N
Decision control for AI execution
You are letting AI
modify your code
without knowing
what it will touch —
until it breaks something.
Nirnex enforces scope, validates decisions, and records every action — before execution, not after something breaks.
Nirnex — Execution Signal
PROMPTUpdate payment validation flow↳ received
↓
CONTEXTScoped to 3 files · 47 symbols verified↳ built
↓
CONSTRAINTSmax_lines: 200 · blocked: legacy/*↳ applied
↓
EXECUTION2 files modified · 0 scope violations↳ complete
↓
TRACEtrace_id: pay-val-0491 · logged↳ stored
What actually happens today
You only discover it
after deployment.
- →AI reads partial context
- →Chooses paths you didn't intend
- →Modifies unrelated modules
- →Leaves no explainable trace
If This Sounds Familiar
You've already
seen this happen.
- →AI modified something outside the intended scope
- →A change looked correct but broke another module
- →You couldn't explain why the agent made that decision
- →You relied on tests to catch issues after execution
This isn't a tooling issue. It's a lack of control over how decisions are made.
What Nirnex Is
Nirnex is not
an AI tool.
It is a decision control system
for AI execution.
01
It decides what AI
is allowed to see
02
It decides what AI
is allowed to change
03
It proves why a
decision was made
Who This Is For
This is for teams
where mistakes matter.
01 ·
Large codebases
Hundreds of interdependent modules where one unscoped change cascades.
02 ·
Multi-agent workflows
Multiple agents sharing a codebase without enforced execution boundaries.
03 ·
Regulated environments
Financial, healthcare, and infrastructure systems that require auditable AI decisions.
04 ·
Critical systems
Production code where a wrong AI decision costs more than the sprint.
What Changes When You Use Nirnex
Three guarantees.
Every run.
Nirnex replaces probabilistic AI behaviour with structural guarantees — before execution, during it, and after.
01
Predictable.
You know the exact scope before a single line changes.
- ✓Scope is verified before execution begins
- ✓No silent side-effects — ever
- ✓Consistent results across every run
02
Bounded.
The agent cannot act outside its declared execution envelope.
- ✓Out-of-scope writes are structurally rejected
- ✓Allowed modules declared upfront
- ✓Constraints enforced, not assumed
03
Explainable.
Every decision is recorded, replayable, and auditable.
- ✓Append-only decision ledger per trace_id
- ✓Failures are explainable, not mysterious
- ✓Full chain available for review or rollback
Why Teams Use Nirnex
What changes when you stop letting AI guess
Most teams discover AI mistakes after deployment. Nirnex makes them impossible to execute in the first place.
For Developers
- ✓No hidden side effects from AI execution
- ✓Know exact scope before any change happens
- ✓Debug failures using full decision trace
- ✓Stop guessing why something broke
For Teams & Business
- ✓Fewer regressions from AI-driven changes
- ✓Reduced debugging time and engineering cost
- ✓Predictable and controlled AI behaviour
- ✓Audit-ready trace for compliance and review
How It Works
Context → Constraints → Execution → Trace
Four steps. No guesswork. Every action declared before it runs.
01 ·Context
Execution Context Object
Parse the live codebase with tree-sitter and ast-grep. Produce a precision-scored ECO — a verified map of every symbol, dependency, and file relevant to the task. The agent never guesses at scope.
02 ·Constraints
Task Execution Envelope
Wrap the task in a TEE: allowed modules, blocked files, max line scope. The agent sees only what it is permitted to see. Writes outside scope are rejected before they touch the codebase.
03 ·Execution
Controlled Execution
The agent executes inside declared bounds. Violations fail loudly, not silently. Confidence scoring flags uncertainty before execution — not after deployment.
04 ·Trace
Decision Ledger
Every decision — what was read, what was chosen, what was written — is logged under a single trace_id. Replay any failure in isolation. Root cause in minutes, not days.
Under the Hood
Three Layers. One Control Surface.
This Is Not Probabilistic Control
Deterministic
by design.
—Deterministic pipeline
Same inputs produce the same execution bounds — every time.
—Explicit constraints
Allowed modules, blocked files, and max scope are declared before any agent acts.
—Append-only decision ledger
Every decision is written once, immutable, replayable from any point.
—Evidence-based gating
Classification requires a verified parse graph. Guesses are structurally impossible.
Without Nirnex
Unpredictable.
- ✕Agent guesses at execution context
- ✕No scope enforcement — agent decides what to touch
- ✕Output looks correct until it isn't
- ✕No audit trail — decisions are invisible
- ✕Root cause: days of manual triage
With Nirnex
Controlled + explainable.
- ✓Verified parse graph — exact scope, not guesses
- ✓Task Execution Envelope — out-of-scope writes rejected
- ✓Explicit confidence score flags uncertainty before shipping
- ✓Full decision trace — replay any step from trace_id
- ✓Root cause in minutes — full chain available
The Objection
"I already have
tests and code review."
Tests validate outcomes. Code review validates intent. Neither validates decision correctness — whether the agent was operating on the right context, inside the right boundaries, for the right reasons.
Tests catch what went wrong after the fact.
Nirnex prevents invalid execution before it happens.
See It in Action
Update payment validation flow.
Same task. Unscoped agent versus Nirnex.
Without Nirnex
TaskUpdate payment validation flow
ScopeUnknown — agent decides at runtime
Files touched12 files across 4 modules
TraceNone
OutcomeRegression in 3 unrelated modules
With Nirnex
TaskUpdate payment validation flow
Scopepayment/validator.ts · payment/schema.ts · tests/payment.spec.ts
Files touched2 of 3 declared — 1 unchanged
Tracetrace_id: pay-val-0491 · full decision log
OutcomeChange applied. Zero unintended writes.
For CTOs · Staff Engineers · AI Platform Teams
Stop letting
AI guess.
Built for regulated enterprises and engineering organizations where a wrong AI decision costs more than the sprint. Read the full architecture specification.
Read the Architecture →