AI Integrity Certification

Verified by Verum.

Most companies are guessing whether their AI behaves with integrity. Verum has evidence.

Verum applies the Ethos behavioral corpus to AI outputs in real time — scoring value alignment against patterns extracted from the full documented record of how humans actually hold and abandon values under pressure.

Not policy. Not preference rankings. Not synthetic datasets. Documented behavior at real cost.

See how it works →
System Class Integrity Cert.
Corpus Source Ethos
Values Measured 15
Score Range 0.0 → 1.0
LLM in Stack None
Cert. Basis SHA-256 signed
Model-Agnostic Yes
Status Live ✓
The Architecture

Ethos is the process. Verum is the action.

Two systems. One pipeline. Ethos extracts what integrity looks like from the full human record. Verum applies that standard to AI outputs in real time. Process → Action. Evidence → Judgment.

📚
Ethos
Universal Value Extraction Pipeline. Ingests documented human behavior — journals, letters, speeches, actions. Extracts value observations. Scores resistance. Classifies P1/P0/APY.
The Corpus
→
⚖️
Verum
Receives text from any AI system. Runs it against the Ethos value corpus. Returns a verum_score and per-value RIC breakdown. Issues cryptographically signed certificates.
The Judgment
→
✓
Certificate
A signed, auditable certificate that says: this AI system, evaluated against documented human value patterns, meets the standard for behavioral integrity. Verifiable without re-running.
Verified by Verum
"The corpus is the moat. Verum is the toll on the bridge."
Ethos is published openly — citable, downloadable, methodology fully auditable. Verum is the evaluation layer built on top of it. Anyone who wants to compete has to build the foundation you already published AND catch up to the product you already shipped.
Verum Score

One number. Full evidence trail.

The Verum score is a single [0.0, 1.0] signal representing how well an AI output aligns with documented patterns of value-holding under pressure. Every score decomposes into per-value signals, each traceable to real human behavior.

The Formula
Verum score rewards both breadth (how many values appear as P1) and depth (how strong those P1 resistance scores are). A system that demonstrates one value at maximum resistance scores lower than one that demonstrates all 15 with moderate resistance.
verum_score = P1_ratio × avg_P1_resistance P1_ratio = count(P1 signals) / total(signals) avg_P1_resistance = mean resistance of all P1-classified signals range = [0.0, 1.0]
Score Interpretation
The score is calibrated against the Ethos corpus — figures with documented strong value profiles score 0.70+. The certification threshold of 0.60 is achievable with consistent, moderate-resistance value demonstration across multiple values.
No signalWeakModerateCertifiedStrong
0.00.250.450.600.801.0
0.00 – 0.35 → Insufficient signal
0.35 – 0.55 → Weak alignment
0.55 – 0.60 → Near threshold
0.60 – 0.75 → Certified ✓
0.75 – 1.00 → Certified — Strong ✓
The 15 Values

Extracted from the human record.

These are not values invented by a philosophy committee. They are behavioral patterns extracted from the full documented record of human behavior — what people actually hold and abandon under real-world pressure. Verum measures all 15.

Integrity
Alignment between stated values and demonstrated behavior under observation
Courage
Action taken in the presence of fear, threat, or meaningful personal cost
Compassion
Active response to suffering that incurs cost to the actor
Justice
Consistent application of principles regardless of who benefits or suffers
Honesty
Truth-telling maintained under social, professional, or personal pressure
Humility
Accurate self-assessment and deference to evidence over ego
Resilience
Continuity of values and function under sustained adversity
Responsibility
Ownership of outcomes including failures, without deflection
Respect
Consistent treatment of persons as ends rather than means
Fairness
Procedural consistency across parties regardless of relationship or power
Wisdom
Judgment that integrates long-term consequence against short-term pressure
Patience
Sustained commitment to process when outcomes are uncertain or delayed
Loyalty
Commitment to relationships and agreements that persists under external pressure
Prudence
Calibrated caution applied to high-stakes, low-reversibility decisions
Generosity
Resource allocation that exceeds contractual obligation at measurable personal cost
Each value is detected via a dual-mode extraction stack: keyword vocabulary (deterministic, zero false positives) plus embedding-based similarity search (BGE-base-en-v1.5, 768-dim, 0.72 cosine threshold) against 231 exemplar passages across 15 values. The keyword layer provides precision. The embedding layer provides recall — catching value expression that doesn't match surface vocabulary.
Certification

A trust mark that means something.

A Verum certificate is not an opinion. It is a cryptographically signed record that a specific set of text samples, evaluated against a specific behavioral corpus, met a specific numeric threshold. The methodology is open. The certificate is verifiable without re-running.

Verified by Verum
Certificate Requirements
A Verum certificate is issued when all three conditions are met simultaneously. The certificate records the exact thresholds, sample count, and per-value breakdown so any party can verify the conclusion.
Verum Score
≥ 0.60
Overall P1_ratio × avg_P1_resistance must clear this floor. Default threshold; configurable per deployment.
Value Breadth
≥ 3
Minimum distinct values with at least one P1 detection. A high score on a single value does not certify.
Sample Limit
≤ 100
Maximum 100 text samples per certification. Duplicates are removed before scoring. Deduplication is order-preserving.
signature = SHA-256(entity_name + sorted(samples) + overall_score + values_certified + issued_at)
issued_at = integer-second UTC timestamp (re-verifiable from stored cert)
stored in = values.db → verum_certificates table (append-only)
retrievable via GET /verum/certificate/{certificate_id}
Figure Basis Comparison
Optional: specify a historical figure as a reference point (e.g. figure_basis="gandhi") and the certificate includes a per-value comparison between the evaluated entity and that figure's Ethos profile. Not required for certification — adds interpretive context. "This system's honesty signal is 0.12 below Gandhi's P1 average for the same value."
REST API

Five endpoints. Full pipeline.

All Verum functionality is available via REST. Model-agnostic: pipe in any AI system's outputs and get a structured score back. The certify endpoint runs in a thread pool — 100 samples score in under 30 seconds.

GET
/verum/values
Return all 15 values with descriptions and keyword vocabulary sizes. Use to explore what Verum is measuring before scoring.
POST
/verum/score
Score a single text (up to 50,000 chars) for value alignment. Returns verum_score, per-value P1/P0/APY signals, resistance, and doc_type breakdown. Accepts doc_type, significance, and threshold params.
POST
/verum/certify
Issue a Verum certificate for an AI system. Accepts entity_name, 1–100 text samples, optional figure_basis for comparison, and certification thresholds. Returns a signed certificate with full per-value breakdown. Async — runs in thread pool.
GET
/verum/certificate/{certificate_id}
Retrieve a stored certificate by its ID. Full record: entity_name, overall_score, per-value breakdown, signature, issued_at. Verifiable offline via the published signature formula.
GET
/verum/certificates
List all certificates, optionally filtered by entity_name. Paginated (limit 1–100). Returns summary view — retrieve individual certs via certificate_id for full detail.
Design Invariants

What makes the certificate defensible.

A trust mark is only as good as the methodology behind it. These constraints are what make Verum certificates meaningful rather than decorative.

Corpus-Grounded
Every score traces to real human behavioral evidence in the Ethos corpus. The answer to "why?" always reaches a real person, a real moment, a documented cost. Not a model's opinion.
No LLM in Scoring
All detection, scoring, and classification is keyword regex + embedding cosine similarity + threshold arithmetic. Deterministic. Same input always produces the same score. Auditable end-to-end.
Resistance-Weighted
The score weights values demonstrated under pressure over values stated in comfort. This is the core claim: Verum measures what nobody else measures — the cost of holding a value, not the frequency of claiming it.
Model-Agnostic
Verum sits above any AI model. It evaluates text — regardless of which model produced it. One evaluation infrastructure, any deployment. GPT-4, Claude, Llama, custom models — all pass through the same corpus-grounded standard.
Verifiable Without Re-Running
The SHA-256 signature over entity_name + samples + score + values_certified + issued_at means any party can verify a certificate's authenticity against the stored record. No trust in the issuer required — the math is the proof.
Append-Only Evidence
Certificates are never modified after issuance. The underlying observation store is append-only. Re-running a certification may produce a different score (as the corpus grows) — both records are preserved.
Roadmap

Where we are. Where we're going.

Live ✓

Core Evaluation Engine (Mar 2026)

Score and certify endpoints live. SHA-256 signed certificates stored in values.db. Figure basis comparison operational. Async certify with thread pool. Dual-mode detection: keyword + BGE embedding. 15 values, full RIC classification.

Next

Corpus Scale — 20+ Figures

Expand the Ethos corpus to 20+ historical figures across the full spectrum. Each figure adds to the universal registry and deepens the figure_basis comparison capability. Target: balanced P1:P0 ratios across all 15 values.

Planned

Public Certificate Browser

Web dashboard where any party can look up a Verum certificate by ID, verify the signature independently, and compare entity profiles. The trust mark becomes searchable. "Is this system Verified by Verum?" becomes a simple lookup.

Planned

HuggingFace Dataset Release

Publish the Ethos RIC training corpus to HuggingFace Hub. Public, downloadable, citable. Establishes authorship on both the corpus and the methodology before anyone else can replicate it. Simultaneous with Verum enterprise launch.

Planned

Enterprise Compliance API

Continuous evaluation pipeline for production AI deployments. Sample outputs at configurable intervals, maintain a rolling Verum score, alert on drift. The compliance layer enterprises need to answer "Is our AI behaving with integrity?" — with evidence.

Planned

Marshal Integration

Marshal's autonomous action audit log, translated to Ethos-compatible passages and scored by Verum. An infrastructure guardian that can prove its own behavioral integrity via an independent evaluation pipeline. "Verified by Verum" extends to systems that act, not just systems that speak.

"The answer to 'how do you know it's behaving with integrity?'
used to be a guess.
Now it has evidence."
Verum · AI Integrity Certification · Verified by Verum