RIG Deviate

Push AI output past the generic median.

🥇 Large language models are gravity wells. Ask for copy, code, strategy, or design and they collapse toward the polished, safe, forgettable median. rig-deviate fights that gravity with 40 named engines and a deterministic scorer for how far you actually moved.

60-second install

pip install rig-deviate

from rig_deviate import deviate, score

seed = "Our product helps teams collaborate better."

result = deviate(seed, "GRAVITON", 10)   # push away from generic gravity
report = score(result)
print(report["rig_l"], report["rig_l_label"])

How it works

_{seed → engine(σ) → deviated artifact → Robust-MAD-Z score → RIG-L grade}

Every engine runs on a ±30σ ladder with 14 anchored rungs — negative σ pulls toward the median, 0σ is the median, positive σ pushes away along that engine's axis. Cognitive and Nature engines operate on a soft ±20σ scale; Physics engines are hard ±30σ state gates, where the negative pole is a BLOCK, not a soft nudge.

-30  -20  -10   -5   -3   -1    0   +1   +3   +5  +10  +20  +30
  |----|----|----|----|----|----|----|----|----|----|----|----|
negative pole              generic median              positive pole

Results: the score system

rig-deviate uses Robust-MAD-Z instead of a normal z-score because the generic-LLM baseline is not Gaussian and is full of outliers:

MAD = median(|x_i - median(x)|)
Robust-MAD-Z = 0.6745 * (score - median(baseline)) / max(MAD, 5.0)

Per-engine σ values combine into a composite RIG-L grade:

RIG-L	σ range	Meaning
block	< 3σ	Still generic or unsafe
marginal	3–5σ	Borderline
review	5–10σ	Promising, needs human review
promote	10–20σ	Strong deviation
doctrine_artifact	≥ 20σ	Exceptional, civilization-grade output

Why it exists

• 40 orthogonal lenses, each tuned to a specific failure mode of generic output
• Fully deterministic — regex and arithmetic only, no network calls, no model inference, no API keys
• CI-safe — suitable for gates, pre-commit hooks, and automated evaluation pipelines
• Measurable, not aesthetic — every deviation ships with a Robust-MAD-Z score against a baseline corpus

The 40 engines

#	Codename	Layer	Full name	Positive pole
1	GRAVITON	Cognitive	Gravity Escape	surprising, category-defying, memorable
2	ANCHOR	Cognitive	Reality Anchor	evidence-dense, source-anchored
3	DARWIN	Cognitive	Evolutionary Selection	iterated, selected, pressure-tested
4	XRAY	Cognitive	Feynman X-Ray	precise, concrete, explainable to a novice
5	FORGE	Cognitive	Mechanism Furnace	mechanism-dense, causal, operational
6	BREAKER	Cognitive	Rupture Engine	contrarian, frame-breaking, orthogonal
7	COLLIDER	Cognitive	Collision Collider	cross-domain recombination
8	VOLT	Cognitive	Voltage Reactor	genuinely felt stakes, no coercion
9	ECHO	Cognitive	Memory Residue	sticky, quotable, durable recall
10	HORIZON	Cognitive	Temporal Horizon Integrity	optionality-rich, reversible bets
11	SOVEREIGN	Cognitive	Autonomy Calibration	agency-respecting, transparent
12	SURPRISE	Cognitive	Predictive Error Calibration	genuinely surprising yet coherent
13	LOOP	Cognitive	Zeigarnik Residue	curiosity loops, serialized intrigue
14	VISCERA	Cognitive	Somatic Marker	physically felt consequences
15	REBOUND	Cognitive	Opponent Process	dynamic contrast, earned resolution
16	PRISM	Cognitive	Signal-to-Noise Discriminability	sharp signal, clean structure
17	WELLSPRING	Cognitive	Hedonic Adaptation Resistance	layered, rewarding revisits
18	GLYPH	Cognitive	Kolmogorov Originality	incompressible, irreducible expression
19	BAYES	Cognitive	Confidence Calibration	appropriately uncertain, well-calibrated
20	SHIELD	Cognitive	Cognitive Sovereignty Shield	AI-augmented, human-gated judgment
21	SWARM	Nature	Pheromone Saturation	diverse, unexplored paths maintained
22	ALBATROSS	Nature	Lévy Flight	occasional long-range exploration
23	SLIME	Nature	Physarum Pruner	efficient, adaptive allocation
24	CLONAL	Nature	Immune Hypermutator	differential mutation by quality
25	LUMINA	Nature	Firefly Attractor	diverse attraction, controlled clustering
26	COLI	Nature	Chemotaxis Climber	gradient ascent with tumble fallback
27	ROOT	Nature	Mycorrhizal Allocator	fair, resilience-preserving allocation
28	HUMPBACK	Nature	Whale Spiral	annealed convergence
29	CUCKOO	Nature	Cuckoo Parasite	disruptive variants pruned or promoted
30	REEF	Nature	Coral Reef Evolver	maximal diversity with selection
31	TUNNEL	Physics	Quantum Tunneling	genuine orthodoxy penetration
32	PAULI	Physics	Pauli Exclusion	state-distinct identity
33	CRITICAL	Physics	Phase Transition	verified regime shift
34	PARSEC	Physics	Fine Tuning	cosmologically precise tuning
35	HAWKING	Physics	Hawking Radiation	information leakage / auditability
36	CASIMIR	Physics	Casimir Effect	deliberate absence produces value
37	KELVIN	Physics	Absolute Zero	honest bounded claims
38	LUMEN	Physics	Speed of Light	latency respects causal chain
39	BELL	Physics	Entanglement	genuine coupled-system effect
40	ZEROPOINT	Physics	Vacuum Fluctuation	healthy baseline variance present

Usage examples

Apply every engine:

from rig_deviate import deviate_all

variants = deviate_all("Our product helps teams collaborate.", sigma=5)
for code, text in variants.items():
    print(f"{code}: {text}")

Score an artifact:

from rig_deviate import score

report = score("Our product helps teams collaborate.")
print(report["rig_l"])        # composite σ
print(report["rig_l_label"])  # block | marginal | review | promote | doctrine_artifact
print(report["weakest_gate"]) # lowest-scoring engine

Use a custom baseline:

from rig_deviate import score

baselines = {
    "GRAVITON": (40.0, 45.0, 48.0, 50.0, 52.0, 55.0, 58.0, 62.0, 65.0, 68.0),
}

report = score("...", baselines=baselines)

Documentation

Resource	Description
CONTRIBUTING.md	Contribution guide
LICENSE	MIT

---

_{Built by Mike Rodgers · Forward Deployed Engineer · rodgersintelligence.com}