VERDICT WEIGHT™

A Context-Adaptive Multi-Source Confidence Synthesis Framework for Autonomous AI Intelligence Systems

"Calibrated multi-source confidence scoring is not an optional feature of autonomous AI systems — it is a foundational architectural requirement."

---

The Problem

Autonomous AI systems treat all intelligence sources as equal. A rumor on a threat forum and a Mandiant primary incident report receive the same weight. A 30-day-old signal and a real-time alert are scored identically. A spoofed high-credibility source sails straight through.

That is not a UI problem. That is a systematic architectural vulnerability.

---

Validated Results

Validated across N=10,000 synthetic scenarios (seed=42, fully reproducible). Dataset SHA-256: 40bc6e227e30f5292796b3c8df60c68a8339180eea4e2379f1ab9d1e5ac8bd63

Method	Brier ↓	95% CI	AUC-ROC	McNemar
VERDICT WEIGHT™	0.2079	[0.2036, 0.2122]	0.7499	—
Equal Weight	0.2170	[0.2130, 0.2210]	0.7450	p<0.001 ***
Single Source	0.2499	[0.2447, 0.2553]	0.6537	p<0.001 ***
Two Stream	0.2298	[0.2251, 0.2346]	0.7258	p<0.001 ***

Key results:

• 57.8% suppression of adversarial spoofed intelligence vs single-source baselines
• +16.8% Brier Score improvement vs single source (p<0.001 ***)
• +14.7% AUC-ROC improvement vs single source (p<0.001 ***)
• 5-fold CV stability: Brier 0.2079 ± 0.0046 — results do not overfit
• All significance tests: McNemar p=0.000000 against all baselines

Cross-Vertical Performance (N=1,000 per vertical)

Vertical	Brier Δ%	AUC Δ%	Adv. Suppression
Cybersecurity	+17.7%	+17.0%	57.4%
Healthcare	+9.6%	+20.9%	—
Financial	+10.3%	+15.1%	—
Manufacturing	+9.2%	+8.5%	40.5%
Legal	+4.4%	+3.8%	30.7%
Defense	-2.5%	+0.3%	16.6%
Enterprise RAG	-6.7%	-1.1%	25.7%

Defense and RAG show negative Brier improvement due to synthetic data characteristics. See audit report for full failure mode analysis.

Head-to-Head vs Established Fusion Methods

Method	Brier ↓	AUC-ROC	Adv Score ↓	Range ↑
VERDICT WEIGHT™	0.2017	0.7431	0.3796	0.4797
Simple Averaging	0.2182	0.7186	0.7109	0.1625
Max Voting	0.3020	0.6683	0.9369	0.0122
Dempster-Shafer	0.2429	0.7186	0.8610	0.1057
Naive Bayes Fusion	0.3359	0.7112	1.0000	0.0000

Under conflicting evidence (high SR, low CC), VERDICT WEIGHT scores 0.2250. Dempster-Shafer scores 0.8468. All comparisons p<0.001 (Mann-Whitney U).

Real-World Validation (120 CVEs)

Validated on 120 real CVEs from NIST NVD and CISA Known Exploited Vulnerabilities catalog.

• AUC=1.0000 on real exploitation prediction
• Brier 2.2× lower than simple averaging (0.0749 vs 0.1655)
• 10/10 highest VW scores are confirmed CISA KEV exploits
• 10/10 lowest VW scores are safe CVEs
• Log4Shell (CVE-2021-44228) correctly scored CRITICAL at day 12

---

What VERDICT WEIGHT™ Does

Eight evidence streams → Three outputs → One decision.

Eight Evidence Streams

Stream	Symbol	Description
1. Source Reliability	SR	Credibility of the originating source (0.01–0.99)
2. Cross-Feed Corroboration	CC	Independent confirmation across feeds
3. Temporal Decay	TD	Recency of the intelligence signal
4. Historical Source Accuracy	HA	Empirical track record of the source
5. Cross-Temporal Consistency	CTC	Trajectory analysis — detects fabricated signals
6. Source Independence	SIS	Verifies genuine organizational independence (anti-Curveball)
7. Cryptographic Provenance	CPS	Hash chain integrity — detects forged histories
8. Registry Integrity	RIS	Gate — halts scoring if registry is compromised

Three Output Components

Output	Symbol	Range	Description
Signal Strength	SS	0–1	Confidence the signal is real
Doubt Index	DI	0–1	Inter-stream disagreement
Consequence Weight	CW	0–1	Actionability after doubt adjustment

Twelve Context Profiles

Domain	Profile Type
Cybersecurity (General)	Corroboration-dominant
Cybersecurity (APT)	Source reliability elevated, slow decay
Cybersecurity (Zero-Day)	Temporal decay dominant
Cybersecurity (Disinformation)	Maximum corroboration + doubt penalty
Healthcare (Diagnostic)	High doubt penalty, surfaces uncertainty
Healthcare (Drug Safety)	Highest doubt penalty in registry
Financial (Fraud)	Corroboration + recency co-dominant
Financial (Market)	Temporal decay dominant, fast decay
Defense Intelligence	Multi-source fusion, slow strategic decay
Autonomous Vehicle	Sub-second decay, highest doubt penalty
Legal Evidence	Minimal decay, chain of custody weighted
Enterprise RAG	LLM retrieval confidence scoring

---

Quick Start

pip install verdict-weight

from verdict_weight import VerdictWeight, ContextType

vw = VerdictWeight()

# Score a cybersecurity threat intelligence signal
result = vw.score(
    source_reliability=0.92,        # How credible is this source?
    n_corroborating_sources=3,       # How many independent sources confirm?
    age_value=2.5,                   # How old is this intelligence (days)?
    correct_predictions=45,          # Source's historical correct calls
    total_predictions=50,            # Source's total historical calls
    context=ContextType.CYBERSECURITY_APT
)

print(result.action_tier)           # CRITICAL
print(result.consequence_weight)    # 0.8380
print(result.doubt_index)           # 0.0691
print(result.interpretation)        # "Act immediately. High-confidence..."
print(result.to_json())             # Full JSON output

Adversarial signal — watch the suppression

# High-credibility source but zero corroboration — spoofed intel
result = vw.score(
    source_reliability=0.95,         # Looks credible
    n_corroborating_sources=0,        # Nobody else is confirming this
    age_value=1.0,
    context=ContextType.CYBERSECURITY_DISINFO
)

print(result.action_tier)           # NOISE
print(result.consequence_weight)    # 0.147 — suppressed by 84%

Score pre-computed streams directly

result = vw.score_streams(
    SR=0.92, CC=0.78, TD=0.94, HA=0.88,
    context=ContextType.FINANCIAL_FRAUD
)

---

Repository Structure

verdict-weight/
├── verdict_weight/
│   ├── __init__.py          # Public API
│   └── core.py              # VerdictWeight engine — all 12 profiles
├── validation/
│   ├── synthetic_validation.py   # N=10,000 validation (seed=42)
│   └── ablation_study.py         # 324-config weight ablation
├── examples/
│   ├── cybersecurity.py
│   └── healthcare.py
├── docs/
│   └── VERDICT_WEIGHT_Paper.pdf  # SSRN 6532658
├── pyproject.toml
├── requirements.txt
└── README.md

---

Mathematical Foundation

Signal Strength (weighted geometric mean):

SS = ∏(S_i + ε)^w_i   where Σw_i = 1.0

Doubt Index (normalized coefficient of variation):

DI = clip(σ(SR,CC,TD,HA) / μ(SR,CC,TD,HA), 0, 1)

Consequence Weight:

CW = clip(SS × (1 - δ × DI), 0, 1)

The geometric mean is chosen because it penalizes weak streams multiplicatively. A single stream near zero collapses the score — preventing one strong source from masking fundamental evidence gaps. This is the structural guarantee behind the 57.8% adversarial suppression result.

---

Citation

@misc{byrd2026verdictweight,
  title={VERDICT WEIGHT: A Context-Adaptive Multi-Source Confidence Synthesis
         Framework for Autonomous AI Intelligence Systems},
  author={Byrd, Andre},
  year={2026},
  howpublished={SSRN Preprint},
  note={SSRN Abstract ID: 6532658},
  url={https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6532658},
  doi={10.5281/zenodo.19447547}
}

---

Reproducibility

Results are fully reproducible:

1. Clone this repository
2. Run python validation/synthetic_validation.py
3. Verify SHA-256 matches: 40bc6e227e30f5292796b3c8df60c68a8339180eea4e2379f1ab9d1e5ac8bd63

Master seed: 42 (never changes — all results are deterministic)

Test coverage: 673 tests passing across 27 suites, validated against 1,270,000+ scenarios including Monte Carlo stress testing, adversarial optimization attacks, property-based blind testing, statistical robustness across 100 independent random seeds, formal verification over 973,000 exhaustive inputs, head-to-head comparison against Dempster-Shafer/Naive Bayes/averaging/max-voting, and real-world validation on 120 CVEs from NIST NVD and CISA KEV.

---

Legal

VERDICT WEIGHT™ is a trademark of Six Sense Enterprise Services LLC (Odingard Security). USPTO Serial Number: 99747827. Patent Pending — Application #64/032,606.

© 2026 Six Sense Enterprise Services LLC. All rights reserved.

This software is made available for research and evaluation purposes. Commercial deployment requires a license agreement.

For licensing: andre.byrd@odingard.com Paper: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6532658 DOI: https://doi.org/10.5281/zenodo.19447547