tau · sealed core · public evidence

agτ

tau

The Ethereum archive that answers while still compressed — built without a single third-party archiver inside.

Brotli-class density — 2.93× on a mixed corpus, up to 5.37× on structured ERC-1155 batches (2026-06-09 · Delta) — reached by Tau’s own core, no third-party archiver in the path, at nearly 1 GB/s where Brotli and Zstd slow to a crawl. It exploits structure both within a transaction and across history; on unstructured noise it falls back to ≈1× and never inflates output. Byte-exact recoverable, independently verifiable, queryable while still compressed — without exposing the engine. The core is sealed. The evidence is not.

Request BYOC evaluation Read evidence brief

Blobs changed the tariff layer. Tau attacks the payload layer.

inclusion is not representation · availability is not compression · hashing is not payload economics

speed at equal density · no third-party archiver

same compression — built without a third-party archiver.

Re-run on 2026-06-09 (Delta · Ryzen 9 7900X), 20,000 transactions, 6.13 MB. Tau reaches Brotli-class density using only its own semantic core — no Zstd, no Brotli, no LZ4 in the compression path. The difference is speed: to reach the same ratio, general-purpose compressors must drop to their slow high-quality levels. Tau holds it at nearly 1 GB/s.

algorithm	ratio	compress MB/s	tau faster by
★ Tau · own core	2.93×	961	—
Brotli q5 (default)	2.91×	66	14.6×
Zstd L11	2.90×	84	11.4×
Zstd L22 (max)	2.96×	3.9	247×
Brotli q11 (max)	2.97×	1.0	952×

The faster general-purpose codecs (LZ4, Snappy) run at 1,400–1,700 MB/s but reach only 2.27–2.28× — roughly half the density. Tau is the single point that holds both high density and high speed at once. On structured batches it pulls ahead on ratio too: 5.37× on ERC-1155 vs Zstd 5.16× / Brotli 5.18× on the same input.

Thirty years of codec research to reach this density. One semantic core — built in months — to match it, and to read the data while it stays compressed.

for the skeptic · verify it yourself

“just delta-compression,” says the skeptic. the code says otherwise.

Tau ships its own honest telemetry. In the same binary, on the same input, it runs Zstd (L3/L11/L22), Brotli (q1/q5/q11), LZ4 and Snappy head-to-head against its own engine — and prints the table. Those codecs are benchmarked as competitors, never called as components: Tau reaches its density with none of them required in the compression path. General-purpose compressors win by hunting repeated bytes. Tau works a layer up — on the structure of the transactions themselves. Same density, reached a different way. Don’t take the claim on trust — run the benchmark on your own corpus and read the table it prints.

built-in benchmarkThe competitor table is part of the engine, not a slide — Zstd / Brotli / LZ4 / Snappy run in-process on every evaluation.independenceTarget density is met without those codecs in the compression path; they exist only to be measured against.different mechanismNot byte-repetition search. Tau models transaction structure — which is why it can answer queries while the data stays compressed.buyer-side proofBring your own corpus. The table regenerates on your data, on your hardware.

evidence brief

a verification challenge, not a compression claim.

Every figure below was produced by running the engine on 2026-06-01 (Mac arm64), not quoted from a doc. Run logs are retained. Tau does not ask the market to believe — it asks the market to reproduce.

real corpus · lossless

100 tx · 6 chains · bit-perfect

head-to-head

2.93×@961 MB/s vs Brotli q11 2.97×@1.0

format coverage

legacy / 2930 / 1559 / 4844-blob / 7702

inclusion witness

76 B constant · O(1) verify

query-in-place

0.01 ms · no decompression

conformance

multichain replay · 100/100 round-trip

repeat-run soak

1000/1000 lossless round-trip · 0 errors

what tau is

A sealed transaction-representation engine.
A bit-perfect recovery path.
A constant-size witness in the tested proof path.
A buyer-side verification model.
Infrastructure-oriented throughput.

what tau is not

Not a blob replacement.
Not a Merkle replacement.
Not a generic compressor.
Not a public-source release.
Not a universal cost-reduction claim.
Not externally crypto-audited yet — the construction is novel; independent audit is scheduled and available to evaluation partners under NDA.

If Tau fails, your corpus proves it. If it holds, your pipeline has a representation problem worth quantifying.

multi-workload matrix · live

easy transfers are not the benchmark. real workloads are.

Ten Ethereum-shaped workload classes, 5,000 transactions each, re-run on 2026-06-09 (Delta · Ryzen 9 7900X). Each row is the same input through Tau, Zstd L11 and Brotli q5 — density side by side, with Tau’s compression throughput. On structured batches Tau pulls ahead (5.37× on ERC-1155 vs Zstd 5.16× / Brotli 5.18×); on unstructured noise it falls back to ≈1× and never inflates output. The gain is a corpus / cross-tx effect — structure found across transaction history, not single-tx magic.

workload class	raw	Tau×	Zstd L11	Brotli q5	Tau speed
ERC-1155 mixed	2.46 MB	5.37×	5.16×	5.18×	2,004 MB/s
NFT transfers	1.64 MB	3.96×	3.96×	3.99×	1,601 MB/s
ERC-20 heavy	1.49 MB	3.69×	3.68×	3.73×	1,680 MB/s
L2 batch	1.44 MB	3.54×	3.53×	3.57×	1,830 MB/s
EIP-1559	1.54 MB	2.89×	2.86×	2.87×	1,985 MB/s
legacy	1.54 MB	2.89×	2.86×	2.87×	1,969 MB/s
NFT mints	1.49 MB	2.56×	2.85×	2.87×	1,428 MB/s
DEX swaps	2.40 MB	1.51×	1.53×	1.54×	4,858 MB/s
adversarial · pure noise	3.60 MB	1.25×	1.25×	1.24×	6,374 MB/s
account abstraction	6.65 MB	1.14×	1.14×	1.13×	10,300 MB/s

Weak classes are shown, not hidden. DEX swaps, account-abstraction and pure-noise payloads carry little exploitable structure — Tau, Zstd and Brotli all land near 1× there, a level playing field. Round-trip stayed bit-perfect across every class.

Best-case is marketing. Worst-case is procurement. Both are on this table.

same input, three engines, side by side — reproduce it on your corpus.

methodology separation

numbers, kept honest.

Figures from different runs and datasets are not interchangeable. They are labelled, not blended.

PUBLIC_FACTSCanonical public figures — full holdout, x86_64 (3.91× / 4.01×, 1.19M proofs/s).2026-06-01 · arm64This breadth / procurement evidence run (2.93× @ 961 MB/s, equal-density vs Brotli/Zstd, bit-perfect round-trip).x86_64 re-runPending — hardware-of-record alignment.BYOC evaluationBuyer corpus, buyer baseline, buyer-side verification. The decisive number.

cross-platform verification

same corpus. four hosts. reproduced.

The same 20,000-transaction / 6.13 MB corpus, replayed on four hardware generations. The figures below are execution-view (read) throughput — the rate at which compressed transactions are served in place. This is a different metric from the 961 MB/s compression throughput above; the two are labelled, not blended.

host	CPU	OS / kernel	exec-view read	P99 latency
Delta · re-verified 2026-06-09	AMD Ryzen 9 7900X	Linux 6.17	12.45 GB/s	380 ns
Mac	Apple M2 (ARM)	macOS	12.27 GB/s	<100 ns
Beta	AMD Ryzen 7 7840HS	Linux 6.17	10.77 GB/s	180 ns
Alpha	Intel i3-9100	Linux	7.34 GB/s	435 ns

Execution-view throughput is an in-place read rate on the tested corpus — not a compression-ratio claim and not a network line-rate claim. Delta was re-verified live on 2026-06-09 (12.45 GB/s); Mac, Beta and Alpha rows are from the prior cross-platform run and were not re-measured in this session.

One corpus, four hosts, one result. Reproduce it on yours.