Framework Spine
One question per stage. One framework per question. Each has one job.
They exist because something went wrong in a way the existing vocabulary could not name.
The stages are the questions. The frameworks below exist to answer them, not the other way round.
| Stage | The question | Framework | Defined |
|---|---|---|---|
| Select | Which use cases deserve funding? | Opportunity Score finds them → BXT ranks them | Finding the work · below |
| Diagnose | Are we solving the right problem? | The Meaning Gap: Run × Reason | below |
| Classify | What are we actually building? | Application Modes → Agent Patterns → Risk Tiers | below · Agent Patterns · below, recorded in the agent registry |
| Build | How does it get assembled? | The Seven-Layer Stack | The Seven-Layer Stack |
| Ship | Is it ready? | Three Proof Gates → Production Gate Question → Pre-Flight Checklist → Agent Card, released by the Deployment Ladder | Three proof gates · below · Published tools · Deployment ladder |
| Operate | Is it still working? | The Three Drifts | The Three Drifts |
| Scale | Where are we as an organisation? | Agent Factory Maturity Model | The Agent Factory |
| (throughout) | What stops it going wrong? | Guardrail Stack · Four Pillars | Governance as code |
Every name in that table resolves. That is a deliberate rule for this page: a spine that names a framework and does not define it is an index of things you are expected to already know, which is the opposite of the job.
These sit beside the spine rather than on it. They answer questions that recur at more than one stage.
| Framework | The question it answers | Defined |
|---|---|---|
| Three Proofs | What evidence do we bring? | below |
| Five Modes | What is the human's job while it runs? | below |
| Seniority Ladder | What kind of agent is this, and what governance does it need? | The seniority ladder |
| Trust Ladder | How far has this particular agent earned its way? | NORTH and the trust ladder |
| The Dignity Clause | What do the people whose work this touches keep? | Value and the CFO |
The distance between what a system optimises and what the organisation actually needs. Two axes:
- Run: can it execute reliably?
- Reason: is it reasoning about the right problem?
Most organisations measure only Run. The dangerous quadrant is Precise but Wrong: high operational confidence in a system solving the wrong problem.
Presented at the Toronto Machine Learning Summit.
"If this agent gives the right answer to the wrong question, how would you know?"
It needs a specific operational answer before architecture work begins.
Four tiers, and they are already shipped: the Agent Card records one per agent and the MCP Catalog organises servers by the same scale.
| Tier | What the agent or tool can reach | The governance question it forces |
|---|---|---|
| 1 | Read-only, public or non-sensitive | Can it be wrong in a way anyone would notice? |
| 2 | Internal write | Who reverses it, and how long does that take? |
| 3 | External or regulated | Which regulation applies, and who signed off that it does? |
| 4 | Irreversible or financial | What is the cap, and what happens when it is hit? |
The tier is a property of the blast radius, not of the model. A small model with a payments tool is tier 4. A frontier model summarising a public document is tier 1.
Socially irreversible counts as irreversible. A message that can technically be deleted has still been read.
BXT scores a candidate use case on three dimensions so funding goes to what can actually ship: Business value, eXecutability, Trust. The BXT Scorecard is the shipped instrument, and it is the artifact of record for what the letters mean.
Three Proofs is what you bring to that scoring. Every initiative must demonstrate three things:
| Proof | Answers | Failure mode when missing |
|---|---|---|
| Technology | Does it work? | Succeeds in POC, fails at scale or in integration |
| Value | Does it matter? | Works, but nobody uses it and no metric moves |
| Competence | Can we run it, and fix it when it breaks? | Delivers value until the first incident, then has no owner and no path back |
Competence is the harder question and the one that predicts whether a pilot survives. A deployment can be entirely compliant and still fail it, because nobody on the client side can operate or remediate it at 2am.
These are not the same as BXT, and were once described as if they were. BXT names the dimension. Three Proofs names the evidence you bring to it. Trust is the outcome; Competence is what earns it. They are companions, not synonyms.
Until 2026-08-20 both were called "Four Modes," and the collision produced a long-running fork.
Application Modes classifies what the AI does: DOING (task automation), DECIDING (decision augmentation), DESIGNING (artifact generation), DIRECTING (agent orchestration). Each carries a distinct risk profile and governance requirement. This is what an Agent Card records.
Five Modes classifies what the human does as capability increases: Doing → Directing → Delegating → Designing → Defining. The human moves from doing the work, to specifying the steps, to specifying the outcome, to setting the environment, to setting standards for work they will never see.
Use both. One tells you what you are building. The other tells you what your job is while it runs, and the asymmetry is the point: autonomy transfers down that ladder and responsibility does not.
A third ladder is not this one. The Seniority Ladder also has five levels and it is a different question again: not what the AI does and not what the human does, but what kind of system this is and what governance that kind requires. The levels are numbered and the modes are not, which is deliberate.