SIGMA by Invariant

We contained O(n3).
Then we made edits nearly free.

SIGMA decomposes knowledge graphs into bounded cells via sheaf cohomology. Streaming edits cost 63 microseconds at one million vertices. Contradiction queries cost 13 microseconds. The scaling exponent is 0.19. No GPU. One laptop.

63 usper edit (V=1M)
13 usper query (V=1M)
0.19scaling exponent
0GPU required
scroll
The Insight

The cube didn't disappear. It got imprisoned inside a constant.

Spectral analysis on knowledge graphs requires eigensolving matrices that grow as O(n3). At 21,000 vertices, the sheaf Laplacian is 170,472 x 170,472. The standard approach: buy more GPUs.

SIGMA's approach: decompose the graph so every eigensolve operates on at most 500 vertices. The O(n3) cost is factored into n/500 independent bounded subproblems. Then eliminate per-edit overhead so streaming updates cost microseconds, not milliseconds. The total cost is effectively constant per vertex.

Why It Matters

A post-generation verification layer.

When an AI agent ingests 50,000 contracts, there is no mechanism to detect that Clause A in Contract 1 contradicts Clause B in Contract 2. The AI doesn't check. It can't. It just absorbs everything and moves on.

That contradiction doesn't surface as an alert. It surfaces as a lawsuit. Or a compliance failure. Or a decision made on information that was never internally consistent.

SIGMA is the verification layer that catches it. Not with AI guessing. With mathematical proof. It sits downstream of any LLM, RAG pipeline, or knowledge system and checks whether the structure holds.

The Proof

Enron. Same dataset. Different century.

The FBI had to read the emails. That's why it took 5 years and 45 agents. SIGMA never read a single one. It mapped every relationship and checked whether the patterns were structurally consistent. Power-law topology. Hub vertices at degree 1,141. The exact graph structure that breaks naive spectral methods.

FBI Enron Task Force

Duration5 years
Personnel45+ agents
Data reviewed4+ TB
Interviews1,800

SIGMA (Run #23)

Duration30.3 sec
Hardware1 laptop
Cells639
Per entity0.61 ms
4x
seeds, identical results. Zero correctness drift across seeds 42, 137, 2718. The partition structure depends only on graph topology. Deterministic. Reproducible. Every time.
April 17, 2026

The streaming breakthrough.

The original system processed the full Enron graph in 30 seconds. That was the build. The question was: what happens when a new edge arrives? What does it cost to update the structural analysis without rebuilding from scratch?

The measured answer: 63 microseconds. At one million vertices. The cost barely grows with graph size. The scaling exponent dropped from 0.55 to 0.19.

Edit Path

63 us
per streaming edit at V = 1,000,000

The old baseline: 966 microseconds. The fix: stop deallocating cached matrices on every edit. Pool the restriction map math. Two changes, 30 lines of code, 15x faster.

Query Path

13 us
per contradiction query at V = 1,000,000

The old baseline: 132 milliseconds. The fix: a hierarchical nerve tree with an O(1) lowest-common-ancestor oracle. 10,504x faster. Sub-10 microseconds at every measured scale.

Scaling exponent: 0.19 (R2 = 0.975, 4-point log-log fit)
Baseline exponent: 0.55
What this means: doubling the graph size increases per-edit cost by 13%.
Under the old baseline, it increased by 46%.
The Architecture

Cellular Sheaf Decomposition

SIGMA doesn't try to analyze the whole network at once. It decomposes the graph into bounded independent cells, checks each one separately, then assembles global structure through the nerve complex. The bigger the network, the more cells. Each cell takes the same amount of time.

01
DecomposeFiedler spectral bisection with shift-invert eigensolver. The Enron giant component (21,309 vertices) becomes 639 bounded cells in seconds. Each cell holds at most 500 vertices.
02
StreamNew edges arrive. The incremental updater routes each edit to its cell in microseconds. Restriction maps from a pre-computed pool. No cache invalidation. No per-edit eigensolve. 63 us at V=1M.
03
QueryContradiction queries via hierarchical nerve tree with O(1) LCA oracle. 13 microseconds at V=1M. No graph traversal. No BFS. The answer is a single array lookup.
Streaming Scale

The cost per edit is flat across two orders of magnitude.

Measured on the Enron email network (power-law, hub degree 1,141) and Barabasi-Albert synthetic graphs at increasing scale. Four seeds per scale. Zero correctness drift at every measurement point.

VerticesEdit MeanQuery p99DriftCells
21,309 (Enron)0.031 ms0.005 ms0639
100,0000.046 ms0.010 ms0421
250,0000.051 ms0.010 ms01,096
1,000,0000.063 ms0.013 ms04,611
V grew 47x. Edit cost grew 2x. Query cost: flat. Zero crashes. All on a single laptop CPU (i9-13900H, 64 GB RAM).
Early Warning

See it coming before it breaks.

Most systems tell you a contradiction exists after the damage is done. SIGMA watches the structural tension building in real time. In testing, it detected stress forming 21 steps before any other diagnostic noticed a problem. That's the difference between a smoke alarm and watching the temperature rise.

Traditional Detection

Contradiction:
Yes / No

Binary. You find out after it happened. Too late to prevent the damage.

SIGMA Spectral Tracking

21-cycle
lead time

Continuous signal. Detected structural stress while every other diagnostic still reported all clear.

In Plain Language

An X-ray for knowledge.

A doctor doesn't read every cell in your body to find the fracture. They take an X-ray, and the break lights up. SIGMA works the same way. Scan the structure. The contradictions light up. Then you know exactly where to look.

Legal Financial Compliance Healthcare Intelligence Defense eDiscovery RegTech Biotech
SIGMA + Proofkit

The truth, and the unforgeable proof of the truth.

SIGMA finds structural contradictions mathematically. Proofkit signs every output cryptographically. Together, every analysis produces a sealed evidence pack: Ed25519 digital signature, SHA256 hash, complete audit trail. Independently verifiable by any third party with the public key.

No trust required. Download the bundle. Verify the signature yourself.

Ed25519
+
SHA256
These documents contain N structural contradictions, located at specific positions, proven mathematically, and here is the cryptographic proof that this analysis is authentic, deterministic, and independently verifiable. The kind of output that could survive cross-examination in court, regulatory audit, or peer review.
01
DetectSIGMA computes H1 over the knowledge graph. When H1 > 0, contradictions exist that no local adjustment can fix. The result is deterministic.
02
SignProofkit signs the analysis with Ed25519 and hashes it with SHA256. The sealed bundle contains the input hashes, the SIGMA configuration, the full output, and the signature.
03
VerifyAnyone with the public key can verify the signature offline. No server. No API. No trust. The math and the crypto are both independently auditable.
Get Started

See it run on your documents.

Upload a document set. Watch SIGMA find the contradictions. Download the signed evidence pack. No commitment. No sales pitch. Just math.

Request a Demo Download Sample Corpus Read the Full Story