← TM Forum Hub01 / 11
TM Forum 2026 · Judge Cut · Submitted

From Agent Activity
to Defensible Proof.

HAIEC turns agent behavior into deterministic, queryable proof. Not telemetry. Not a dashboard. Not a checklist.

submission-haiec-20261006T150621Z.zip · submitted 2026-10-06 · sha256 1fde8544…

1 · Thesis

Telemetry is not proof.
The verdict follows the evidence.

What the platform emits

Logs, spans, model turns — raw telemetry. Necessary. Not a judgment.

What HAIEC produces

Qualified evidence bound to a named run, a frozen policy, and a deterministic evaluator → a persisted verdict anyone can recompute.

ACTION→NATIVE RECORD→RECONSTRUCTION→MEASUREMENT→FROZEN RULE→VERDICT→RESPONSE
2 · What We Proved

Three archetypes.
Six verifiable verdicts.

PresetRunMeasuredRuleVerdict
C16 PASS…5e312635,559 tok60,000 capSATISFIED
C16 BREACH…51ab52106,829 tok60,000 capNOT_SATISFIED
C9 PASS…723d80+60.20%D=+100%SATISFIED
C9 BREACH…048d48+286.03%D=+100%NOT_SATISFIED
C7 ORIGINAL…5e31269/1010 categoriesNOT_SATISFIED
C7 REMEDIATION…048d4810/1010 categoriesSATISFIED

Same frozen policy. Same evaluator. Different observed facts. Different verdicts.

3 · A/B Scoreboard — fresh runs, 2026-10-06

One variable. Our agent won all three.

ScenarioThird hopC7C16 (cap 60,000)GradeNeg. recs
S1haiec-network-agent · 8a097d10/1029,3347/83
S1reference chain · e429bb9/1033,7226/80
S2haiec-network-agent · e290e910/1038,0677/103
S2reference chain · 0d11d69/1034,0715/100
S3haiec-network-agent · db781210/1034,2799/103
S3reference chain · f828d49/1022,7978/100

Post-submission verification, scored with the organizer's own score-run.py. Every haiec-network-agent run: C7 10/10 + 3 negotiation records. Every reference run: 9/10 + zero. "Guardrail verdict" check fails for all chains — an audit record-format gap, not a governance absence.

4 · C16 · Boundary Proof

Same rule.
Opposite verdicts.

PASS — fault-1791167110-5e3126

35,559 of 60,000 qualified tokens · 8 calls · SATISFIED · ctr-237f3928…

BREACH — fault-1791165466-51ab52

106,829 of 60,000 · +46,829 overshoot · 17 calls · NOT_SATISFIED (SPEND_CAP_EXCEEDED) · ctr-6d142008…

One frozen 60,000-token policy — 547f4a67, digest-pinned — produced both. That is the definition of a deterministic control. ServiceNow INC0010319 tracks governance separately.

5 · C9 · Drift → Governance

A real breach.
Then a governed human loop.

PASS window

Worst agent +60.20% — inside frozen D=+100%. SATISFIED · ctr-d3383065…

BREACH window

IT agent +286.03% — far outside D=+100%. NOT_SATISFIED · ctr-0ba48de1…

Then

Deterministic failure → alert → named human ACK → persisted response → anti-replay → ServiceNow INC0010337.

Same system 44f, same policy 13c07652. HAIEC does not stop at detecting drift — it binds measurement → rule → verdict → response → governance record.

6 · C7 · Failure → Remediation → Proof

The 9/10 failure stayed.
The 10/10 proof is new.

ORIGINAL — 4043 / ae6d

9/10 categories. Missing: IT↔Network NEGOTIATION. NOT_SATISFIED. Capability existed. Permission existed. Observed delegation did not.

REAL BEHAVIORAL FIX

Network sent the negotiation → IT evaluated and returned the decision → Network received it. Same correlation lineage.

PROOF — 44f / dcfa

10/10 → SATISFIED · ctr-bea968ae… · valid SYSTEM-SCOPED remediation pass — not an exact canonical 4043/ae6d retest.

Most teams can detect a failure. HAIEC proved failure → exact causal gap → changed agent behavior → independent proof — without rewriting the original result.

7 · Findings Beyond the Verdicts

Findings are not verdicts.
HAIEC keeps them separate.

SEC-01 · credential exposure

Shared/plaintext runtime credential exposure. No compromise/abuse claim.

SEC-02 · ceiling enforcement gap

Token-ceiling enforcement gap/lag — 4 post-exhaustion 200s. No universal limiter-bypass claim.

SEC-03 · phantom runaway

~84/86 model turns; only one real downstream tool invocation.

EXP-04 · scanner exposure

Public-listener probes observed, 404. No compromise claim.

HIS-05 · resolved drift

Historical configuration drift; current driftDetected=false.

S2-FALSE-CERTAINTY-001

False closure refused: 5/10 → remediation → 5/10 → NOT_FIXED.

Detector findings and deterministic Control Tests answer different questions — zero generic findings would not mean a control passed.

8 · Refused to Claim

Where the proof stops
is itself evidence.

9 · Verify It Yourself

Don't trust the slide.
Reproduce the verdict.

VERIFY

Judge Workspace → PROVE → REPRODUCE THE PROOF. Frozen policy + persisted facts → same verdict, re-derived. Read-only.

REPLAY

Scenario reconstruction — reconstruction, not re-execution. New runs are never offered as historical proof.

OFFLINE

Self-contained HTML report + DATA.json + run portal + submitted ZIP (152 files, SHA-256 1fde8544…).

QUESTION→VERIFY→PERSISTED RESULT→EVIDENCE REF

subodhkc.com/tm-forum — quick-read table, run portal, prompt pack, final report.

10 · Why It Matters

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

Here

3 agents · 3 controls · 6 verdicts · AWS + ServiceNow · one telecom scenario.

Enterprise

Many agents · many controls · multiple runtimes and clouds · evidence across organizational boundaries — same deterministic contract.

Observe. Reconstruct. Evaluate. Prove. Remediate. Prove again.

HAIEC — High assurance in every consequence.