Roadmap
Three capabilities are designed but not yet shipped. They are not part of the current product. This page exists so you know what’s coming and can plan around the boundary between what’s measured today and what is on the way.
Run on compressed
Section titled “Run on compressed”Lazy-decompression of a compressed program or model so that CPU and memory track the active query rather than the full artifact. The architecture supports it — the rule table and residual stream are independently addressable — but no CLI surface or runtime exists today.
The natural test domains are model weight tensors (operating on a compressed LLM during inference) and large structured archives (running query workloads against a compressed columnar store).
Cross-domain pattern matching
Section titled “Cross-domain pattern matching”A query in the formula vocabulary of one artifact resolved against another in a different domain — a phrase extracted from medical notes searched against the rule table of a financial corpus, or a motion signature from one telemetry stream matched against another sensor’s deltas.
The shared symbolic substrate makes this tractable: every artifact lives in the same coordinate space regardless of its origin domain. The index and query language to expose this property are not yet built.
Additional domain coverage
Section titled “Additional domain coverage”Domains where Bindu has plausibly applicable architecture but no measured benchmark numbers yet:
- Video frames. Each frame is similar in structure to the imagery cases that already work; the question is how the inter-frame redundancy interacts with our pipelines, and how we compare against
cjxl/AV1on the relevant axes. - RNA / genomic reads. Base-pair packing plus the patterns specific to read alignment. The domain has a strong specialist codec (
genozip); the comparison has to be measured. - Large neural network weight tensors. A natural fit for the dictionary and stride pipelines, but tensor layout reshaping changes the problem; preliminary work, no claims.
- Audio waveforms. Similar story to video; FLAC is the lossless reference and is competitive enough that the comparison needs to be done seriously.
How we’ll ship these
Section titled “How we’ll ship these”When these capabilities ship, they will be released with the same standard as the current ones:
- Published numbers against named baselines.
- A public corpus anyone can reproduce on.
- Bit-exact round-trip verification by SHA-256.
No claims until measured. No “should work” benchmarks. The capability ships with reproducible numbers or it doesn’t ship.