← TM Forum Hub HAIEC · DETERMINISTIC ASSURANCE
01 / 15
TM Forum 2026 · Agentic Assurance · Pre-Event Technical Thesis

When AI Can Act,
Proof Has to Catch Up.

Deterministic assurance for agentic systems — from distributed telemetry to independently reproducible control proof.

Pre-event architecture · synthetic examples are not assessed event evidence

3 agents

Customer · IT · Network

3 controls

C7 · C9 · C16

Cross-platform

AWS · ServiceNow · OTel

One proof layer

HAIEC deterministic Control Test

The Problem Has Changed

We gave AI agency.
Proof still fragments across systems.

A consequential action can cross multiple agents, gateways, models, tools, APIs, a digital twin and a human workflow. The evidence cannot remain trapped in separate dashboards.

Runtime

AgentCore, gateways, tools, models, ServiceNow and network systems execute the work.

Observability

OTel, CloudWatch, gateway records and AICT show pieces of what happened.

Assurance

Someone still has to bind the declared control, frozen threshold, exact run and exact supporting records.

The Central Distinction

Observability is necessary.
It is not assurance.

Observability tells us

  • what emitted telemetry
  • what traces exist
  • what requests occurred
  • what systems were involved

Assurance must establish

  • what governed the run
  • which threshold/baseline/cap was frozen
  • what actually occurred
  • whether the control held
  • which exact evidence supports that result
TELEMETRY≠EVIDENCE≠FINDING≠ASSURANCE
The Architectural Wager

AI can operate the system.
AI does not establish the verdict.

The organizer requires Nemotron in the agent path. That strengthens our separation: the agent may reason with Nemotron; HAIEC does not ask Nemotron whether the agent complied.

Nemotron-backed agent path→Runtime evidence→Frozen control→Deterministic calculation→Persisted result
AI may help

Navigate evidence, compile a natural-language question into bounded query parameters, or explain a persisted result.

AI may not silently decide

Threshold, assessed-run selection and authoritative control verdict stay outside model judgment.

Validated Event Gates

The challenge has three integration gates
before our first assessed run.

A · Nemotron

Verify actual model identity of supplied agents. If Nemotron-backed, preserve the binding. New/replacement agents use Nemotron.

B · AWS ↔ ServiceNow AICT

Create a unique team connector, prove agent discovery, capture stable identities, and qualify model/trace discovery limitations.

C · Telemetry

Prove real gateway + OTel/CloudWatch evidence and decide direct HAIEC ingestion versus LogSense normalization/export fallback.

HAIEC's Model of Truth

Five questions. Five evidence planes.
No shortcuts.

REQUESTEDWhat was asked or declared?
POLICY AUTHORIZEDWhat did an approved, effective policy/change permit?
EFFECTIVELY GRANTEDWhat could the acting identity actually exercise?
CODE CAPABLEWhat could deployed-bound source/config cause?
OBSERVEDWhat actually happened?

A CR does not automatically prove approval. IAM/config existence does not automatically prove effective authority. A copied repo does not prove deployed code. One plane cannot prove another.

Freeze Discipline

HAIEC is ready.
The event policy is not frozen yet.

Pre-event technical baseline

Evaluator semantics, evidence intake, C7/C9/C16 logic, result vocabulary, query surface and judge workspace stay stable.

Onsite governing instance

Freeze exact scope, threshold/baseline/cap, evidence requirements, source/binding profile and assessed-run selection rule only after discovery + calibration.

DISCOVER→QUALIFY→CALIBRATE→DECLARE→FREEZE→RUN→CONTROL TEST

Threshold change after freeze = new policy version + new run. Never rewrite the prior result.

Fit the Environment

We do not replace the lab.
We make its evidence coherent.

AWS / AgentCore

Runtime + agent identity. HAIEC needs exact run/session/deployment and model/tool invocation IDs.

Gateway + OTel / CloudWatch

Observed actions, calls, timing and usage; possible C16 enforcement point if a true refusal hook exists.

ServiceNow AICT

Inventory, AWS connector, sessions/traces where available, and human-loop evidence. AICT does not become the HAIEC verdict engine.

ODA / Cedar / Guardrails

Authorization/effective-state evidence when approval, scope, revision and effective time are actually established.

Digital Twin / KPI

C9 metric/calibration/baseline source.

LogSense

Ingest, normalize, correlate and reconstruct when direct HAIEC ingestion is unavailable or when forensic exploration is needed.

Controls as Executable Contracts

Three controls.
One assurance contract.

C7 · AIA-LOG-001

Expected-event coverage, required zones/enforcement points, timing gaps and exception allowance.

C9 · AIA-ARC-006

Frozen KPI/baseline, comparable windows, drift threshold/allowance, alert and named human path.

C16 · ACN-COST-001

Provider-executed calls, token/cost accounting, hard cap and genuine refusal/enforcement evidence.

Control IDFrozen scopeThreshold/versionEvidence requirementsDeterministic testResult vocabulary
Judge Query

Ask a factual question.
Get evidence, not a story.

“What did this AI system do between 14:32 and 14:35?”
Identityrun / trace / agent
Actiontool / model / API
TimeeventTime / clock
Sourcegateway / OTel / AICT
Boundarycoverage / gaps
QUERY

Retrieves raw normalized facts and exact evidence refs. No verdict recomputation.

CONTROL TEST

Uses the frozen governing instance and reconciled evidence to return the persisted deterministic result.

Authority vs Reality

Permission is not execution.
Execution is not authorization.

Declared / authorized

  • approved ODA change / effective policy
  • Cedar / Guardrail / configuration
  • frozen control threshold
  • approved model/tool or human workflow

Observed

  • OTel trace
  • gateway/model call
  • tool/API invocation
  • ServiceNow session/alert
  • human acknowledgement/action or silence

HAIEC reconciles these planes without pretending one proves the other.

Architecture Generality

One assurance architecture.
Multiple enforcement points.

Model / agent gateway

Potential in-transaction C16 check + runtime evidence.

Tool boundary / MCP

Second genuine enforcement point only if it can actually allow/refuse the action.

ServiceNow human loop

C9 alert → named queue/recipient → acknowledgement/action or documented silence → evidence back into HAIEC.

The bonus is not “we have multiple dashboards.” It is reuse of the same bounded control/evidence architecture at real enforcement points.

Failure Is Still Proof

Assurance is not a green dashboard.

SATISFIED

Required evidence is sufficient and the frozen test holds.

NOT SATISFIED

The frozen test does not hold for the assessed run/window.

NOT EVALUATED

Prerequisite, scope or evidence requirements are not sufficient for the Control Test.

EVIDENCE / COVERAGE CONDITION

Incomplete, limited or conflicting evidence remains a separate condition — not a fourth control verdict. A trustworthy system must be able to say what it cannot establish.

What Survives

The hackathon is an instance.
The architecture is the product.

Here

3 agents · 3 controls · AWS + ServiceNow · a bounded telecom fault scenario.

Enterprise scale

Many agents · control catalogs · multiple runtimes/clouds · multiple enforcement points · evidence across organizational boundaries.

HAIEC is not a dashboard for these three controls.
These three controls are instances of an assurance architecture.
HAIEC · Human AI Evidence Company

AI can move faster than human review.
Proof cannot become optional.

Declare→Observe→Bind→Evaluate→Prove

High assurance in every consequence.

Appendix
Illustrative schemas + synthetic examples · replace with assessed event evidence onsite

Event freeze checklist

COMMON: run/session identity, event-time semantics, clock domain/skew, stable agent/gateway/AICT identities, source profile, assessed-run selection rule
C7: expected-event manifest/scope + gap/exception rule
C9: KPI profile + calibration + frozen baseline + threshold/allowance + human path
C16: provider-call/token/retry accounting + cap + enforcement scope

Source hierarchy

1. ORGANIZER FACT — official brief / direct organizer answer
2. EVENT OBSERVED — measured live environment fact
3. TEAM SPEC — our declared control/threshold/binding contract
4. HAIEC CAPABILITY — what our platform can do
5. ASSUMPTION / ONSITE VERIFY — never promote without evidence

C16 edge states

A. EXECUTED WITHIN CAP
B. OVER-CAP ATTEMPT PREVENTED
C. EXECUTED OVER-CAP BREACH
Runtime DENY = preventive evidence, not automatically a breach.
No atomic/no-overshoot claim unless proven.

Judgment-Day package

Evidence File
Threshold / Governance Document
Control Test / Judge Operator Card
Test Cases + Named PASS/Breach Runs
One-Page Architecture
Gap / Remediation / Retest Register