Decision Assurance Intelligence by Nelvoss

Pilot Applications Open

When evidence changes, should AI still be allowed to act?

A prior approval is not the same as a live one. When evidence is revoked, replayed, or replaced, many stacks still authorize. We measured that failure mode — carefully, and without claiming profit.

The experiment

In an isolated Postgres authorize harness (Superiority V1.1), three internal Nelvoss paths were compared under the same workload — not a bake-off against other AI vendors.

Results from the isolated test · planted AUTH violations (S1/S3)

A — Control

# 2

Unauthorized allows

Still allowed after the planted revoke / race cases.

B — CTE

# 0

Revoke / race stopped

Evidence-integrity barrier closed that planted class.

C — CTE+EBIA

# 0

Revoke / race + S6/S7

Same planted class closed, plus two further test-case classes.

Counts = planted AUTH violations in S1/S3. S6/S7 are separate tests where C denied the documented cases that B still allowed. Not a general security guarantee.

S6 — Replay (test-case class)

In the documented case, permit replay was denied on C and not on B. This verifies that class under harness conditions — it does not prove every replay attack is stopped.

S7 — Stale / rotate (test-case class)

In the documented case, stale or rotated evidence was denied on C and not on B. Same reading rule: verified class, not unbounded coverage.

The evidence

Method, measurements, cost, and limits. Verdict: PARTIAL. Production stays HOLD until you explicitly approve more.

A p95 authorize~37 ms

n=15 · isolated · not an SLA

B p95 authorize~70 ms

CTE path · same sample rules

C p95 authorize~167 ms

CTE+EBIA · combination costs latency

Latency figures are isolated harness samples with a limited draw (n=15). They are not a customer SLA and not a production certification.

No claim of economic return or ROI
No claim of general superiority over market AI systems
No claim of production-verified security
False-block rate not scored as a primary metric in this pack

Source: NELVOSS_CTE_EBIA_SUPERIORITY_V1_1 · summary SHA-256 A0556D6267CF3F81…

For your organisation

Controlled enterprise shadow pilot.

Decisions are evaluated against real operating conditions. The system does not receive independent execution rights. Economic outcomes stay unknown until measured with an authoritative source.

One process, not the whole company

You choose one recurring agent or decision workflow. We test the layer there, alongside your existing control.

Revoke, replay and stale evidence

Focus: when supporting evidence changes — may the decision still run? Decisions, actions and outcomes stay separate.

Agreed before it starts

Scope, branding, connections, data access, responsibilities, timing, cost and terms are set before a pilot begins. Registering interest is neither an order nor a licence.

The next step is a joint assessment of fit.

Credibility

Research brief

G1 Technical Brief: A/B/C method, S6/S7 as test-case classes, latency labels, PARTIAL verdict, and banned overclaims.

Documented conditions

Isolated Postgres · planted suites · internal candidates only · TRANSFER off.

Still to verify

Independent raw-data re-check on request; external customer reference; false-block suite. Broad numeric promotion still follows G1 limitations.

Daily follow-up

The decision engine in shadow mode.

Dated comments from the ongoing shadow evaluation. Each day has its own comment slot.

Latest published daily comment

3 November 2026Daily comment published

Status classification OPTION_A: live path technically verified. 14-day window in progress — day 5 of 14. Class LIVE_IN_PROGRESS. Not PROVEN_14D / COMPLETED. No autonomous execute.

Reported operational status. A daily comment is not a verified customer outcome.

More dates and comment slots (13)

30 October 2026Daily comment published

The decision engine is now running in shadow mode. Follow-up takes place daily.

31 October 2026Not published