-Phase Discovery

AESOP's discovery pipeline splits into two distinct phases, each with a different relationship to truth:

Phase 1: Inherited.
Maps what exists. The current process. The org chart. The handoffs. The approval chains. This is the organization as it is — captured faithfully, without judgement. Every step documented. Every stakeholder identified. Every constraint recorded. This is what you'd give a consultant on day one.
Phase 2: Authored.
Questions structural assumptions through 7 decision-first categories: Decision Mapping (what decisions are actually being made, by whom, with what information), Context Requirements (what knowledge does each decision require, where does it live, how fresh must it be), Coordination Needs (handoffs, timing, failure modes), Constraint Identification (which constraints are real vs. inherited vs. flexible), Success Definition (ideal outcome, metrics, minimal viable), Structural Assumptions Audit (role necessity, org design gaps, eliminable steps), and Autonomy & Accountability (ideal autonomy level, accountability chain, decision audit trail). Phase 1 told you what the organization does. Phase 2 asks whether it should.

The Decision Surface

The output of authored discovery is a decision surface — a structured, actionable JSON artifact that captures:

Decisions
The specific choices the organization faces. Not recommendations. Fork points.
Friction
Where the current structure creates resistance. The hidden invoice. The cost the org chart doesn't show.
Proposed Change
What shifts. A new handoff. A removed approval. A collapsed hierarchy. Specific, falsifiable.
Boundaries
What cannot change. Regulatory. Contractual. Political. The hard walls vs. the soft ones.
Owner
Who carries the decision. Named. Accountable. A person, not a title.
Corridors
The acceptable range. What good looks like, what too far looks like. Guardrails with taste.
Receipts
What would prove the change worked. Measurable. Time-bound. Hard numbers.
Structural Challenges
What in the current org design makes this hard. The thing nobody says out loud.
Next Step
The first action. Concrete. Doable this week. Because decision surfaces without first steps are just philosophy.

Why This Matters

Most AI transformation work automates the organization that exists. Authored change asks what the organization should become when the constraints that created it no longer apply. This is the difference between deploying chatbots and redesigning how legal work gets done. Between faster contract review and questioning whether contract review needs to be a separate function at all.

The decision surface connects directly to executive decision-making: it gives leadership a structured artifact they can act on, with named owners, measurable receipts, and explicit boundaries. It turns "we should think about how AI changes our operating model" into "here are the 7 decisions you need to make, here's what depends on each one, here's who owns them, and here's what happens after." Not advice. A surface.

Section VI

The Living Corpus

Why it was built. How. How it was graded. How it was watched.

Platform What It Is & Why It Matters Scale
MVPs
AESOP OS Enterprises deploy AI agents with no independent evaluation. Vendors grade their own homework. AESOP is the external evaluation layer: a full pipeline from discovery through authored organizational change to deployment, with 6 evaluation modes and a scoring engine that blocks unsafe output before it ships. Certification tiers stakeholders can trust. Drift detection that catches degradation before it becomes liability. 990 tests
6 eval modes
Legal AI OS Harvey agents deployed across the enterprise need independent evaluation. Harvey can't grade its own homework. Legal AI OS provides the external governance layer: 9 built legal functions, independent 4-dimension agent evaluation, drift monitoring, and an agentic help system. A reusable 5-layer architecture for any enterprise legal AI program. 62 tests
9 functions
SuperAssistant AI assistant platform for a professional services firm. Multi-tenant architecture where client data can never leak between tenants — enforced at the foundation, not in application code. Chat interface with knowledge management. Verified against production with 428 automated tests. 428 tests
TBG-owned
TSE-Stripped Voice interviews produce rich qualitative data. Processing it manually doesn't scale. A fully automated serverless pipeline: voice agents capture the interview, structured analysis is delivered automatically. The architecture for any voice-to-insight pipeline. Production deployed
Walter Persistent AI analysis that survives individual sessions. Knowledge that lives beyond the conversation. Session analysis, insight extraction, social-safe excerpts. A firm's collective intelligence, indexed and retrievable. TBG-owned infra
POCs & Portfolio
Matter Intake Evaluator Legal intake evaluation with configurable rubrics, not hardcoded rules. Staffing teams and evaluation criteria that adapt to the firm, not the other way around. Deployed and in use. 24 tests
132 Visual Explainers A portfolio of thinking tools. Architecture overviews, process maps, decision flows, maturity models — each a single self-contained file that opens anywhere and survives the platform it was built on. Custom typography, intentional color, dual light and dark themes. The medium enforces portability. Nothing rots. 132 pages
Section VII

Reusable Patterns

The patterns that port. Built, tested, deployed across multiple projects. Ready to drop into anything.

Evaluation
Independent Agent Evaluation
A scoring engine that grades AI output across multiple dimensions. Veto power on safety-critical failures. Certification tiers that stakeholders can trust. Ported across platforms.
Evaluation
Behavior Drift Detection
Agents change over time. Subtle shifts in tone, scope, or accuracy compound into real risk. Automated monitoring catches drift before it becomes a liability.
Help
Agentic Help System
Documentation that answers questions. A chat interface that knows the system and retrieves relevant context. Works offline. Ports across platforms in an afternoon.
Infra
Provider Independence
No lock-in to any AI provider. The system selects the right model for each task. Failures degrade gracefully. Costs optimize automatically. Ported across platforms.
Infra
Tenant Data Isolation
Client data cannot leak between tenants. Enforced at the foundation layer, not in application code. Architecture guarantees what policy can only promise.
Pipeline
Structured Decision-Making
Discovery to decision in a single pipeline. Vague language flagged and rejected. Prompts versioned and evaluated. Intuition converted to actionable output.
Architecture
Enterprise AI Operating Model
A reusable 5-layer architecture for any enterprise AI program. Knowledge. Governance. Functions. Operations. Organization. Stack them. Ship them.
Pipeline
Human-Above-the-Loop Automation
Every automated output has a human review gate. Beyond review: the human makes authored decisions about structural change. The machines propose. Humans dispose.
Pipeline
Authored Organizational Change
Phase 1 maps what exists. Phase 2 questions whether it should. Produces a decision surface executives can act on: what changes, who owns it, what boundaries apply, what happens next.
Visual
Self-Contained Visual Communication
Every diagram, every explainer, every portfolio page is a single file. No dependencies. No build step. Opens anywhere. Survives the platform it was built on.
Pipeline
Voice-to-Insight Processing
Voice interviews processed end-to-end without human touch. Structured analysis delivered automatically. The architecture for any voice-to-intelligence pipeline.
Ops
Honest ROI Measurement
Cost impact measured against real rate cards. Baselines calibrated with documented methodology. Quality tracked by override rate, not satisfaction surveys. Numbers you can defend.
Ops
AI Portfolio Management
Track every AI initiative from discovery through graduation. WIP limits. Department visibility. The operating model for managing an AI portfolio at scale.