Public-safe model / Synthetic throughout
One source. Several representations. Independently checked.
Most data platforms end up holding the same information more than once — as records, as vectors, as a graph. Each copy is useful, and each one can drift. This is an illustrative model of how I think about keeping them honest: what gets captured, what gets derived, what happens when the source changes, and what a check actually proves.
01Engineering systems
What gets captured, what gets derived, what a check proves
Every quantity below is illustrative. The interesting failure is not a store going missing; it is two stores agreeing about a total and disagreeing about everything else.
Illustrative system model. Synthetic data throughout; it describes no particular system.
Captured before it is parsed
The bytes are retained first, so everything built afterwards has one fixed thing to be compared against — rather than being compared against another copy of itself.
Capture retained
Structured data
Ordered records with stable identities, and the relationships found in their fields.
- Derived from
- The captured source, parsed once
- Checked by
- Reconstructing each record from the bytes it came from
Vector space
One embedding per record, placed so that distance stands for similarity.
- Derived from
- The structured records
- Checked by
- Comparing every point against the record it represents
Knowledge graph
A node per record and typed edges for the relationships between them.
- Derived from
- The structured records
- Checked by
- Comparing the node and edge sets against the records
Illustrative model, not a measurement. One writer per derived view makes a divergence attributable to a single path; it does not on its own establish that the views agree.
Rows written by the ingestion path, each with its own identity.
One embedding per record, written by the projection path.
Totals agree
Totals still agree
- Totals agree
Both sides report the same number of records, and a dashboard turns green.
- One record is removed, one is added
The source changes. The total does not.
- Totals still agree
Nothing about which records are present, or what they say, has been checked.
Illustrative quantities, spelled as words so they are not mistaken for a measurement. A total survives one removal paired with one addition, so equal totals establish quantity alone.
What each earlier version was built from, kept exactly as captured.
- Record A
- Record B
- Record C
- Record D
- Record E
- Record F
What was being served until the updated source arrived.
- Record A
- Record B
- Record C
- Record D
- Record E
- Record F
What is served now, after converging on the update.
- Record A
- Record B
- Record C
- Record D
- Record E
- Record F
- UnchangedRetained
The record is identical to what is held, so it keeps its identity and everything derived from it.
- NewAdded
A record the previous version did not contain is stored and projected into both derived views.
- ChangedReplaced
The record keeps its place and its content is replaced, along with its embedding and its edges.
- RemovedPruned
The record leaves the current version. What is removed is its place in what is served, not the record that it existed.
Current state converges on the update; what earlier versions were built from is kept apart from it rather than brought forward as though it were current. The model does not claim the update itself was correct.
Quantity
- Rules out
- A view that is obviously incomplete or obviously duplicated.
- Still open
- Which records are present, and whether any of them says the right thing.
Identity
- Rules out
- A substitution that preserves the total — one record removed, another added.
- Still open
- Whether the content behind a correctly named identity matches its source.
Content
- Rules out
- Stored text that does not reproduce the bytes it was captured from.
- Still open
- Whether those bytes were the right thing to capture.
Relationships
- Rules out
- An edge the current extraction no longer produces from the record it came from.
- Still open
- Whether an edge that reproduces cleanly points at the right target.
Provenance
- Rules out
- A version built by an undeclared pipeline, or by more than one of them.
- Still open
- Whether declaring that pipeline was the correct decision.
Each comparison is recorded as its own result rather than folded into one status, because passing one is not evidence for another — and none of them turns into proof.
02How the work connects
From data to a decision someone can act on
Select any node to see what it means here and which public artifact backs it. Dashed nodes are directions of study with no public project yet.
Data
Validated, fingerprinted, versioned inputs
Models
Trained approximations of an expensive process
Retrieval & agents
Grounding a response in retrieved evidence
Production systems
Gates, registries, containers, and served contracts
Reliable decisions
Calibrated confidence and an explicit review path
Evidenced
Reliable AI
Uncertainty that ranks likely error, calibration that survives its own protocol, conformal intervals with a stated coverage, and a review queue for the rest.
See the evidenceBack to the work
The model is the frame. The case studies are the evidence.
Nothing on this page establishes anything on its own — it describes how the problem is shaped and where each project sits in it. What a claim rests on is in the case study that makes it.
Selected work