How do you know what you know — and what happens when you're wrong?
An open framework that makes consequential AI claims verifiable and correctable by treating evidence, authorization, challenge, reconsideration, correction, and escalation as explicit system states.
It is not a separate body of work from the products. It is why they are shaped the way they are: NextConsensus holds the evidence state, Ambit the authorization state, Refract the change state. Each one makes explicit a state that consequential systems usually leave implicit.
The method
Every AI system deployed in high-stakes environments makes implicit claims: that a guideline is current, a drug protocol is safe, or a patient needs immediate prioritization.
A claim is verifiable when you can see what backs it, challenge is possible, and corrections reach the systems that relied on it.
Most AI governance work starts with the operating model — who approves what, under what conditions, for how long. That is downstream. First: how do you know what you know, and how is an error repaired?
NextConsensus freezes each forecast before the outcome is known. Refract keeps source changes verifiable, and Fast Harm, Slow Repair applies the same discipline to recovery evaluation.
Ambit implements the operating model in agent infrastructure: authorization held apart from capability, changes with their own inverse, verification before trust, and bounded tool scope.
Ethotechnics publishes numbered, versioned standards. Each is crosswalked to NIST AI RMF, ISO/IEC 42001, and the EU AI Act; the mapping is mine, not an endorsement by those bodies. No outside institution has adopted one yet, so they remain proposals.
Indexed, not yet readableJustice SLOs, verifiable-credential schemas, a minimum viable contestability standard, and an institutional failure postmortem template. They are named so the count is checkable; links appear when the documents resolve. A FHIR profile set was drafted and is now marked deprecated upstream, so it is not counted here.
—
Indexed, not yet readable
—
—
—
What the standards bind
The twelve lawsEach law carries the invariant a standard clause binds: capability does not imply authority, authority decays unless renewed, every delegation creates a correction obligation, and nine more.
The Ethotechnical invariantOne sentence the laws compress to: no system may accumulate consequential agency faster than the institution accumulates the capacity to inspect, challenge, revise, and survive its decisions.
Core axiomsFive commitments the laws are owed to, with the argument for why they are entitlements rather than preferences.
TheoryEssays on why the laws hold. Motivation only: no standard cites one as a requirement, and a reader can adopt a clause without the argument.
Principles
What I require from the teams I lead. Each one links to the work it
came out of.
01Human override is a product feature.
Override, escalation, and return-for-review are core capabilities that determine whether a high-stakes system can safely deploy.
Recurs in two deployments
AndwiseAutomated clause analysis surfaced and explained each issue. An accountable human reviewed every analysis before a physician could act on it.
EthotechnicsAuthorization states written down as an open standard: who may override, and what the override obliges them to record.
02Provenance beats confidence scores.
Knowing where a recommendation came from — and what evidence backed it — matters far more than an opaque confidence score nobody can audit.
Recurs in two deployments
RefractEvery change replayed into a verifiable event carrying its provenance, with the judgment left to the caller.
NextConsensusA public ledger of forecasts registered and frozen before the outcome is known, scored against the record afterwards.
03The decision the AI feeds into matters more than the AI's output.
When an automated system causes harm, the decisive question is never just 'was the model wrong?' It is 'who owned the decision the model fed into?'
Recurs in two deployments
EpicInstalled, signed off, and live — and still not doing the job, because nobody owned what happened after the go-live.
The Crumple ZoneEssays on the gap between an automated recommendation and the person who has to carry it out.
04Governance belongs in the product.
Evaluation, monitoring, escalation, and correction work best when designed directly into daily product workflows rather than managed via external committees.
Recurs in two deployments
EthotechnicsAuthorization, correction, and escalation published as open, versioned standards.
AndwiseCompliance review was a routed step in the flow with accountable sign-off.
05A deployment is finished when corrections reach every record it touched.
An error corrected at the source remains active everywhere it already propagated.
Recurs in two deployments
RefractDownstream systems receive what changed and when.
Fast Harm, Slow RepairA protocol that measures how far a wrong output travels before the correction catches up with it.
06Evaluate workflows before models.
A model that performs well in evaluation can still fail at the integration point where busy clinicians have to use it.
One instance so far
TranscarentFour programs shipped on one shared decision architecture because routing was the binding constraint.
07Missing approval can be earned back; a harmed patient can't.
A premature launch creates downstream costs that dwarf any speed advantage when the risk is clinical.
One instance so far
EpicThe escalation route that had to exist before the software could honestly be called live.
08Hospital IT trusts the system around the model.
Hospital IT, security, and legal teams evaluate the whole operational system surrounding a model — reviewing access boundaries and audit trails as closely as the interface.
One instance so far
DoximityHospital security reviews cleared the product on documented policies and procedures.
The operating model
AI deployment requires continuing authorization: when a system may influence decisions, what conditions limit that authority, how people can challenge it, and what evidence or behavior reopens review.
A model recommends a treatment change. Authorization stays versioned and conditional the whole way.
State transitions for the treatment-change recommendation. Rows are in order; a state is entered only when the row above it exits.
State
Trigger
Who acts
Evidence gate
What propagates
State 0Not authorized
A use is proposed.
Named deployment and escalation owners
Defined clinical or operational decision
Named deployment and escalation owners
Specified affected population and exclusions
Nothing. The proposed use cannot influence care or workflow.
State 1ObservedThe recommendation is observed but cannot influence care.
State 0 exits when A bounded use case, accountable owner, and evaluation plan are approved.
—
Silent or retrospective evaluation
Error taxonomy and exception review
Baseline comparison against current practice
Nothing to care or workflow. Outputs are captured for evaluation.
State 2Constrained useAfter a silent evaluation, it may inform a narrow workflow under human review.
State 1 exits when Observed performance and failure modes justify a limited prospective deployment.
Explicit human review on each recommendation
Prospective workflow validation
Documented override and escalation paths
Monitored safety, equity, and operational indicators
Recommendations into one narrow workflow, within limited population scope and predefined stop conditions.
State 3Routine reliance
State 2 exits when The deployment performs acceptably inside its stated scope and the institution can pause, correct, or roll it back.
Independent review, with challenge rights preserved
Stable prospective performance
Independent evaluation appropriate to the use
Operational readiness for correction and rollback
Routine reliance within the defined scope, with change detection and periodic reconsideration.
Any stateReopenedA new safety signal suspends it immediately.
01 Detect Identify a potentially material change in evidence, model behavior, policy, data, workflow, population, or observed outcomes.
The clinical safety lead opens the reconsideration case and issues the disposition.
04 Review → Reconsideration case Present the source-traced change, prior rationale, observed performance, dissent, and unresolved uncertainty to the authorized reviewers.
06 Propagate → Corrected operating state Update the authorization record and each connected workflow, instruction, interface, monitoring rule, and affected stakeholder.
The loop opens on a change, not on a review cycle. A system that can only be reconsidered on schedule is unrevisable in between.
Three tests
Explainable
A clinician, compliance officer, or patient can see what the system considered and recommended, with enough detail to understand the basis for action.
Challengeable
The recommendation can be overridden, escalated, or sent back for review without halting care or creating a compliance incident.
Correctable
When the evidence, policy, or model changes, the decision can be updated and the record corrected. Someone is responsible for that correction.
Key concepts
Terms I use for AI deployment, institutional decision-making, and the boundary between human judgment and automated systems.
Term
In a deployment
The property What is at stake.
Revisability
Ensures automated decisions have explicit correction paths so mistakes don't become permanent policy.
Ensuring that when evidence, policy, or models change, relying systems automatically update and correct affected downstream workflows.
2011Georgia Tech RNA
Modeled thermodynamics and barrier kinetics deciding whether a molecular process completes or stalls.
View record →
2024Refract & NextConsensus
Built deterministic change-detection bots and claim-trajectory scoring against public records.
View record →
2025Fast Harm, Slow Repair
Designed a recovery-evaluation scaffold measuring harm duration and correction propagation; the draft covers one of twelve development cases.
View record →
I study what changes when a decision that used to end with a person ends in software instead: where errors travel, who absorbs the work they create, and whether the system can repair itself when the information under it changes.
Three cases, and which of the questions each one asks
Case
Mechanism
Who absorbs the cost
Which questions
Emergency department loudspeaker
Patients are called for triage by name over a loudspeaker. For a Deaf patient the mechanism deciding who waits is inaudible.
The Deaf patient; what it costs is measured in mortality.
Who waits?
What cannot be allowed to fail?
Epic: a misrouted clinical alert
A misrouted alert could bury a critical lab result.
Whichever clinician trusted the queue. The routing system carried no matching accountability.
Who absorbs error?
What gets buffered?
Andwise: monetization
The monetization paths most likely to fund growth would have made employers or financial institutions the customer.
In 2012 I helped build a medication-recommendation system for type 2 diabetes, and we measured the thing that now worries me most about deployed models: not whether the recommendation was right, but what happened to the clinician's own judgment once they had seen it.
Internists n=262%→92%change +30 pts
Familiar with the rules n=464%→86%change +23 pts
Endocrinologists n=468%→76%change +8 pts
Unfamiliar with the rules n=271%→71%change +1 pts
Agreement with the algorithm, before and after the clinician saw its recommendation. Blinded validation, twenty patient data sets per reviewer, six reviewers in total. Percentages are rounded; the subgroups nest rather than partition, so the reviewer counts do not sum to six. Rows are ordered by how far each group moved.
The specialists barely moved and the generalists moved most. The group that moved least of all was the one that did not know how the algorithm worked, and the group that understood its rules moved almost as far as the generalists. Understanding the tool predicted deferring to it.
A recommendation moves the judgment beside it, and it moves different readers by different amounts. This course project's data shows the direction but cannot size the effect. That is the mechanism the authority-boundary work exists to constrain: a model's output becoming institutional policy without an explicit decision.
The constraint the measurement above argues for. Evidence may move an assessment on its own;
nothing moves an institutional position without a person crossing the boundary and leaving a
record of having done it. “No action” is one of the dispositions, because a system where
declining to act is not a recordable choice will drift into acting by default.
Jain, K., Patel, K., Rowland, J., Yong, C. DiaMonD (DIAbetes MONitoring and Dosing) System. Wallace H. Coulter Department of Biomedical Engineering, Georgia Tech / Emory University. Advised by Dr. Lawrence Phillips, MD.
Go deeper
Case studiesWhere this came from: five deployments through hospital security review, federal quality reporting, and clinical sign-off.
Decision recordsFour decisions with the reasoning as I wrote it down before the outcome was known, and the contemporaneous document attached where one survives.
PrinciplesWhat 14 years of clinical products left me unwilling to ship without.
Reliance LabOne composite deployment, six months in, with new information on the table. Make the call and see what it commits the organization to.