tau
Ethereum-class lossless compression for transaction calldata. Bit-perfect round-trip across 11 EVM chains.
Real-time transaction pipelines. Latency profile under NDA.
Archival and on-disk retention. Latency profile under NDA.
what tau is
Tau is a lossless semantic compression engine for Ethereum-class transaction streams. On a corpus of 320,612 real on-chain transactions across 11 EVM chains, Tau compresses calldata 3.91× in the hot configuration (or 4.01× in archival), with bit-perfect byte round-trip verified across the entire corpus.
A documented set of Ethereum transaction envelope formats and standard ABI call shapes is natively understood. The full list of supported envelopes and per-shape coverage is disclosed under NDA.
A cryptographic inclusion proof is produced for every compressed transaction. An independent verifier is implemented and tested with zero dependency on the engine itself; public release of the verifier is in preparation. Proof construction details are disclosed under NDA.
deterministic boundaries
EVM transaction streams. 11 chains measured: ethereum, arbitrum, optimism, base, polygon, bsc, avalanche, linea, scroll, blast, mantle.
Any decode mismatch invalidates the run. Inputs outside the documented domain are still preserved losslessly with no compression gain.
x86_64 Linux single-thread, in-memory. Per-tx speed is hardware-bound; ratio is hardware-independent.
2026-05-25. Corpus refresh cadence: quarterly. Next refresh due 2026-08-25.
re-execution posture
A sceptical engineer reproduces every headline number from the source in under one hour.
Five reproduction surfaces exist, each documented with its own gold-reference output. Without source access, the eval-kit Docker image is the canonical buyer-side reproduction path — same scripts, same expected output, no source disclosure.
- ·Synthetic engine bench against in-tree reference dataset
- ·Property-test suites with multiple seed populations per release
- ·Real-corpus replay against frozen mainnet + L2 fixtures
- ·Cross-chain benchmark against the documented 11-chain corpus
- ·Independent verifier that consumes the encoded output without depending on the encoder
Exact commands, paths, fixture names, corpus paths and the eval-kit Docker tag are released after NDA. The reproduction itself is open to anyone with source access; the rituals to invoke it are not advertised publicly.
Patent priority filed, PCT process underway. Status: patent pending. Filing numbers and primitive identifiers available under NDA.
Documented corpus with per-block cryptographic hashing. Train/test holdout split documented. No leakage.
Internal property-test suites plus real-fixture replay verified. External cryptographic audit of the accumulator construction: pending.
Memory-safety hardened. An independent verifier is implemented with zero dependency on the engine itself; public release of the verifier is in preparation.
Independent benchmarks against major general-purpose codecs on this same corpus are available under partner access.
Patent filing numbers, primitive codes, per-chain benchmark matrix, audit dates, dataset SHA chains, eval-kit Docker image — disclosed under bilateral NDA.
enter NDA gatewaywhat tau does not claim
Boundaries stated openly — not as fine print, not at the bottom of a card.
·Not a general-purpose compressor. Adversarial random bytes fall through to opaque encoding.
·Not externally cryptographically audited. The accumulator construction is novel.
·Not "beats every codec on every workload". Per-axis comparison versus general-purpose codecs is documented; on selected axes / chains other codecs win. Full matrix under NDA.
·Not free / open-source. Engine is proprietary. Verifier publicly hosted, not yet open-licensed.
·Not "production-ready". Called "pilot / production-candidate" — no external cryptographic audit yet, no SOC 2.
·Does not compress fully encrypted or already-compressed streams. Lossless is still preserved; no compression gain is possible.
measurement log
Engine source, full benchmark matrices, IP detail, eval-kit Docker image — gated.