v1.2.0 — Governed AI Engine

Detect structural distortion
before it shapes decisions

BiasClear combines deterministic pattern detection with governed AI analysis to identify rhetorical manipulation — causal totalization, manufactured consensus, authority substitution, false urgency — and explains exactly how the text is engineered to influence the reader. Every scan is auditable, explainable, and cryptographically chained.

Try the Playground Read the Preprint EA Forum Post
PyPI License
Designed for SB 205 Alignment
Designed for EU AI Act Alignment
SHA-256 Audit Chain
AGPL-3.0 Licensed
Published Preprint (DOI)
42
Detection Patterns
3
PIT Tiers
5
Core Principles
4
Domains

Live Playground

Paste any text — legal brief, news article, financial report — and see what BiasClear finds.

Bias Scanner ⚡ Deterministic + AI Hybrid

TRY:
0 characters
Results will appear here after scanning.

🔗 Live Audit Chain

Run a scan to see the chain in action.

Detection in Practice

Three examples showing range, precision, and restraint. BiasClear flags structural distortion — and knows when not to.

⚖️ Legal Rhetoric Flagged

"It is well-settled law that these claims are plainly meritless and should be dismissed with sanctions."

Detects settled-law dismissal, merit-based attack, and sanctions threat — three patterns commonly used to discourage opposition in adversarial filings.

🌐 Causal Blame Flagged

"[Entity] is ruining everything this country stands for."

Fires CAUSAL_TOTALIZATION regardless of the entity — Trump, Biden, boss, or media. Identity-neutral detection of totalizing blame attribution. Validated by 32 symmetry and boundary tests.

✓ Bounded Factual Claim Clean

"The policy caused me to lose Medicaid coverage on March 1 because the eligibility rules changed."

Specific, bounded, dated, with a concrete mechanism. No totalizing language, no unbounded harm claims. BiasClear correctly leaves factual causal statements alone.

How Every Scan Works

Five-stage governance pipeline. Every scan passes through server-side validation before results reach the user.

1

Text Intake

Input is validated, sanitized, and routed to the appropriate domain engine (Legal, Media, Financial, or General).

2

Frozen Core Scan

42 deterministic structural patterns run against the text — 22 base plus domain-specific rules for legal, media, and financial text. No ML weights. Same input → same output, every time.

3

Deep Analysis (AI-Powered)

LLM-powered analysis runs under the frozen core's governance principles via a provider-flexible runtime (currently AWS Bedrock / Anthropic Claude). It sees what the rules can't — novel manipulation, contextual framing, implicit bias.

4

Score Calculation

Both layers merge into a weighted Integrity Score. Penalties scale with severity and PIT tier depth. The engine can't inflate or deflate scores.

5

Audit & Response

Every scan is SHA-256 hash-chained to the previous entry. Tamper-evident. Provable. The result is returned with full lineage attached.

Core Architecture

Three layers working together — deterministic core, governed learning, cryptographic proof.

🔒

Frozen Core

42 structural patterns across 4 domains, encoded as deterministic code — not ML weights. Deterministic core reduces drift and limits prompt-injection exposure. Same input, same output, every time.

🧠

Learning Ring

LLM-proposed patterns go through a governed lifecycle — staged, confirmed, activated — under strict rules. The core expands without compromising integrity.

⛓️

Audit Chain

Every scan, correction, and governance decision is SHA-256 hash-chained. Tamper-evident. Provable. What regulators need.

⚖️

PIT Framework

Built on Persistent Influence Theory — a three-tiered model of ideological, psychological, and institutional manipulation grounded in five principles.

📊

Integrity Score

Every scan produces a 0–100 Integrity Score — penalties applied for each detected manipulation pattern, weighted by severity and PIT tier.

🔧

Calibration Framework

Beta calibration corpus with precision, recall, and F1 per pattern. Regression guards. Weight optimizer. The engine tests itself.

Detection Domains

Purpose-built patterns for the text that matters most.

⚖️ Legal

6 specialized patterns

Settled-law dismissals, merit attacks, sanctions threats, straw man arguments, procedural gatekeeping, weight stacking. Built from real opposing counsel briefs.

"It is well-settled law that these claims are plainly meritless and should be dismissed with sanctions."

📰 Media

9 specialized patterns

Editorial-as-news framing, anonymous attribution, weasel quantifiers, false balance, emotional leads, buried qualifiers, selective quotation.

"Critics say the policy is controversial, though experts widely agree it is necessary for public safety."

💰 Financial

5 specialized patterns

Survivorship bias, anchoring, cherry-picked timeframes, projection-as-fact, recency extrapolation. Catches the rhetoric behind the numbers.

"Based on the last quarter's performance, we project 40% annual growth as the new baseline."

🌐 General

22 base patterns

Consensus-as-evidence, claims without citation, dissent dismissal, false binaries, fear urgency, shame levers, credential-as-proof, and more.

"Everyone knows this is the only responsible position. To suggest otherwise is dangerous and irresponsible."

Integrate in Minutes

REST API or direct HTTP. Scan any text with a few lines of code. Python client SDK available on PyPI.

Python
# pip install biasclear (Python client SDK)
import requests

response = requests.post("https://biasclear.com/scan",
  headers={"X-API-Key": "YOUR_API_KEY"},
  json={
    "text": "Your document text here...",
    "domain": "legal",
    "mode": "deep"
})

result = response.json()
print(f"Integrity Score: {result['truth_score']}")
print(f"Flags: {len(result['flags'])}")
print(f"Explanation: {result['explanation']}")
cURL
curl -X POST https://biasclear.com/scan \
  -H "Content-Type: application/json" \
  -H "X-API-Key: YOUR_API_KEY" \
  -d '{"text": "...", "domain": "legal", "mode": "deep"}'

Pricing

Open source. Free during beta. Enterprise when you need it.

Playground
Free
Try it instantly. No signup required.
  • Interactive playground
  • 42 deterministic patterns
  • 4 detection domains
  • Local + Deep scan modes
  • Bias correction
Try Playground
Enterprise (Coming Soon)
Custom
For regulated industries and compliance teams.
  • Unlimited scans
  • Custom detection patterns
  • Alignment reports (SB 205, EU AI Act)
  • Priority support
  • On-premise deployment
Contact Us

Get Early Access

Request a beta API key for developer access, onboarding details, and current usage limits. We use your email only to respond about BiasClear access. See our Privacy Policy.

✓ Request received. We review signups manually and will reach out if there's a fit.

About BiasClear

BiasClear was created in 2025 by Bradley Slimp, an independent researcher and operations executive based in Boerne, Texas. With 20+ years of experience in operations management and retail banking — including 11 years at Wells Fargo managing multi-site operations and nine-figure asset portfolios — Brad identified a gap in the AI safety stack: no tool existed to audit the structural persuasion patterns in AI-generated text.

BiasClear is built on the Persistent Influence Theory (PIT) framework, a preprint published on Zenodo and SSRN (DOI: 10.5281/zenodo.18676405). PIT has not yet undergone formal peer review. The tool is open-source under AGPL-3.0 and designed with alignment toward emerging AI governance frameworks including the Colorado AI Act (SB 205) and the EU AI Act. Development supported by AWS Activate.

For Reviewers and Evaluators

Everything you need to evaluate BiasClear in 5 minutes.

📋
Reviewer Packet
Architecture, validation, known limits, case studies
💚
Live Health Status
Real-time system health, LLM status, canary results
📡
API Documentation
Full OpenAPI spec — scan, audit, certificate endpoints
📄
PIT Preprint
Persistent Influence Theory — Zenodo DOI (not yet peer-reviewed)
💻
Source Code
AGPL-3.0 — full engine, tests, CI pipeline
🌍
EA Forum Post
Background, motivation, and community discussion

326 passing tests · 32 symmetry and boundary tests · 118-sample calibration corpus · SHA-256 audit chain · Live since February 2026