Faramesh Tesseract

Pre-governance observation engine for AI agents.

Tesseract is the on-ramp to Faramesh Core governance. It wraps your agent's tool calls in zero-policy observe mode, records everything, analyzes behavioral patterns, and auto-generates a calibrated initial policy — so you can enable enforcement with confidence instead of guessing.

The Problem

Without observation, the path is: guess at policies → deploy → break production → disable governance.

With Tesseract: instrument → observe → auto-generate policy → review → enforce.

Quick Start

go get github.com/faramesh/tesseract

import "github.com/faramesh/tesseract"

// 1. Create an observer
obs := tesseract.NewObserver(100000)

// 2. Record tool calls (wrap your existing tool execution)
obs.Record(tesseract.ObservedCall{
    AgentID:   "payment-bot",
    SessionID: "sess-001",
    ToolID:    "stripe/refund",
    Args:      map[string]any{"amount": 150.00},
})

// 3. Check governance readiness
score := tesseract.CalculateReadiness(obs)
fmt.Printf("Readiness: %d/100 (%s)\n", score.Score, score.Level)

// 4. Auto-generate a policy
gen := tesseract.NewPolicyGenerator(obs)
policy := gen.Generate()
fmt.Print(policy.ToFPL())

Features

Observe Mode

Zero-policy observation of all agent tool calls. Nothing is denied or deferred — everything is recorded for analysis.

• Ring-buffer storage (configurable max calls)
• Per-agent and per-session grouping
• Tool frequency analysis
• Error rate tracking

Governance Readiness Score

Analyzes observed patterns and computes a 0-100 readiness score:

Score: 67/100 ⚠️  Needs Attention
──────────────────────────────────
✓ 23 distinct tools observed
✗ 4 tools have global blast radius
✗ Error rate exceeds 10%

Levels: not_ready → basic → moderate → ready → production

Policy Generation

Auto-generates FPL policy drafts from observed patterns:

• Classifies tools by blast radius (none/local/network/global)
• High-blast-radius tools → DEFER (require human approval)
• High-frequency routine tools → PERMIT with rate limits
• Unknown tools → DEFER (conservative default)
• Each rule includes confidence score and evidence

Drift Detection (planned)

Post-enforcement continuous monitoring for behavioral drift — argument distribution changes, new tools, frequency anomalies.

CLI (planned)

faramesh tesseract observe --agent payment-bot --output ./observations/
faramesh tesseract score --agent payment-bot
faramesh tesseract generate --agent payment-bot --last 7d --output policy.yaml
faramesh tesseract drift --agent payment-bot --since 30d

Privacy

Tesseract stores structural signatures and statistical summaries only — never raw argument values. An observation store can be shared for analysis without leaking production data.

Relationship to Faramesh Products

Product	Layer	Purpose
Tesseract	Pre-governance	Observe → generate policy
Core	L1 enforcement	Deterministic tool call governance
Sverm	L3 detection	Cross-agent behavioral analysis
Hub	Distribution	Policy pack registry
Horizon	Enterprise	Fleet management, compliance, SSO

License

Apache License 2.0 — see LICENSE.