DETERMA does not generate software changes, orchestrate agents, or replace the systems that build, scan, and ship code. It answers one question at execution time: does this exact change still hold sufficient authority and evidence to proceed? This page describes the model. It does not describe a production-ready, generally available, or customer-validated deployment.
Hard stops
WAITING_APPROVAL is a hard stop.A pending state is terminal until authority is granted. It is never treated as an implicit allow, never timed out into execution, and never bypassed by retry. Human approval alone is also insufficient: an approval that is stale, out of scope, reused, or unbound to the exact change does not authorize execution.
When any blocking condition is present, the outcome is no-proceed. There is no partial execution, no blind retry, and no degraded-mode allow. An unknown outcome is treated as a halt, not as success. The absence of a decision is never an authorization.
Human approval is a required input for the change classes that demand it — it is not a substitute for the scope, freshness, witness, release, and verification requirements described below.
Mutation preconditions
These conditions are conjunctive. Satisfying six of seven does not produce a partial authorization; it produces a block.
The change is normalized to one unambiguous description: the specific repository, branch, object, and field or operation permitted.
Authority is bound to that scope, has not expired, and can be consumed exactly once.
The target state is re-read at decision time and compared to the approved before-state. A mismatch fails closed.
A release scoped to the envelope, its idempotency key, and its expiry is issued, and is marked consumed on use.
A narrow adapter applies only the allowed field or operation. Nothing broader is reachable from the release.
The decision, its inputs, and the release are written to an append-only evidence ledger before and after execution.
The target is re-read after execution to confirm the outcome matches the approved after-state.
The governed mutation path described here is a provisional isolated-sandbox implementation. It is not a customer capability, not a production capability, and not evidence of broad enforcement.
Interactive product demo
Concept demonstration — not live authority or production enforcementAI agents can propose software changes. DETERMA determines whether the exact change has sufficient authority and evidence to proceed. Choose a decision to see the flow run. An ALLOW state here is an illustrative diagnostic outcome, not permission to merge or execute.
01
AI Change Proposal
02
Evidence Binding
03
Approval Context
04
State Witness
05
Change Release
06
Verification Receipt
Choose ALLOW or a DENY reason to see the deterministic receipt.
Authority is evidence-based and deterministic. It is not derived from model confidence. This flow has been validated in an isolated Git sandbox and is one mechanism inside the broader AI Software Change Control product.
Evidence Model
AI-generated software changes can be reviewed, approved, and still drift before they reach main. Branches move. Files change. Policies shift. Prior approvals can be reused. DETERMA binds provenance, approvals, evidence, and current repository state to the exact change so authority can be recomputed before it proceeds.
What it proves:An approval is not enough if the software change or repository state moved after approval.
Why it matters:AI-assisted workflows often approve a change before the repository advances. DETERMA revalidates the exact change against current state before it proceeds.
What it proves:A change release is short-lived and cannot be replayed.
Why it matters:An old release cannot be reused to advance a second mutation, in the sandbox or in future enforced deployments.
What it proves:Every allow or deny decision is reconstructable from structured evidence: provenance, approvals, state, and policy signals.
Why it matters:Engineering, security, and audit stakeholders need reconstructable evidence, not model explanations.
Truth boundary
The evidence model supports metadata-first assessments and bounded Shadow Mode. Governed repository mutation has been validated only in an isolated Git sandbox. Enforcement is introduced only after scope, evidence, and controls justify it.
Integrations & inputs
DETERMA connects to the systems engineering, security, and platform teams already operate. These tools provide the evidence and context that authority decisions are computed from. DETERMA does not replace them.
GitHub, GitLab, or comparable
Build, test, and delivery signals
Scanners and SBOM tools
Human and machine identity systems
Platforms such as Rein Security
Policy engines and change tickets
Integration is progressive: metadata assessment first, Shadow Mode next, evidence binding and approval orchestration after, exact-action authority and governed mutation only when scope and controls justify it.
Where DETERMA fits
Existing systems answer whether the agent, code, pipeline, or identity is acceptable. DETERMA answers whether this exact change still has sufficient authority and evidence to proceed now.
DETERMA is complementary to these categories, not a replacement, and no claim of superiority, production readiness, or customer validation is made here.
Illustrative sample report
A structured deliverable engineering, security, and leadership can act on. The sample below is illustrative structure only — not customer evidence.
Subject
A qualitative structure showing what a completed assessment deliverable contains. No numbers, scores, or savings are implied.
AI Change Inventory
Which changes originated with or were advanced by AI agents.
Provenance Gaps
Where agent and task attribution is incomplete.
Evidence Coverage
Which SCM, CI, AppSec, IAM, and policy signals are present or missing per change.
Approval Mismatch
Where the change that shipped differs from what was reviewed.
Review Burden
Where AI-generated PR volume exceeds review capacity.
Recommended next step
Scope a bounded Shadow Mode evaluation
This is an illustrative report structure. It is not a production customer audit and should not be represented as one.
Report sections
Executive Control Brief
Workflow Map
AI Change Inventory
Evidence Gaps
Approval-to-Action Mismatches
Shadow Mode Recommendation
Current status & readiness
Each capability has a precise readiness label. Expansion is earned by evidence from the prior stage.
Available for bounded design-partner evaluation
Proposed next-stage design-partner evaluation
Sandbox-validated
Provisional isolated-sandbox implementation; not a customer or production capability.
Requires bounded customer validation before production use
DETERMA does not claim general availability, production deployment, customer traction, or broad production enforcement. Portfolio enforcement is a future expansion earned by evidence from bounded rollout.
Progressive path to enforcement
The safest entry point is historical analysis, followed by live non-blocking observation, then increasingly active controls. Exact-action authority and governed mutation are a provisional isolated-sandbox implementation; they are not a customer or production capability, and bounded enforcement remains dependent on bounded customer validation.
A proposed fixed-scope, read-only assessment path. Metadata-only reconstruction of past AI-assisted software changes. Evaluates what DETERMA would have allowed, denied, or routed for review at the time. No code execution, write access, secrets, or production blocking.
Proposed next-stage design-partner evaluation — requires exact workflow scoping and technical release gates. DETERMA would evaluate live changes alongside existing workflows and produce deterministic verdicts and receipts without blocking the pipeline.
DETERMA recommends ALLOW, DENY, or REVIEW and routes exceptions to humans. Existing systems retain final enforcement authority.
Exact-action authority and governed mutation enforced for one explicitly scoped repository, branch, change class, or workflow with fail-closed behavior, rollback, operator controls, and auditable receipts.
Multi-repository policy and authority governance, after the bounded workflow proves reliable and valuable.
Expansion is earned by evidence from the previous stage. Historical Replay is a metadata-based reconstruction — not code replay and not execution replay.
Architecture Incubator
The exact-action authority primitives behind AI Software Change Control are being evaluated for AI-agent-initiated actions across other external systems. This is a bounded architecture incubator — not a change of DETERMA's primary product positioning, and not a production offering for any external vendor.
Next step
Assess one AI-assisted software delivery workflow with metadata-first evidence. No production credentials, write access, secrets, or source code by default.