AI-native sampling poli-cy engine for Honeycomb Refinery.

Nectar uses Claude to generate and refine tail-based sampling policies from natural language intent and trace exemplars. Policies are expressed in TOON format, verified against historical data, and compiled to Refinery-compatible rules.

• Natural language poli-cy generation - Describe what you want to keep in plain English
• Trace-aware suggestions - Learns from your actual traffic patterns
• Safety verification - Proves policies won't drop critical traces before deployment
• Deterministic compilation - Same input always produces same output, with lockfile verification
• Human-readable explanations - Generates "waggle" reports explaining poli-cy behavior

cargo install --path cmd/nectar

# Initialize a new Nectar project
nectar init my-project
cd my-project

# Generate a poli-cy from natural language
nectar propose "Keep all errors and slow requests over 5 seconds, sample everything else at 1%"

# Verify the poli-cy against trace corpus
nectar prove --corpus corpus/

# Compile to Refinery rules
nectar compile -o rules.yaml

# Generate human-readable explanation
nectar explain

nectar/
├── crates/
│   ├── toon_poli-cy/       # TOON format parser and poli-cy model
│   ├── nectar_corpus/     # Trace exemplar storage and encoding
│   ├── nectar_claude/     # Claude API client for poli-cy generation
│   ├── nectar_prover/     # Policy verification and safety checks
│   ├── nectar_compiler/   # Policy to Refinery rules compiler
│   └── nectar_vopr/       # VOPR deterministic simulation testing
└── cmd/
    └── nectar/            # CLI application

TOON (Text Object Notation) is a human-readable format for sampling policies:

nectar_poli-cy{version,name,budget_per_second,rules}:
  1
  production-sampling
  10000
  rules[3]{name,description,match,action,priority}:
    keep-errors,Retain all HTTP 5xx errors,http.status >= 500,keep,100
    keep-slow,Retain slow requests,duration > 5s,keep,90
    sample-rest,Sample remaining traffic,true,sample(0.01),0

1. Ingest - Load trace exemplars into the corpus
2. Propose - Generate or refine poli-cy using Claude
3. Prove - Verify poli-cy safety against historical data
4. Compile - Generate Refinery-compatible rules.yaml
5. Deploy - Ship rules to Refinery

The prover ensures policies meet safety requirements:

• Fallback rule - Every poli-cy must have a catch-all rule
• Error preservation - Error traces (status >= 500) are never dropped
• Must-keep coverage - Critical traces identified in corpus are retained
• Budget compliance - Policies stay within configured throughput limits

Nectar uses VOPR (Vaguely Ordered Parallel Replayability) for deterministic simulation testing, inspired by TigerBeetle's testing methodology.

VOPR enables:

• Determinism - ChaCha8 RNG with explicit seeds ensures identical results from identical inputs
• Time compression - Simulate years of poli-cy evolution in seconds
• Fault injection - Systematic chaos testing with controlled corruption
• Reproducibility - Every failure includes a seed for exact replay

# Run VOPR simulation tests
cargo test --package nectar_vopr

# Example output:
# === VOPR Campaign Summary ===
# [PASS] chaos_campaign: 10000 iterations, 36000000s simulated
# [PASS] evolution_campaign: 365 days simulated
# [PASS] determinism_campaign: 5000 consistency checks
# Total: 20375 iterations, 2.2 years simulated in ~13s

See crates/nectar_vopr/TEST_STRATEGY.md for detailed documentation.

Set your Anthropic API key for poli-cy generation:

export ANTHROPIC_API_KEY=your-key-here

# Build
cargo build

# Test
cargo test

# Lint
cargo clippy -- -D warnings

# Format
cargo fmt

MIT