1,000,000 real tx, start to finish, real AWS Nitro Enclave
Windows authorized
1000 / 1000 100%
by a real, separate authority-service process on a real AWS CloudHSM cluster
Key rotation
3.12 µs/step (mean)
1000/1000 unique session-signing addresses -- zero repeats
On-chain relay
17 / 17 succeeded
real release() broadcasts, sent from OUTSIDE the enclave (no shell/network inside a real Nitro Enclave) -- 0 failed
Everything real at once, for the first time: real Nitro
Enclave (non-debug mode, so PCR0 is the genuine build-time measurement, not the zeroed
value --debug-mode produces), real Kafka/Redpanda intake, real re-attestation checked by a
real, network-separated HSM authority backed by real AWS CloudHSM -- not the in-process
HsmAuthority stand-in. Per-window throughput averaged 36,588 tx/sec (p10-p90:
20,523-48,609 tx/sec), well above Visa/VisaNet's own published real average (1,700-3,000
tx/sec) and real peak (4,000-7,000 tx/sec) load, though below their published theoretical
maximum (>65,000 tx/sec) -- see the record below for that comparison.
On-chain relay from outside the enclave, verified 2026-09-17:
a real Nitro Enclave can't shell out or reach the network itself (FROM scratch,
vsock only), so the same real signature + real HSM-issued token is sent over vsock to a
dedicated host-side process, which broadcasts it -- opt-in, host-side policy
(TACHPAWL_RELAY_ONCHAIN=1), not baked into the enclave build. A separate run
against a fresh VelocityArbiter deployment (real CloudHSM cluster's own key, see
docs/tachpawl-key-architecture.svg): 1,034,000 real transactions, 100% HSM
authorized, first confirmed release() tx
0x031528903eceb64ccbabfc39fd57d320edea703700703cfd2bf935686cf70a64.
Best Recorded Run — Dev-Laptop Throughput Ceiling
2026-09-16, fixed record
Steady-state
187,989 tx/sec
real Kafka intake → ratchet → Merkle → HSM auth
Averaged
126,296 tx/sec
1,000,000 real tx in 7.92s wall clock, cold start included
Key rotation
0.4 µs/step
the ratchet step alone, isolated from network/queue time
Not a production or HSM-authorized number: measured on
one developer laptop (WSL2) with the in-process HsmAuthority stand-in, not a real,
separate HSM. This is the pipeline's own ceiling with everything off the hot path removed
-- it's what tells you the crypto and Kafka intake are never the bottleneck, not a
deployment-shaped claim. For that, see the real Nitro + real CloudHSM run above.
vs. Visa/VisaNet's own published figures: real average 1,700-3,000 tx/sec, real peak
4,000-7,000 tx/sec, theoretical maximum >65,000 tx/sec -- this dev-laptop ceiling
exceeds even that theoretical maximum. Reproduce it yourself:
scripts/wsl-benchmark-kafka-intake-throughput.sh.
Real Threshold-ECDSA — Cross-Continent Signing
2026-09-26, real hardware, 3 AWS regions
Keygen (2-of-3 DKG)
93-105 sec/party
indistinguishable from same-region/in-process -- CPU-bound safe-prime search dominates, not network RTT
Signing, cross-continent
3.99-5.85 sec
Virginia <-> Singapore, relayed through Stockholm -- byte-identical signature both sides
Signing, in-process baseline
~1.2-1.3 sec
same protocol, one process, for comparison -- real WAN adds roughly 3-4.5x, still single-digit seconds
Public key agreement
3 / 3 matched
every party independently converged on the same shared public key
What this proves: a real 2-of-3 threshold-ECDSA (cggmp21)
DKG + signing ceremony run across three genuinely separate AWS EC2 instances in three
different regions on three different continents (eu-north-1/Stockholm hosting the relays,
us-east-1/Virginia and ap-southeast-1/Singapore as the two signing parties for this run) --
not simulated, not same-machine. Both signing parties independently arrived at the
byte-identical signature and independently verified it before reporting it.
What this does NOT prove, stated plainly: one operator
(one AWS account, one SSH key) controlled every party's process in this run. The protocol
itself never assembles the complete private key at any point -- each party only ever
computes its own contribution -- but an operator who can read every party's own key-share
file could, outside the protocol, combine any t of them via Lagrange interpolation and
reconstruct it. Genuine trust separation needs each party under a genuinely different
controlling organization, which this run does not demonstrate and no infrastructure test
can -- see TODO.md item #11.
Reproduce it yourself: scripts/wsl-deploy-and-build-party.sh,
scripts/wsl-run-keygen-on-party.sh, scripts/wsl-run-sign-on-party.sh.
Real TDX Attestation — Second TEE Backend (Phala Cloud / dstack)
2026-09-27, real hardware, Intel TDX
Real DCAP quote
5010 bytes
header 04 00 02 00 81 00 00 00 -- version 4, ECDSA-P256, tee_type=TDX, matching Intel's real layout
Cert-chain verification
VERIFIED tcb_status: UpToDate
dcap-qvl against real Intel PCCS collateral, not a stubbed check
mr_config_id across Restart/Stop-Start
unchanged same deployment
byte-identical across a real Stop->Start and a real plain Restart of the same CVM
mr_config_id after delete + redeploy
changed new deployment
real, deployment-scoped identifier -- not always-zero, not one fixed value regardless of deployment
What this proves: a second, genuinely different TEE
backend works end-to-end, not just boots -- a real TDX DCAP quote fetched from a real
Phala Cloud confidential VM over its guest socket, independently verified against Intel's
real certificate chain (not the platform's own say-so). The restart-uniqueness test
settles a question raised by TDX's attestation shape: mr_config_id/
mr_owner/mr_owner_config are real, per-deployment identifiers set
at CVM/App creation time -- confirmed by comparing three real events (Stop->Start, plain
Restart, full delete-and-redeploy-from-scratch as a brand-new App/CVM). The verification
CVM itself was torn down once this test completed (the delete-and-redeploy step, above,
required that anyway) and is not kept running continuously -- redeployed on request for a
live demonstration.
What this does NOT prove, stated plainly: these fields
are not a per-boot-random, Nitro-module_id-equivalent restart-detection
signal -- they cannot catch a stealthy relaunch of the *same* deployment, only distinguish
one deployment from another. The genuine per-launch-unique field TDX's own module spec
defines (TDID256) is not reachable today: an Intel engineer confirmed on a
public forum that it "has not been upstreamed" through QEMU, and independently, the
real DCAP report structs this project parses (TDReport10/TDReport15
via dcap-qvl) have no such field at all. See TODO.md item #11 for the full
account, including the real ca-certificates bug found and fixed along the way.
Opened directly as a file, or via a local server next to tachpawl-tee-status.json — when reachable, the page polls it every 5s.