Observation → HAVE
Where are we?
Observations record what was found, with evidence and uncertainty. HAVE records what a person has ratified as the accepted reference. An observation does not become HAVE by itself.
Reference architecture · Software research pilot
Generation got cheap.
Vigilance did not.
AI can produce changes faster than people can judge them. You still need to understand what changed, whether it remains sound, and whether it advances the intended purpose.
SolventAgent explores Matthew Gradient’s reference architecture for keeping intent, execution and evidence visible—so you can direct your attention where judgment matters and retain informed control.
SolventAgent is a software research pilot in development. This page introduces the architecture and illustrates one part of its proposed review experience.
01 · From generation to vigilance
You can approve the work and still lose your bearings.
The feeling of lost control can begin before anything visibly fails. The output looks plausible. Tests pass. Changes are accepted. Yet the connection between the original purpose and the current result becomes harder to explain.
An approval button gives you a decision to make. Making that decision well requires a clear view of what changed, why it changed, and what the available evidence establishes.
Aviation addresses the vigilance challenge through training and instrumentation. Pilots are taught Aviate, Navigate, Communicate (ANC): maintain control of the aircraft, establish and follow the intended course, and communicate without allowing that demand to displace the first two. The priority order helps direct limited human attention when several tasks compete.
Cockpit instrumentation supports that discipline by making the aircraft’s state readable and distinguishing alerts by urgency. Training and instruments work together: the pilot needs to know what deserves attention and have the information needed to act.
Aviate: maintain control. Navigate: keep the intended course. Communicate: make the state and decisions clear.
The reference architecture applies this principle to AI-assisted work. ANC supplies the attention priority; explicit records and decision points keep intent, accepted state and evidence available as the work changes.
There is an echo of Attention Is All You Need here. That paper concerns computational attention. The question here is human: as execution accelerates, what deserves our attention, and what do we need to see to judge it?
02 · Staying oriented
Keep the state, the purpose and the next decision in view.
The architecture keeps observations, accepted state and intended work distinct. A model’s plausible account of the project cannot silently become the reference everyone relies on.
PLANNED → bounded work
Where are we going?
Intent is made explicit and developed into bounded increments. Each increment identifies the outcome sought, the scope of the work and what would count as progress.
Attention → decision
What needs judgment now?
Consequences, departures and important unknowns stay visible. ANC helps prioritize what threatens control or direction before lower-priority detail consumes attention.
Approve the plan. Agree on the intended outcome and proposed approach.
Authorize the action. Permit a bounded piece of work to proceed.
Ratify the result. Assess the evidence and decide what becomes part of HAVE.
Plans can change. When an increment takes a different approach, its rationale stays alongside the intent and evidence. Each accepted increment leaves a snapshot that later work can refer to, revisit or branch from.
Only ratification moves HAVE. Unknowns remain visible after acceptance.
03 · Two judgments
Preservation and progress need separate answers.
What remains sound?
Compare the candidate with HAVE. Which accepted guarantees and constraints remain intact? What changed, and what was left unchecked?
This is the preservation question.
What moved us forward?
Compare the result with PLANNED. Did the increment achieve its intended outcome? If the approach changed, is the reason visible and justified?
This is the progress question.
A change can preserve every checked boundary and still leave the intended feature unbuilt. It can also deliver the feature while weakening something the system depended on. Both judgments belong in the acceptance decision.
04 · One instrument within the architecture
A boundary change, made visible.
Software gives us concrete structures to inspect. Some boundary conditions can be machine-checked; others require runtime evidence or human interpretation. A contract card can make those findings easier to examine.
The illustration below focuses on preservation. It does not answer whether the proposed work is useful or complete. That still requires comparison with the intent for the increment.
Four predefined scenarios. This illustration does not analyse a repository. Selecting a candidate never changes the fixed HAVE₀ reference.
HAVE₀ · Fixed ratified reference · Inspect the starting contract
Inputs: cartId: CartId; clock: Clock; config.checkout.*; SessionRepository (port).
Guarantees: resolve(): Session — always; throws SessionExpired; emits session.opened; no other effects.
Permitted capabilities: pure computation, reading config, emitting events and calling ports. Direct network and database I/O are outside this role’s envelope.
This record remains unchanged throughout the illustration. A candidate finding is not ratification.
src/checkout/session.ts
What may enter inputs
- cartId: CartId
- clock: Clock
- config.checkout.*
- SessionRepository (port)
Implementation
Omitted from this boundary view. Targeted code review and tests remain part of verification.
What must remain true guarantees
- resolve(): Session — always
- throws SessionExpired
- emits session.opened
- no other effects
Permitted capabilitiespure computationread configemit eventscall portsno network i/ono database i/o
Review a candidate
HAVE₀ loaded · no candidate selected
Choose a proposed change to compare it with the starting contract. The baseline stays fixed.
These are scripted findings. They establish neither implementation correctness nor progress toward the intended outcome.
Respected: the listed boundary is preserved. Widened: an additional capability remains within the permitted envelope. Violated: a guarantee is weakened or the envelope is exceeded. Every finding still needs evidence and a human decision.
05 · Architecture and implementation
A reference design.
A pilot in software.
Matthew Gradient’s reference architecture describes how intent, bounded execution, evidence and human decisions fit together. SolventAgent is a research pilot exploring selected parts of that design in software.
The pilot is scoped to advisory review. The wider architecture includes authorization and ratification controls that require separate implementation. The illustration above demonstrates a proposed interaction, rather than verified extraction or repository coverage.
A finding must state what was checked and what remains unknown. Tests, runtime observation and targeted code review supply evidence that a boundary view alone cannot provide.
Apply the questions to your organisation
Where is AI making the work harder to judge?
Matthew Gradient’s €900 preliminary diagnostic explores how your team notices unexamined acceptance, misplaced effort and drift from its mission.
A 45-minute guided questionnaire is followed by a short written judgment profile. It identifies concerns suggested by your answers, the reported examples behind them and important unknowns, prioritised through Aviate, Navigate and Communicate.
Your company’s mission and distinctive value proposition guide what matters most. Workflow inspection, codebase evaluation and implementation are separate engagements.
Explore the €900 diagnosticThe diagnostic is based on your answers. The underlying argument is developed in Judgment Insolvency and the Matthew Gradient.