Runtime Execution Authority
AI can act. DETERMA decides whether it may act now.
DETERMA is an external execution authority layer. It checks whether the exact action an AI is about to perform is still covered by its approval, its scope and the current state — before it changes a real system.
Let AI work. Keep authority with you.
01
Proposal
02
DETERMA — Execution Authority
03
State change / external effect
What DETERMA checks before execution
Six checks, evaluated at execution time.
Not a score, not a heuristic. Each check is evaluated against the exact action that is about to run.
01
Intent
Is this the action that was actually proposed and understood?
02
Scope
Does the requested effect stay inside the bounds that were reviewed?
03
Approval
Is there an approval that covers this exact action, unconsumed?
04
Current State
Does the system still look like the state the approval was bound to?
05
Policy
Do the rules that govern this class of effect still permit it now?
06
Executor / Tool
Is this executor allowed to perform this action, in this scope?
Why existing controls are not enough
Each layer answers a different question.
These layers are necessary and DETERMA assumes they exist. None of them decide whether a specific action may take effect right now.
Identity proves who.
DETERMA validates what may happen now.
Sandboxing controls where.
DETERMA controls authority.
Orchestration keeps work moving.
DETERMA decides whether it may continue.
Observability explains what happened.
DETERMA can block before the side effect.
Category-level comparison. DETERMA sits beside these layers, not against any vendor.
Use cases
One current wedge. Several evaluation tracks.
The first proof surface is deliberately narrow. Everything else is presented as a Design Partner evaluation track, not availability.
Current — first evaluation wedge
GitHub / software change
Open PR Authority Continuity: one historical or sanitized pull request where approval existed, then the world changed before merge. Evaluated against intent, scope, approval, current state, policy and executor before anything is applied.
See all use casesDesign Partner evaluation tracks
MCP / tool calls
Authority evaluation for agent tool invocations.
Cloud operations
Bounded parameter changes with state revalidation.
CRM / revenue operations
Allowlisted field writes with readback.
ITSM / operations
Status transitions gated on evidence.
Finance / payments
Draft preparation with single-use releases.
Procurement
Bounded record actions with exact-action approval.
Administrative workflows
Repetitive state changes under defined authority.
These tracks describe evaluation scope with a design partner. They are not statements of availability and imply no named vendor integration.
First product / first proof
GitHub Authority Check.
Open PR Authority Continuity: a governed evaluation of one historical or sanitized pull request where approval existed and the world changed before merge. The starting path is read-only or sanitized replay by default — no production credentials, no write access, no production mutation.
The first evaluation checks for three drift classes: Target/Base Drift, Approver Authority Drift, and Policy/Ruleset Drift. GitHub is the first wedge, not the boundary of the platform.
- No autonomous merge.
- No permanent model credentials.
- No success claim before verification.
- Read-only by default during assessment.
Governed lifecycle
AI may propose. Authority decides. Constrained executors mutate. Evidence proves.
Evidence and validation are not authority.
01
Proposal
02
Validation
03
WAITING_APPROVAL
04
Human approval context
05
Exact scope binding
06
State Witness where required
07
Fresh single-use Execution Release
08
Constrained Executor
09
Append-only audit
10
Post-execution verification
Evidence, not claims
Three levels of proof, stated separately.
We separate what the mechanism demonstrates from what has been implemented and from what a market has validated. The three are never blended.
Mechanism proof
The authority model itself: exact-action binding, state revalidation, single-use release, fail-closed behaviour on ambiguity, drift or replay.
Verified implementation slice
Internal evaluation of a narrow software-change slice. Internal findings only; not customer validation and not a production capability.
Market proof
Shown only when real, linked customer evidence exists. Nothing is presented here until then.
We do not display customer logos, deployment counts, savings figures or risk-reduction percentages. Where evidence does not exist, the section stays empty.
See the evidence modelDesign partner path
Four steps. Enforcement only under a separate approval.
Each step is bounded, reversible and evidence-producing. Nothing escalates automatically.
01
01 · Read-only assessment
Metadata-first review of an existing workflow. No writes.
02
02 · Historical replay / shadow mode
Replay past actions, or observe live proposals without effect.
03
03 · Bounded evaluation
A defined scope, a defined executor, a defined set of actions.
04
04 · Enforcement
Only with a separate, explicit approval. Never a default.
What this does not require
- No production credentials by default.
- No write access during assessment.
- No secrets handed over.
- A defined workflow, agreed in advance.
- An evidence report at the end.
Business outcomes
What teams are trying to change.
Qualitative outcomes only. We publish no ROI figures, no savings estimates and no quantified risk reduction.
- Less manual coordination around bounded approvals and revalidation.
- Fewer broad standing permissions held by agents and automation.
- Less ambiguity and less after-the-fact reconciliation.
- Traceable evidence from intent through to outcome.
- More autonomy for AI inside defined authority boundaries.
These are the outcomes an evaluation is designed to test, not results we claim on your behalf.
Why this is hard to copy
Five properties, stated plainly.
Authority continuity
Authority is carried from proposal to effect, not re-guessed at the end.
Exact-action binding
Approval attaches to one action, not to a role or a session.
Single-use release
A release authorises one execution and cannot be replayed.
Verified effect
The effect is read back and compared against what was authorised.
Evidence lineage
Intent, approval, state and outcome remain linked as one record.
Next step
Which action would you want AI to be able to perform safely?
Tell us the workflow and the action. We will tell you whether a bounded evaluation makes sense — and what evidence it would produce.
