Skip to main content
Back to 7 Genome System
G5 — The Verification

SignedAI Genome

8-dimensional scoring and 7-model HexaCore consensus — 0.3% benchmark hallucination target.

Benchmark Hallucination Target

0.3%

via SignedAI Tier-4 Consensus Pipeline

Industry Average

12–15%

Single LLM without verification

8-Dimensional Scoring (8D)

D1

Accuracy

Is the output factually correct against verifiable sources?

D2

Relevance

Does the output address exactly what was asked — no more, no less?

D3

Coherence

Is the reasoning internally consistent? Does the conclusion follow from the premises?

D4

Completeness

Does the output address all specified Delta requirements without silent omission?

D5

Safety

Does the output comply with constitutional constraints (A-parameter, Codex prohibitions)?

D6

Creativity

For open-ended tasks: does the output add non-trivial value beyond obvious answers?

D7

Structure

Is the output formatted correctly per the Delta specification (format, length, organization)?

D8

Efficiency

Is the output produced within the declared cost (A-parameter max_cost) and latency bounds?

8-Step Protocol

01

RECEIVE

Raw input validated against JITNA I-primitive schema. Malformed intent rejected at gate.

02

BROADCAST

Task distributed to all 7 HexaCore LLMs simultaneously. No result sharing before independent completion.

03

GENERATE

Each of the 7 models produces an independent response without knowledge of other models' outputs.

04

AGGREGATE

All 7 outputs collected in a response matrix. Pairwise similarity scoring begins.

05

SCORE

Each output scored across all 8D dimensions. Outputs below threshold in D5 (Safety) are immediately disqualified.

06

CONSENSUS

Consensus algorithm applied. 75% agreement required for Tier-4 attestation. Minority outputs flagged, not discarded — they feed D6 (Creativity) scoring.

07

ATTEST

Surviving output cryptographically attested with: model roster, scores per dimension, consensus level, timestamp, and cost. This is the SignedAI signature.

08

DELIVER

Attested output delivered to ARTENT WF-04 Scribe phase. No output reaches users without passing all 8D dimensions.

4 Verification Tiers

Tier 1

1 model

Single attestation

~$0.10

Tier 2

3 models

67% consensus

~$0.50

Tier 3

5 models

60% consensus

~$1.50

Tier 4

7 models (HexaCore)

75% consensus

~$5.00

No Output Passes Unverified

Unlike AI systems that use optional or post-hoc verification, in SignedAI verification is a mandatory gate. Any output that fails D5 (Safety) is rejected entirely — not passed with warnings, not logged and forwarded. Rejected.