CellID
An open-source identity-resolution experiment that uses REIM to study how noisy pair evidence becomes coherent, measurable, and revisable identity over time.
About the project
CellID is an independent, open-source experiment in probabilistic record linkage, also known as entity resolution.
It asks how a system can decide whether emails, phone numbers, devices, cookies, and other identifiers refer to the same underlying entity when every available signal is partial, noisy, or potentially biased.
The project began as a synthetic-data testbed. On real identity data, the correct answer is usually unknown, restricted, or entangled with personally identifiable information. CellID generates populations in which identity is known by construction, while keeping that ground truth structurally separate from the evidence seen by matching systems. This makes false merges, false splits, calibration, graph coherence, and temporal continuity measurable rather than anecdotal.
How CellID uses REIM
CellID employs REIM, the Reticular Epistemic Inference Model, to investigate how evidence from sources with different reliability should affect a candidate match.
In the original bridge, each candidate pair becomes a REIM system, each signal source becomes an observer, and each score becomes an observation. REIM estimates the latent value of the pair together with the reliability of the sources reporting on it.
The experiment later introduced IdentityREIM, an identity-specific extension owned by REIM. It learns class-conditional source behavior, identifier-type domains, and the difference between an applicable source that remained silent and one that was never applicable. Its output is calibrated pair evidence with source-level provenance.
REIM does not assign the final identity. That distinction is central to the experiment. Pair inference estimates the credibility of a relationship. A separate constrained graph resolver must turn many relationships into one complete and coherent partition.
What the experiment contains
CellID keeps the identity pipeline divided into explicit layers:
- a deterministic synthetic generator with structurally isolated ground truth;
- observations and complete source-opportunity states;
- optional IdentityREIM pair calibration;
- constrained and risk-aware graph resolvers;
- complete-partition metrics for over-merging, over-splitting, and transitive error;
- a temporal registry that assigns stable opaque CellIDs across merges and splits;
- an append-only persistence contract with in-memory and LadybugDB implementations;
- reproducible internal, external-transfer, scale, and temporal benchmarks.
No layer silently assumes the responsibility of the next one. IdentityREIM does not create entities. The resolver does not issue persistent CellIDs. The registry does not score evidence. The graph store persists decisions without defining what an identity means.
What the results show
On the frozen synthetic benchmark, IdentityREIM materially improved complete-partition quality over an unchanged mean-score baseline. It reduced large false-merge components while improving both pairwise and entity-level measures.
The same pipeline was also tested on FEBRL3, an external synthetic record-linkage dataset. There, IdentityREIM produced a large improvement in probability calibration but almost no improvement in final partition quality over a simple field mean. The mean retained advantages on some measures and avoided all false merges at its selected operating point.
That result defines the experiment as much as the internal improvement does. CellID is designed to separate better-calibrated uncertainty from universal model superiority, and to expose when a more sophisticated method does not materially improve the identity partition.
Status and boundaries
CellID is currently a 1.0.0rc1 research prototype with a frozen public API, reproducible benchmark reports, deterministic replay, and a tested persistence boundary.
It is not a production identity service, a universal commercial identity graph, or a clean-room security product. It has not been validated on independently collected production PII with a trustworthy gold standard. Tokenization, tenant isolation, query policy, minimum cohorts, deletion workflows, and aggregate release controls remain outside the CellID boundary.
The purpose of the project is narrower and more foundational: to make identity inference inspectable as a sequence of evidence, graph, lifecycle, and persistence decisions, and to test precisely where REIM contributes to that sequence.
The implementation, contracts, benchmark reports, and reproduction scripts are available in the CellID repository. The underlying inference framework is available in the separate REIM repository.