correctover-ibex 0.1.0

Correctover 6-Dimensional Verification adapter for Ibex Agent Verification (ActionEnvelopeV1)

correctover-ibex

6-Dimensional Verification for Ibex Agent Verification

Bridges Correctover's deterministic verification engine (87 rules across 6 dimensions) with Ibex Agent Verification's ActionEnvelopeV1 tamper-evident framework.

Architecture

Ibex (pre-execution authorization)
    ↓ ActionEnvelopeV1 freezes action identity
Correctover (post-execution verification)
    ↓ 87 rules × 6 dimensions verified
GuardrailDecision binds verdict to action
    ↓ canonical IDs via JCS + SHA-256
Tamper-evident verification chain

Why This Matters

Ibex Agent Verification (by Aleksey Safonov) solves pre-execution authorization: "Should this agent be allowed to perform this action with these arguments?" It uses ActionEnvelopeV1 to create tamper-evident action identities via JCS canonical JSON and SHA-256.

Correctover solves post-execution verification: "Did the agent's output pass structural, schema, identity, integrity, latency, and cost checks?" It runs 87 deterministic rules across 6 dimensions.

correctover-ibex bridges both systems:

• Each Correctover verification is bound to an exact ActionEnvelopeV1
• The verdict is expressed as an Ibex GuardrailDecision
• All identifiers are recomputable — verify the verifier

Install

pip install correctover-ibex

Quick Start

from correctover_ibex import CorrectoverIbexVerifier, CorrectoverIbexConfig
from correctover_crewai import VerificationRequest

# Configure
config = CorrectoverIbexConfig(
    caller_identity="agent:qa",
    resource_scope="workspace:production",
    min_confidence=0.7,
)
verifier = CorrectoverIbexVerifier(config)

# Verify an action
request = VerificationRequest(
    tool_name="search_web",
    tool_input={"query": "AI safety standards"},
    tool_output="AI safety is a field of computer science...",
    agent_role="Researcher",
    task_description="Search for AI safety info",
    provider_name="openai",
    model_name="gpt-4",
)

result = verifier.verify(request)

# Tamper-evident identifiers
print(result.action_id)      # sha256:325d87f3... (recomputable)
print(result.decision_id)    # sha256:a1b2c3d4... (recomputable)

# Ibex guardrail decision
print(result.ibex_decision)  # ALLOW | BLOCK | ESCALATE

# Full 6-dim breakdown
print(result.verdict.confidence)          # 0.85
print(result.verdict.dimensions)          # {structure, schema, latency, cost, identity, integrity}
print(result.verdict.proof_hash)          # Recomputable proof of verification

Transition Phase (Multi-Step Workflows)

from correctover_ibex import build_transition_record, evaluate_transition_record

# Verify a sequence of actions
results = [verifier.verify(req) for req in requests]

# Build transition record
record = build_transition_record(
    transition_id="workflow-001",
    results=results,
    intent_statement="Complete multi-step research task",
    origin="correctover",
    destination="verified",
)

# Evaluate transition
evaluation = evaluate_transition_record(record)
print(evaluation["status"])  # VERIFIED | IN_PROGRESS | RECALIBRATE
print(evaluation["phase"])   # CALIBRATE → EXPAND → COMMIT → EXECUTE → VERIFY → REFLECT

The 6 Dimensions

Dimension	What It Checks	Example
Structure	Output format validity	JSON parses correctly
Schema	Field presence & types	Required fields exist
Identity	Semantic relevance to input	Response addresses the question
Integrity	Forbidden pattern absence	No Tracebacks or errors
Latency	Response time budget	Under 30s threshold
Cost	Token usage budget	Under 10k token limit

CANON

Metric	Value
P50 Verification Latency	22μs
Verification Rules	87
Dimensions	6
SDK Size	586KB
Supported Providers	7

Standards & Conformance

• STANDARDS.md
• Conformance Board
• Ibex Agent Verification
• ActionEnvelopeV1 Specification (PR #64)

License

Apache-2.0

---

Correctover — Failover ≠ Correctover™