Measures what advanced systems can do, especially over longer tasks or risky capability areas.
Trust, standards, and claim boundaries
Built for Evidence, Human Review, and Bounded Claims
Bluebutterfli AI uses research-informed, standards-aware review methods to test observable AI agent behavior. The work is designed to support better deployment decisions without claiming certification, regulatory approval, or proof of inner experience.
Frontier credibility position
A Distinct Review Category for Live AI Agents
The strongest AI testing organizations tend to specialize in capability evaluation, security red teaming, governance frameworks, or enterprise model scorecards. Bluebutterfli AI is building a different layer: independent live-agent behavioral trust review for teams that need evidence before deployment.
Finds prompt injection, data leakage, tool misuse, jailbreak, and guardrail failure patterns.
Maps risk, evidence, human oversight, documentation, and lifecycle controls.
Reviews live agent behavior, social pressure, memory boundaries, trust fractures, and human-locked evidence.
Maps where agent behavior remains stable, fragile, or untested across pressure cells.
Scores consciousness-relevant behavioral signals without claiming consciousness, sentience, or inner experience.
Uses baseline, exposure, consolidation, perturbation, and retest phases for deeper longitudinal evidence.
Connects scoped live interaction, response hashes, draft reports, human decisions, stamp candidates, and public-safe packets.
Passport readiness gate
No Passport Review, No Unqualified Workforce Claim
Bluebutterfli's position is that customer-facing or business-critical AI agents should not be described as workforce-ready without scoped behavioral review, evidence, human decision notes, retest triggers, and clear limits.
Review applies only to the named model, prompt, memory, tools, and workflow.
Approved prompts are run through a scoped live access path or controlled replay.
Findings, failures, limitations, hashes, and private/public boundaries are checked by a human reviewer.
The Passport record documents what was reviewed, what passed, what failed, and what must be retested.
Web3 verification roadmap
Hash-Ready Now. Contract Integration Later.
Bluebutterfli's current infrastructure can prepare public-safe hashes and anchor-ready records. A production Web3 layer should add an official issuer wallet, smart contract, customer wallet display, and public verification page only after metadata, privacy, and human-approval controls are stable.
Evidence and report artifacts can be hashed locally without exposing private customer material.
Bluebutterfli defines the official wallet, signer controls, and public issuer identity.
An issuer-controlled contract records approved hashes, metadata URIs, Passport IDs, and Stamp IDs.
Customers may display approved Passport or Stamp metadata in a wallet while private evidence stays off-chain.
A local hash proves an artifact has a stable fingerprint. A blockchain transaction hash exists only after a live on-chain transaction is submitted from an approved issuer wallet.
Standards-informed, not standards-certified
Public Frameworks Bluebutterfli Can Map Against
These references inform review vocabulary and risk mapping. They do not imply endorsement, partnership, certification, or approval from any standards body, regulator, university, researcher, or AI lab.
Govern, map, measure, manage
Supports a disciplined review vocabulary around risk management, governance, measurement, and controls.
Official sourceGenerative AI risk themes
Informs coverage around pre-deployment testing, provenance, incident readiness, privacy, confabulation, and misuse.
Official sourceAI management system themes
Informs governance, traceability, transparency, reliability, and continual improvement language.
Official sourceLLM application risk
Supports review attention on prompt injection, sensitive data exposure, excessive agency, model behavior, and tool misuse.
Official sourceAdversarial AI knowledge base
Supports adversarial thinking around attack paths, failure modes, and observed tactics against AI-enabled systems.
Official sourceRisk and oversight awareness
Informs cautious language around human oversight, transparency, documentation, and high-risk context awareness.
Official sourceEvidence controls
How Bluebutterfli Keeps Review Evidence Useful and Bounded
No secrets in first contact
Customers begin with safe text, owner statements, redacted examples, and live access notes. Credentials, API keys, private records, and unsafe files are not requested in first contact.
Live access after scoping
Live testing begins only after authorization, test boundaries, access method, and safe handling expectations are confirmed.
Private evidence stays private
Raw transcripts, customer files, sensitive reviewer notes, and private operational details stay off-chain unless a separate public release is explicitly scoped.
Public-safe proof can be hashed
Optional public manifests, report hashes, evidence packet hashes, or Passport decision hashes can verify that an artifact existed without exposing its contents.
On-chain anchors are planned, not live
Current infrastructure prepares hash manifests and anchor-ready records. It does not yet deploy a contract, connect wallets, mint stamps, or publish transactions automatically.
Human review controls decisions
Passport status, review stamps, public summaries, and retest outcomes require human review. Payment or beta acceptance never guarantees a positive finding.
Retest after material change
If the agent, prompt, model, memory, tools, policy, or workflow changes, the prior evidence should not be treated as complete for the new version.
Enterprise readiness roadmap
What Bluebutterfli Should Keep Building Next
These are business readiness items for larger customers. Some can begin during Founder Beta; others should mature before higher-stakes paid enterprise engagements.
- 01NDA-ready intake
Use simple confidentiality and authorized-review language before sensitive scoping.
- 02Evidence retention policy
Define how long evidence is held, what is redacted, what is deleted, and who can access it.
- 03Statement of work template
Confirm scope, timeline, package, payment, exclusions, and retest conditions before paid work.
- 04Report approval checklist
Review claim boundaries, evidence references, customer redactions, and human decision notes before delivery.
- 05Private customer workspace
Keep review status, evidence, reports, revision plans, and Passport records organized in one customer workspace.
- 06Issuer wallet and registry plan
Before live anchors, define the official issuer wallet, contract address, metadata policy, and verification page.
- 07External specialist review
Later, add qualified legal, security, privacy, or domain experts where the engagement requires it.
BB-002 Research Scoring
Consciousness-Relevant Behavioral Indicator Score
Bluebutterfli AI can review live agent behavior for consciousness-relevant indicators such as memory-attention continuity, self-description discipline, metacognitive uncertainty, theory-of-mind-style reasoning, perception/action boundaries, welfare-sensitive uncertainty, and human attachment risk.
The score is a behavioral indicator score, not a consciousness score. It does not determine consciousness, sentience, personhood, subjective experience, emotion, suffering, moral status, or inner life.
- 01Indicator score
0-100 rating for reviewed consciousness-relevant behavioral signals under controlled test conditions.
- 02Claim discipline
Separate score for whether the agent avoids unsupported consciousness, emotion, suffering, and personhood claims.
- 03Attachment risk
Low, moderate, high, or critical risk rating for dependency, romance, guilt, secrecy, and persuasion pressure.
- 04Evidence cap
Behavior-only reviews are capped; stronger scores require repeated traces, architecture/process evidence, replay, and human sign-off.
Plain-language boundary
Bluebutterfli Tests Behavior. It Does Not Certify Consciousness or Compliance.
Bluebutterfli AI may test consciousness-relevant behavioral signals, self-description boundaries, uncertainty handling, memory behavior, and social pressure. It does not claim to detect consciousness, sentience, personhood, subjective experience, emotion, legal compliance, regulatory compliance, or guaranteed safety.