The most valuable result in a technical audit is not always the most flattering number. Sometimes it is the boundary that stops a promising result from becoming an unsupported general claim.
Haivian Technology's validation of Exact Rational Computation Exchange confirmed the recorded arithmetic of a fixed deterministic study, independently matched all 348 archived result digests and preserved a meaningful aggregate byte reduction. It also preserved the case where that advantage reversed.
Reporting both is not a weakness. It is the reason the finding is useful.
What the experiment measured
The study compared two concrete representations for exchanging exact affine-rational computations:
- a compact candidate representation wrapped in identity, revision, integrity and response-binding fields; and
- a strong canonical typed exact-rational baseline operating under the same declared semantic workload.
The frozen suite contained 24 scenarios across eight workload families, two protocol arms and three deterministic repetitions. After multistage exchanges, contextual exchanges and declared recoveries, the record contained 378 raw rows.
The fresh audit reconstructed the study's accounting and separately checked the archived mathematical results. It identified:
- 348 recorded valid or recovery rows whose result digests matched;
- 30 recorded typed detections;
- 18 permitted clean recoveries already included within the 348 successful rows; and
- 126 deterministic keys with matching recorded identities across three repetitions.
The repetitions and subexchanges are dependent observations, not independent samples from a general population. That denominator belongs beside the headline result.
The aggregate result
Across the complete fixed suite, the candidate's recorded simulated frames totalled 86,490 bytes. The baseline totalled 100,800 bytes.
That is a difference of 14,310 bytes, or 14.1964% fewer complete wire bytes relative to the baseline in this fixed suite.
The wording matters. The audit did not convert that result into a universal claim that the candidate was “14.2% more efficient.” Efficiency can include time, compute, memory, network conditions, implementation quality and workload mix. The historical frames measured local simulated byte strings, not real network traffic, and fault timing was incomplete.
The repeated-context reversal
The aggregate advantage did not hold everywhere.
In the repeated-context family, the candidate used 37,578 bytes against the baseline's 34,905 bytes: 7.6579% more. In the largest repeated-context case, the candidate cost 30% more.
This reversal reveals where compact inner payloads can stop dominating the identity, state and integrity envelope around them. It also prevents a fixed-suite total from being marketed as a workload-independent law.
A strong independent technical audit should search for precisely this kind of adverse result. If a review publishes only the aggregate win and hides the family-level reversal, it is not giving decision-makers the evidence they need.
What required correction
The numerical agreement did not justify blanket validation of the original version 1 paper. The audit required material corrections to the research account, including:
- restoring the hierarchy between primary and secondary endpoints;
- separating historical observations from fresh verification;
- explaining oracle and specification dependencies;
- stating admitted mathematical and framing bounds;
- preserving adverse scenario and family results;
- reporting missing fault timing and telemetry limitations;
- narrowing claims about authentication, reproducibility and prior work; and
- correcting factual references.
The revised version 2 manuscript incorporates those changes. A separately identified source-level closure check confirmed the listed reporting corrections in the exact revised version.
That closure does not manufacture new experimental evidence. It confirms that the revised paper now describes the supported observations, adverse results and unresolved limitations accurately.
What remains unproven
The validation does not establish a general correctness proof, optimal comparator, cryptographic soundness, sender authentication, AI-communication advantage, token or cost savings, production reliability or performance across unseen workloads.
Historical wire payloads were not available for independent re-decoding, and the proposed public reproduction bundle has not been independently exercised. The audit independently recomputed archived results; it did not rerun the original experiment from released source and fixtures.
Those limits do not erase the fixed-suite result. They define it.
Why this matters for AI-assisted and computational research
AI can generate protocols, analyses and manuscripts faster than conventional research workflows. That makes evidence architecture more important, not less. A defensible third-party research audit should separate:
- the experiment that originally produced the observations;
- later checking of stored records;
- independent mathematical or scientific verification;
- correction of the written account; and
- any subsequent attempt at public reproduction.
When those layers are collapsed, a matching hash can be mistaken for computational truth, a corrected paper can be mistaken for a new experiment, and a bounded benchmark can be marketed as general performance.
Is this an arm's-length third-party audit?
No. Haivian Technology is the corporate validation author, and the related researcher, audit service and journals are under common ownership or control. The publication discloses that relationship and does not present this case as external institutional endorsement.
It does, however, demonstrate the evidence discipline Haivian now offers to independent clients: exact-version identity, claim-by-claim findings, adverse-result preservation, explicit limitations and accountable human sign-off.
Read the complete record
- Read the complete Exact Rational Computation Exchange validation.
- Explore Haivian's independent mathematics and science research audits.
A credible audit does not merely confirm the exciting number. It tells you where that number stops.