The answers are already in your systems — they just can't agree on who's who. We resolve the same person, facility, and contract across dozens of disconnected repositories, and surface the links between them — without moving or replacing anything. Then we validated it across a terabyte-scale campaign — 3.95 billion records, 5,329 independent runs.
Federal agencies have accumulated the data and the systems — but decisions wait, because the same person, facility, and contract look different in every repository, and nothing connects across them.
Every vendor shows a beautiful front end. The hard part — the part nobody demos — is making the backend actually work the data: resolving identity across systems that share no common key, and proving each connection is right.
Platform-agnostic. It sits on top of whatever you run — Oracle, SharePoint, cloud, mainframe — and deploys inside your security boundary. We don't migrate your data; we resolve across it where it lives.
Connect the same person / facility / provider / contract across systems that share no common ID — and prove every connection against evidence.
Which contracts touch a facility, its operational status, whether that tracks with care delivery — the enterprise question you can't answer today.
Only the resolved links persist. No central copy, no rip-and-replace, no proprietary lock-in — open, standards-based, in-boundary.
Every resolution is inspectable and human-reviewable — the foundation a Chief AI Officer and data-governance role can build on.
The question any data person asks is "does it hold at scale?" So we didn't promise — we ran it: 3.95 billion records across 5,329 independently generated estates, each scored against its own answer key. Across the entire campaign, the score never drifted:
Honest scope: synthetic, PHI-free estate. On the co-designed mess model these figures validate mechanics and per-layer lift; on a hold-out corruption vocabulary the matcher was never built for, F1 is 93% with precision held at 99.4% — unknown mess costs missed matches, never wrong ones. Live-data accuracy is what a CRADA validates. The campaign is 5,329 independent fixed-size runs totalling 1.10 TB of generated JSON (≈795 GB compact); single-population sharded resolution is built and validated — blocking-key partitioning reproduces the unsharded result exactly. One laptop, 19.5 hours, measured.
What the 1.10 TB is, precisely. It is the cumulative volume the pipeline generated and resolved — not a single corpus that sat on a disk. Each 50k-veteran population (~206 MB) was generated, resolved, scored against its own answer key, and discarded before the next one began; peak disk never exceeded ~1 GB and peak RAM was bounded the same way. A 1.10 TB estate never existed at any one moment, and we are not going to imply otherwise. That is a property of the design, not an apology for it: the pipeline is streaming and bounded, which is exactly why it runs on a laptop and why it scales linearly on a cluster. If what you want is one big population resolved as one problem, that is the 500,000-veteran flagship below — 1.6 M identity records in a single ~5.8 GB estate, every record eligible to match every other.
Fair hit. Partitioning a terabyte into independent estates proves throughput, not that identity resolves across an enterprise. So we ran the hard version at the size of the actual problem: one 9,000,000-veteran population, resolved as a single problem — roughly VA's enrolled population, every record eligible to match every other, no partition walls to hide behind.
The degradation is not diffuse. It is one layer, and we can name it.
On the flagship: blocking-key partitioning is provably lossless here — on an identical estate, the partitioned and single-machine resolvers return the same F1, precision, and edges. The five-size curve is measured on true single populations (not independent shards), which is why it is the honest answer to the scaling question and the terabyte campaign is not. All five points measured on one engine and one generator. A caveat that cuts against us: this 9M estate carries 45,000 facilities. Real VA has on the order of 1,300 for a comparable population — roughly 35× fewer — and fewer facilities means more L3 signature collisions, so a realistically-proportioned estate would likely score worse than 92.7% / 2.5%, not better. We are treating this as an optimistic bound and saying so. The 1.10 TB campaign figures above predate a correction to the graph layer and now understate the engine; we publish the older, worse numbers rather than restate a campaign we have not re-run.
Resolving identity is the means. The deliverable is what a program office actually needs on a Monday: named discrepancies, with dollar figures, each with a stated basis and the evidence one click away. On the same synthetic estate (10,000 veterans, 12 systems, keys drifting the way real systems drift), the fabric volunteers:
The same physical device, in three systems that never agreed on a tag. Nobody re-keys a machine on migration day:
Honest scope: every dollar figure in this section is measured on the synthetic, PHI-free testbed estate against its answer key — they demonstrate the method and its honesty machinery, not VA's actual leakage. The attribution-recovery figure is flattered by synthetic GL lines posting claim amounts exactly; live ledgers split and adjust. What survives contact with real data is the shape: exact joins reach a fraction of your money, resolution recovers a large share of the rest, and the residual is queued for people rather than guessed. Sizing YOUR numbers is the first deliverable of a CRADA.
Everything above is browsable: the incident chains, the evidence behind every merge, the findings ledger, and the review queue where the engine hands a person what it will not guess. Access is by invitation while the product is in testbed.
Wrong-merge is the error that matters: it fuses two veterans into one record. So we instrumented the engine against the answer key and asked which layer produces it. The result was specific, and uncomfortable, and fixable.
Measured on the 10,000-veteran messy estate, where the engine's partitioned and single-process paths produce identical scorecards — the invariant that lets a small run predict a large one. The 9M confirmation run has not completed yet, and until it does this fix is measured, not proven at scale. When it finishes, that number goes on this page whatever it says. That is the same promise we kept when the last run made us look worse.
Every clue is weighted by how telling it is — a matching SSN is proof, a matching city is coincidence. The production engine (Splink) learns those weights from your data; our published figures use expert-set weights — the conservative baseline. Same statistical method national census bureaus use.
Comparing every record to every other is impossible at scale. We only compare records that could plausibly match — so it stays fast from a laptop to a terabyte.
Resolve bounded partitions in parallel — accuracy constant per partition, throughput linear in cores. A laptop measured 19.5 h for the campaign; server-class parallelism should cut that to about an hour — an extrapolation we'll validate.
IBM Data Fabric, Palantir Foundry, Snowflake — platforms you buy, marry, and migrate into. Big, slow, lock-in. Another store to move your data out of its systems and into theirs.
Platform-agnostic, in-boundary, no lock-in — and we prove the hard part (resolution + correlation) on your data, fast. We can even ride on top of the platform you already bought. No rip-and-replace.
We're not asking you to buy anything or migrate anything. Through a Cooperative R&D Agreement (CRADA), we co-develop and validate the fabric on your data, in your environment — no funds exchanged, data never leaves.
The data fabric, the engineering team, and the build — co-invested, at no cost to the agency.
Data access, environment, and domain expertise — inside your boundary; data never leaves.
A short confidentiality agreement to scope feasibility and share safely.
Co-develop & validate — synthetic → de-identified → real, under protections.
Prove on real data; ready the enterprise rollout.
On your own data, inside your boundary, with zero acquisition risk. The next step is a short conversation.
Email Metrics →