Clamp first, then floor to $5,000 units.
One report, two envelopes
Choose a measurement, split it, then let both aggregators contribute a verifier share. Only their combined verifier can accept.
24-bit range: 0 to 16,777,215. Client-side limits are convenience, not the defense.
- 1Enter
- 2Shard
- 3Prepare
- 4Collect
Collector tally
The protocol result is computed from output shares. The plain sum is computed separately and compared.
One noisy release, contributed before collection
Exact secure aggregation hides each input from one server but can still disclose it through related totals. This bounded lesson lets honest clients add independently sampled discrete-Laplace noise inside their Prio3 reports.
One slot may change from 0 through 50 units, so sensitivity is 50 units.
Each honest client activates a full scale-50 draw with probability 9/10.
The no-honest-noise event is at most 1 / 1,000,000, plus the disclosed output-magnitude-cap coupling term.
INSECURE EXACT-RELEASE WORLD
Subtract two exact 12-slot totals
Both queries use real Prio3Sum reports. The second keeps every slot but replaces Lane's bounded value with zero. Their difference reveals Lane's clipped, quantized contribution.
BOUNDED TEACHING ADAPTATION
Commit one cohort, release once
The collector receives one noisy aggregate. Honest-count labels and the per-client decomposition below are an omniscient teaching trace, not information the collector is allowed to receive.
No batch has been released.
SEPARATE EXACT MODEL · EPRINT 2026/2059
Who sees an abort changes the mechanism
This bounded-retry, exact-rational lesson is separate from the Prio3 route above. It compares one release, an outcome-dependent abort, publication conditioned on success, a complete capped retry transcript, and an independent-abort control.
No abort-model ledger has been computed.
Meisingseth, “On Aborts in Differential Privacy,” ePrint 2026/2059, received 16 September and approved 18 September 2026. The paper's outcome-leakage model motivates the selective case; this lab's finite retry transcript is an explicit bounded adaptation.
Same goal, different trust boundary
These models are comparisons, not interchangeable labels.
| Model | Who sees inputs? | Where noise is trusted | Failure boundary |
|---|---|---|---|
| Central DP | Trusted curator | Curator adds calibrated noise | Curator compromise exposes raw records |
| Local DP | No curator sees raw input | Every client randomizes its own record | Usually pays substantially more utility cost |
| This distributed lesson | Two non-colluding Prio3 aggregators each see only a share | Privacy comes from an honest activated draw; the no-activation event is charged to δ | Aggregator collusion invalidates the aggregate-view claim; fewer than six honest contributors invalidate the advertised activation-failure term δ ≤ 10−6 (a weaker h-specific term may remain), while the total also adds the disclosed cap-coupling term |
| Shuffle DP | Analyzer receives unlinkably shuffled reports | Clients randomize; shuffler hides linkage | Comparison only: this lab implements no shuffler or amplification theorem |
Make the verifier say no
These controls bypass the friendly input path and attack the real proof preparation code.
Oversized salary
Submit 4,000,000,000 into a 24-bit Sum circuit. The malicious client constructs a malformed encoding directly.
Damaged proof share
Flip one Field64 element after sharding. The salary shares are untouched; proof consistency fails.
Replayed report
Send the identical report twice. Preparation verifies it again and accepts it again; only this lab's nonce registry, outside the VDAF, refuses the duplicate.
TRUST ASSUMPTION
Put both envelopes on one desk
One aggregator sees only a random-looking share. If both aggregators collude, their shares add back to the encoded input.
What a valid proof cannot prove
A validity proof checks the circuit statement, not the world outside it.
NEGATIVE CLAIM 01
A well-formed lie still passes
Prio3 checks that a report is in range and correctly shared. It cannot check that the report is true.
NEGATIVE CLAIM 02
A private share can reveal a public total
Prio3 does not add differential privacy. With one accepted report, the aggregate is the report.
Pinned evidence
Runtime checks use the same deterministic inputs as the committed unit tests.
Protocol scope and extension seams
Real here
- Field64 arithmetic and serialization
- TurboSHAKE128 with domain byte 1
- BBCGGI19 FLP prove, query, and decide
- Prio3Count and bounded Prio3Sum
Not this lab
- DAP transport or HTTP
- Poplar1 or IDPF
- SumVec or Histogram
- A production DP deployment
- More than two aggregators
Extension seams
- Vector circuits in
src/flp/circuit.ts - Other audited distributed-noise constructions
- Additional VDAF instances beside Prio3