SPECTR
Structural Primitive Extraction & Compositional Topology for Replay
A pattern-finding engine for any kind of data.
In plain words: whatever goes in comes back out, exactly.
Feed it text, an image, a sensor signal, anything. It learns the building blocks that show up over and over. When the same pattern appears again, it recognises it instead of relearning it. No training, no schema, no guessing.
↓ See the evidence
Live demo
See it in action
Three short, self-contained walkthroughs. Pick a tab, click Run, watch what happens. Nothing leaves your browser.
Read a tiny image, find its building blocks, rebuild it byte for byte.
The Promise
What SPECTR guarantees
Three plain promises. They hold for every input, every time. Not approximate. Not probabilistic.
What goes in comes back out, exactly.
Same input, same answer, every time. On any machine, any run. No randomness, no rounding, no guessing.
Nothing is thrown away or summarised. Every bit that went in comes back out. The original is fully rebuildable from the record alone.
Finds the building blocks that repeat. Records what they are, in what order, and how often. That structural record is what makes everything else possible.
How It Works
Three steps. One promise.
What SPECTR actually does, in plain language.
Read your data
Hand SPECTR any file, stream, or signal. It reads the bits exactly as they are. No conversion, no schema, no assumptions about what the data means.
Spot the building blocks
SPECTR notices the patterns that repeat. The first time it sees one it learns it. The next time, it recognises it. The record of what repeated, in what order, is the structural identity of the data.
Rebuild it perfectly
From that structural record alone, SPECTR can rebuild the original data byte for byte. Nothing is lost. Nothing is approximated. The same record that detects patterns also reconstructs the input.
Common misconceptions
What SPECTR Is Not
There are many benefits to SPECTR, but it's important to understand which are byproducts of the engine, not its primary purpose.
SPECTR naturally deduplicates recurring primitives, which can reduce size, but that is a byproduct of how it records structure, not a design goal. It makes no guarantees about size reduction, and the ledger grows with novel data.
The original stream is fully reconstructible from the ledger. Nothing is one-way. The seed is a structural identifier, not a content fingerprint.
No AI is used at any point. No machine learning, no training data, no weights, no neural networks, no gradient descent. SPECTR is a deterministic algorithm: same input and configuration always produces identical output on every machine.
SPECTR makes no confidentiality guarantees. The ledger encodes structure, not secrets. It is a substrate layer; what you build on top handles access control.
What it is: A deterministic, append-only substrate that decomposes any binary stream into a primitive vocabulary, records the arrangement as a replayable timeline, and reconstructs the original stream exactly. replay(ingest(X)) == X, always.
What we have actually shown
Three things SPECTR found
that other tools missed.
A hidden pattern that gzip cannot see
We hid a repeating pattern inside one file and shuffled the bytes of another. Both files had identical byte statistics, so they looked the same to a normal compressor. Only the engine could tell them apart.
98 times out of 100, SPECTR picked the right file. Gzip got it right half the time, no better than a coin flip.
Same letters, different order. The engine noticed.
Two families of human DNA sequences share 100% of their structural building blocks. Only the order in which those blocks appear is different. Standard tools that look at what appears (not how it is arranged) cannot tell the families apart.
Two DNA types share 100% of their building blocks. Only the order differs. SPECTR told them apart 96 times out of 100.
The standard randomness test says they are identical. They are not.
We made 1,000 pairs of files. Each pair was carefully built so that both files had the exact same randomness score. One had a hidden repeating pattern, the other was shuffled. We then asked two methods to tell them apart.
Identical randomness scores. SPECTR was right 99 times out of 100. Shannon entropy was right 50 times out of 100.
Built for Research
Every result is reproducible.
The engine is Rust. The bindings are Python.
REST API
All engine capabilities available over HTTP. Ingest, replay, structural profiles, batch analysis, and noise-class normalisation through one authenticated endpoint.
curl -X POST /v1/engine \
-H "X-Api-Key: $KEY" \
-d '{"data": "<base64>", "mode": "auto"}'Python Wheel
The Rust engine ships as a native Python wheel. Run experiments locally without any network round-trips. Full API parity with the hosted service.
import ufm
engine = ufm.InvariantIdentityEngine(
bit_depth=21, storage_path="ledger.bin"
)
seed, status = engine.process(data)
original = [bytes(seq) for seq in engine.replay(seed)]Open Claim Registry
Every quantitative result on this site maps to a claim ID in a public test registry. Tests are deterministic, seeded, and re-runnable. Claims that turn out wrong are listed honestly with their root causes, not removed.
Browse the full registry →