A semantic inference workbench for language under transformation.
SIGNAL·GRT runs natural language through reversible transformation, latent encoding and context-aware decoding. Inputs enter as plain text. Outputs read as ordinary technical prose. Reconstruction requires the key.
- latency
- ~480 ms
- manifolds
- 9
- decode k
- 3 ranked
- key
- device-local
$ grt run --pipeline transform+embed
[00.012] tokenizer ok (48 tok)
[00.104] manifold technical/v3
[00.288] embed dict:kdf-sha256
[00.471] emit surface="thermal drift in the
secondary loop stayed nominal"
$ grt decode --k 3 --context thread:7f2a
0.94 reconstructed payload
0.71 alternate reading
0.38 null hypothesis
status: key required for full reconstructionsystem modules
Six subsystems, one deterministic pipeline.
Semantic transformation
Rewrite any input across nine stylistic manifolds — technical, lexical, cipher, metaphoric — while preserving the underlying vector.
Latent encoding
Deterministic key-derived dictionaries and LSB carriers embed payloads inside text and raster data.
Context-aware inference
The decoder reconstructs intent from conversational state, returning ranked interpretations with confidence.
Persistent context graph
Every session accumulates a retrievable memory graph the models reference on subsequent runs.
Multi-channel I/O
Text, speech and image inputs normalise into one pipeline with reversible transforms.
Zero-visibility mode
Workspace renders a neutral technical surface until a local key is supplied. Nothing is legible without it.
method
Plain input. Technical surface. Recoverable intent.
- 01
Ingest
Text, speech or image is normalised into a single token stream.
- 02
Transform
A chosen manifold rewrites surface form; the semantic vector is retained.
- 03
Embed
A key-derived dictionary maps residual meaning into carrier tokens.
- 04
Reconstruct
With the key and thread context, the decoder returns ranked readings.