← Back to home

Technical detail

VCEP stands for Verifiable Computation over Independently Evidenced Populations. This page shows what the EVP-1 (Population coverage) proof of concept does today, what the verifier checks, and what we don’t claim.

01 EVP-1 today

Four steps, one file.

This is EVP-1 as designed. The packet below and the Status section show what the proof of concept verifies today.

  1. 01

    The control is written down.

    A readable plan, signed by a governance authority.

  2. 02

    The population is committed.

    A signed extraction states the rule, the time window and the totals. Every record left out is counted and tied to a stated reason.

  3. 03

    One fixed runtime executes it.

    A published program. The plan is data, not new code for each control.

  4. 04

    The Evidence Packet is handed over.

    The recipient runs a standalone verifier: offline, with no account and no call to Ellipe 3.

From our proof-of-concept runs on AMLworld, IBM Research’s public synthetic AML dataset.
Evidence PacketAMLworld daily runs
source rows read
Oct 2 run: 5,078,345
Oct 16 run: 5,078,345
anchored (in population)
Oct 2 run: 4,932,279
Oct 16 run: 5,078,345
of which marked malformed
Oct 2 run: 78
Oct 16 run: 146,144
excluded
Oct 2 run: 0
Oct 16 run: 0
rejected at ingest
Oct 2 run: 146,066
Oct 16 run: 0
population vs signed commitment
Oct 2 run: ✓
Oct 16 run: ✓
epochs
Oct 2 run: 18 / 18
Oct 16 run: 18 / 18
VERDICT, BOTH RUNSACCEPT

In the Oct 2 run, 146,066 rows with an unmapped currency were rejected at ingest and reported in an unsigned report. Since Oct 16 those rows are anchored and marked malformed, so every row read is inside the signed population.

What the verifier checks today

  1. Scope and evidence signatures are valid, and the scope covers the period.
  2. The population matches its signed commitment, in both root and count.
  3. Classification counts reconcile within each daily epoch.

Planned

  1. The rule behind each exclusion, and each remaining ingest rejection, is signed, not only stated.
  2. The source total is reconciled inside signed evidence.
  3. Signatures come from a governance authority with validity periods.
  4. The runtime matches a published, reproducible build.

02 Status

Where we are.

Built

  • EVP-1 proof of concept in Rust, with a standalone verifier and a file-based trust store.
  • Testing on AMLworld, IBM Research’s public synthetic AML transaction dataset: 18 of 18 daily epochs ACCEPT in both runs. On Oct 2, 4,932,279 records were anchored and 146,066 rows rejected at ingest were reported in an unsigned report. Since Oct 16, all 5,078,345 rows are inside the signed population, 146,144 of them marked malformed.
  • A tamper demonstration of 10 cases (dropped rows, altered totals, broken bindings). Each is rejected at its expected check.

Next

  • Sign the rule behind each exclusion, not only the count.
  • Testing on real-format institutional data.
  • An independent review of the verifier.
  • A reproducible build with a published digest.
  • Continue EVP-2 (Routing and attribution) from prototype, and EVP-3 (Execution) research.

03 Limits

What we don’t claim

  • No regulator has endorsed this, and we don’t represent otherwise.
  • ACCEPT confirms each record has exactly one outcome, not that it is the right one.
  • “Built” means code exists and has run on public test data. Every result so far is on public synthetic data, not on any institution’s records. In our tests, one machine played every role.
  • Layer status: EVP-1 is working as a proof of concept. EVP-2 is a prototype. EVP-3 is active research.
  • How EVP-2 and EVP-3 work internally is not described here while patent filings are in progress.
  • The claim depends on the source population being evidenced by a party independent of the pipeline being checked. Some banks hold such a record today and some don’t. Where none exists, the evidence is only as strong as the pipeline’s own account of itself.
  • The rule behind an exclusion is stated but not yet signed.
  • Rows whose amount cannot be read are still rejected at ingest and recorded only in an unsigned report. The AMLworld export had none.
  • We don’t prove your extraction query was the right one.
  • A snapshot is not a full period. Late postings and restatements are recorded; the general case is still open.
  • We don’t judge whether your policy is wise, only that the approved control covered the full population.

← Back to home

See it on your own controls.

For sponsor banks and fintechs that want completeness evidence on their BSA/AML data. 20-minute walkthrough, no prep needed.