LEF Ai.E — Public Services — Changelog
What's changed in the engine's structural read — what it reasons about and what it says — in reverse chronological order. (Platform changes — access, infrastructure, page copy — live in the internal platform log, not here.)
LEF Ai.E — Public Services, on federal data, points the LEF Ai Engine at the federal mission: a structural read over multi-source, unclassified records — the landscape, the structural voids (combinations that never occur), each void falsified against reality and refuted before it surfaces, and, across sources, where a structure lived in one corpus but never another. Built on the same engine substrate as DCFN-Patents, DCFN-Research, and DCFN-Bio.
2026-11-16 (later): a change some records made is no longer called a rename
- The site's engine moved from 0.100.0 to 0.101.3. The change is in check C14, which compares a set of records with an earlier copy of the same records. On this site that comparison runs in the forward grader, where a past run's records are read against the same window pulled again.
- A value only some records changed is now reported as changed. Before, when most of the moves out of a value went to one other value, the engine called it a rename, even though most records still held the old value. It now calls something a rename only when nearly every record that held the old value moved to the new one, and it names the old spelling first.
- Records named by words are now read for possible renames. Where a record named by words left and another arrived, the engine says whether the name was possibly respelled, split, merged, narrowed or widened, and never decides which. Awards here are identified by codes, so this does not apply to them.
- A total that matches between copies is checked against its records. The engine now says when a total is the same in both copies but the records making it up have changed.
2026-11-16: the engine refreshed to 0.100.0; eight more checks in the validity report
- The site's engine moved from 0.84.3 to 0.100.0. The validity report beside every run now answers 75 checks; eight are new since 0.84.3 (C15 to C22).
- What would strengthen each convergence (C15). For each pattern the engine holds, it names the contrary records that would end it, a third value the pair rides on, and the years it holds. A third value is named, never called the cause.
- Patterns carried across record sets (C16), and the same things recorded by three sources or more (C17), where a source that differs from the others is named without saying which is right.
- The record followed across its earlier copies (C18, C19), field by field, when two earlier copies or more are given. A run on one copy says so.
- Each gap read along who kept the records (C20) and where each finding stands in its own corpus (C21): how firm each gap is against the rest of this record set, with the weakest held gaps and the firmest leads named. No level changes.
- The engine's own past outcomes as labels (C22): reported and kept, never applied to what a run says.
2026-10-28: no likelihood percent; what each outcome rests on, in words
- Each projected outcome no longer shows a likelihood percent or a band. The number was the engine's own score from its verdicts, never measured against outcomes, so it could be read as accuracy it is not.
- In its place, each outcome says what it rests on: the verdict (confirmed, contested, refuted or not adjudicated), whether the read was powered, and whether the proposed fill was re-tested against the records and what that found.
2026-10-26: Engine 0.62.1, combinations of four, and values a record carries from another
This site now runs the LEF SI Engine at version 0.62.1.
- The engine now also reads combinations of four values that never occur together where every three of them do,
and reports each as a lead, saying how far it read. A large run can take a few minutes longer for it.
- Where records point at other records, such as an order at its contract, the engine names the fields a record
carries from the one it points at, so a reader knows which values were recorded for the record itself.
- What a reader sees is otherwise unchanged in kind: on the engine's known-answer benchmark, nothing moved.
2026-10-25: Engine 0.59.0, every run opens with a short read in plain words
This site now runs the LEF SI Engine at version 0.59.0.
- Each results page now opens with "Read first": what the engine read, how many gaps it found and how they stand,
what to look at first, what it could not check here and why, and where the detail is. Every number on it is a count
the run made. The engine's own technical summary follows it, under its own name.
- The engine now names fields that hold the same value in most records, such as two dates that fall on one day, so a
reader knows how the records were kept before reading their gaps.
- The validity report now answers every check, including the three a visitor's run never takes, each saying why.
2026-10-25: Engine 0.57.1, what the validity report may say about a gap
This site now runs the LEF SI Engine at version 0.57.1, and the footer reads that version from the engine itself.
- The validity report no longer calls a gap a feature of the world where only the record's layout was tested. Where
a record has room for both values, the report now says only that the layout is not what keeps them apart.
- Codes written several to an entry are read one by one, and a field that only ever holds one code per record no
longer yields an absence that is true by construction.
- Dates written out in words or with slashes are read as dates, each field the way it writes them, and a field of
dates is never compared against others as a category.
- A quote is held only when it is whole in its source, and a list pulled through a filter never has its silence
read as meaningful.
- A field naming who kept a record is read as how the records were kept, not as something about the world.
- The report also names the checks it could not run here and why: each needs something a visitor does not supply
(known answers, a plan, a survey design).
- What a reader sees is otherwise unchanged in kind: on the engine's known-answer benchmark, nothing moved.
2026-10-15 — a saved copy is named as one
- When a federal source does not respond, the run still reads the last saved copy of it, and now says so: the page, the PDF and the JSON name the source and the date the copy was taken, and say the records are as they stood then. Until today a run on a saved copy read as a live pull dated the run day.
2026-10-11 — Engine 0.39.0: which witness speaks for each finding
This site now runs the LEF SI Engine at version 0.39.0, and the footer names that version.
- Where a source's own documentation is supplied, each finding now says whether the documentation, the records, or both
explain it, or whether they disagree.
- A rule the records keep that the documentation never states is named, so a dataset's steward can see what the
documentation leaves out.
- What a reader sees is otherwise unchanged: on the engine's known-answer benchmark, nothing moved.
2026-10-11 — Engine 0.38.0: a source's documentation read against its records, rule by rule
This site now runs the LEF SI Engine at version 0.38.0, and the footer names that version.
- Where a source's own documentation is supplied, the read now says, rule by rule, what agrees with the records, what the
records contradict, what is documented but never used, and what is written differently, with counts.
- A documented missing-value code changes nothing until the dataset's steward accepts it; until then the read counts
the cells it would affect.
- Only the source's quoted documentation, another named authority, or the dataset's steward can set a finding aside.
- What a reader sees is otherwise unchanged: on the engine's known-answer benchmark, nothing moved.
2026-10-10 — Engine 0.37.1: three corrections
This site now runs the LEF SI Engine at version 0.37.1, and the footer names that version.
- Rules about a source are now checked against its records as written.
- A few stray dates no longer make a short file look long enough to judge over time.
- The date a record was last edited is never mistaken for the date of what it describes.
- What a reader sees is otherwise unchanged: on the engine's known-answer benchmark, nothing moved.
2026-10-09 — Engine 0.37.0: a source's own definitions of its categories
This site now runs the LEF SI Engine at version 0.37.0, and the footer names that version.
- The engine now takes a source's own definitions of its categories, quoted from the source's documentation, and
checks every record against them.
- Each finding's page shows whose rule set it aside or explains it.
- What a reader sees is otherwise unchanged: on the engine's known-answer benchmark, nothing moved between 0.36.0
and 0.37.0.
2026-10-09 — Engine 0.36.0: a rule about a source says who it rests on
This site now runs the LEF SI Engine at version 0.36.0, and the footer names that version.
- A rule the engine is given about a source now says who it rests on: the source itself, another authority, or a
named person can set a finding aside, and the page says which. A rule only LEF believes can explain a finding but
never remove it.
- Where a source's own documentation and its records disagree, the engine says so.
- What a reader sees is otherwise unchanged: on the engine's known-answer benchmark and its hard cases, nothing
moved between 0.35.0 and 0.36.0.
2026-10-08 — Engine 0.35.0: a stretch of a measurement that holds far fewer records than it should
This site now runs the LEF SI Engine at version 0.35.0, and the footer names that version.
- The engine now finds a stretch of a measurement that holds far fewer records than the records on both sides
predict, not only a stretch that holds none, and a few extreme values no longer hide one.
- What a reader sees is otherwise unchanged: on the engine's known-answer benchmark and its hard cases, nothing
moved between 0.34.1 and 0.35.0.
2026-10-07 — Engine 0.34.1: rules say whether a second source keeps them; pages safe to hand on
This site now runs the LEF SI Engine at version 0.34.1, and the footer names that version.
- A rule the engine confirms now says whether a second, different source keeps it, or only files from the same
source.
- A measurement is no longer read as a date because of its column's name, and a dataset's history is compared only
with its own earlier runs.
- The pages a person reads no longer show anyone's email address or ORCID iD.
- What a reader sees is unchanged on the records tested: on the first 3,000 records of the FDIC directory the
structural read gives the same 24 gaps, pair for pair, each in the same state, as the previous engine.
2026-10-06 — Engine 0.33.1: a record told two ways cannot hide a gap
This site now runs the LEF SI Engine at version 0.33.1, and the footer names that version.
- A record filed twice under one identifier, told differently the second time, can no longer hide a gap the rest of
the records show, while records that only share a name are still read as distinct.
- A correction to the entry below: it said rules the records keep are filed by the source they come from. The engine
did so, but this site's own read did not use it until today; it does now. A codebook supplied for a source is also
read by this site's own read now.
- What a reader sees is unchanged on the records tested: on the first 3,000 records of the FDIC directory the
structural read gives the same 24 gaps, pair for pair, each in the same state, as the previous engine.
2026-10-06 — Engine 0.33.0: the site carries the current engine
This site now runs the LEF SI Engine at version 0.33.0, and the footer names that version.
- A date written as a number is read as a date in every part of the report.
- Rules the records keep are filed by the source they come from, so one source's rules never stand in for another's.
- What a reader sees is unchanged on the records tested: on a 3,000-record FDIC sample the structural read gives
the same 29 gaps, pair for pair, each with the same standing, as the previous engine.
2026-10-05 — Engine 0.31.0: no record is left out because it shares a name or number
This site now runs the LEF SI Engine at version 0.31.0, and the footer names that version.
- No record is left out of the read because it shares a name or a number with another record. When the field that
looks most like an identifier repeats, each record is now read on its own.
- What a reader sees is unchanged on the records tested: on a 3,000-record FDIC sample, whose identifier never
repeats, the structural read gives the same 29 gaps, pair for pair, each with the same standing, as the previous engine.
2026-10-05 — Engine 0.30.5: the site carries the current engine
This site now runs the LEF SI Engine at version 0.30.5, and the footer names that version.
- A build's findings are now checked against its engine's own read, and a gap the engine ruled out cannot be
presented as standing.
- Each run's report names the model behind every piece of prose it checked, and says when a piece names none.
- What a reader sees is unchanged on the records tested: on a 3,000-record FDIC sample the structural read gives
the same 29 gaps, pair for pair, each with the same standing, as the previous engine.
2026-10-05 — A gap the records rule out is never listed as standing
- A gap the records themselves rule out no longer appears on the worklist as confirmed. Before this, the gaps the
read examines most closely could be listed as confirmed even when the records showed that one of the two fields is
set by the other, so the two values could never appear together. They are now listed as ruled out, with the reason.
- Each finding now says plainly what supported it ("the full-records count"), where it used to print a bare
internal label.
- What a reader sees: tested on 97,946 public news-event records in LEF's federal build, which shares this code,
the worklist went from 48 gaps to 40; each of the 8 removed was a field set by another.
2026-10-04 — Engine 0.30.3: codes are kept as written, and a stray value is not a gap
This site now runs the LEF SI Engine at version 0.30.3, and the footer names that version.
- A field of codes written as numbers is kept as written. A ZIP code, a FIPS code, a telephone number or a record
identifier is no longer read as a measurement, so a region missing from your records can no longer appear in the
validity report as an empty range in one, and a leading zero is no longer lost.
- A single stray value far from the rest no longer makes the report say a measurement has an empty range.
- When records cover too short a time to be read by period, the report now says so, and names the field it read.
- What a reader sees is unchanged on the records tested: on a 3,000-record FDIC sample the structural read gives
the same 29 gaps, pair for pair, each with the same standing, as the previous engine.
2026-10-03 — Engine 0.30.0: the site carries the current engine
This site now runs the LEF SI Engine at version 0.30.0, and the footer names that version. Its connector was rebuilt from the engine, keeping the parts that are this site's own (its harness, with the living profile library on, and its federal source readers).
- What a reader sees is unchanged on the records tested. On a 3,000-record FDIC sample and the three example files, the structural read gives the same gaps, pair for pair, as the previous engine. The engine's later checks run inside its attached read, which this site's pipeline uses only in part, so the update mostly changes what the site carries, not what it shows.
- Carried now: a strong association between two fields is no longer read as one field written twice, and the engine's record checks are current to 0.30.0.
2026-09-28 — Engine 0.5.2: spellings read as one value, traced gaps, and a clean image
This site now runs the LEF Ai Engine at version 0.5.2, and the footer names that version. What changed in the structural read:
- One value written several ways is read as one value. "TX", " TX " and "tx" were three values; when one spelling clearly dominates a field, the read now uses it and says so. A field whose case carries meaning can be named to be read exactly as written.
- "Not collected" is recognised in any spelling, so a marker such as NOT_AVAILABLE is never read as a real value.
- A field that restates another in different codes is recognised even with a few stray records, and how surprising a gap is now depends only on the records that fill in both fields.
- Each gap is traced. Backward, to the value that holds it apart; forward, to what a record filling it would have to carry, and whether any record consistent with the file could.
- Wide fields are searched over their common values, and a gap that restates a stronger regularity in the same records says so, so several gaps on one regularity read as that one regularity.
- The validity report says a lead's empty cell could be a coincidence, where it had said the absence was expected.
On the recovered input of the 2026-08-05 banking run, the engine still surfaces 54 candidates and refutes 22 against the records themselves.
2026-06-24 — Confidence label matches the verdict for non-gap voids
A void with no proposed outcome — one field restated in another's codes, or an exclusion rule — still printed its probability under “Outcome confidence,” the same label a fillable gap uses for confidence in its proposed fill. For a non-gap there is no outcome; the number is the engine’s confidence in its structural verdict. The line now reads “Verdict confidence” for those voids, so a high number no longer reads as confidence in an outcome the engine just said doesn’t exist.
2026-06-24 — A void that is one field restated no longer prescribes a collection in the next-move
A void that is one field restated (the two columns are the same field — Crm Cd / Crm Cd 1, or a · populated / · multiplicity facet pair) reported “do not task a collection,” yet its next move still said to collect the record carrying both. It now omits the collection prescription entirely — the structural read and the next move agree.
2026-06-24 — Civil-liberties gate extended to the next-move and the precondition
A protected-attribute void already withheld its authored outcome and suppressed its collection task; the gate now also neutralizes the two places that still prescribed collecting the protected join.
- The next-move no longer prescribes collecting a protected-class join. Where a void touches a protected attribute (race/ethnicity, sex/gender, religion, national origin, age, disability, sexual orientation), the next-move reads “no collection tasked — reportable only with an independent criminal predicate and civil-liberties review (28 CFR Part 23); not a collection target,” instead of “collect the record carrying both …”.
- The precondition reads as a reportable potential, not a collection mandate. The named condition and survivability for a protected-class void no longer frame closing it as a collection target.
The structural fact (the two values never co-occur) still surfaces as a reportable potential — reported, never prescribed against a protected class. Surfaced by running real protected-attribute data (LA crime: Vict Descent × Vict Sex).
2026-06-24 — Five seam fixes after the live-artifact critique
The fed143/fed144 read held up under review; this pass closes the localized seams it surfaced.
- Single-source runs no longer point at an “inter-agency join named above” that isn’t there. The bridgeable-gap survivability names the inter-agency join only when the void actually crosses two authorities; a single-source run reads “a record carrying both values would close it.”
- Same-field aliases and facets no longer surface as collectable gaps. Two facets of one column — a numbered alias (
Crm Cd / Crm Cd 1) or a structural facet (X · populated / X · multiplicity) — now read as one field at detection time, matching the authoring guard, so they no longer appear as fillable gaps carrying a collection mandate and an inflated probability. They read as “the two columns are one field,” not a gap to collect.
- A genuine single-source gap reads honestly. When two distinct columns of one source have a value pairing that is simply absent, the precondition says both values are already co-recorded and names the real alternatives (a sparse slice, or a coding rule that excludes the pairing) — no false “collect the joined record.” (Codebook corpora like crime classifications.)
- Candidate-exclusion voids say why no collection is tasked. The DO block now closes with “No collection is tasked — absent a schema change that lets one record carry both, no record can occupy this cell,” so an empty task reads as intentional, not omitted.
- Multi-corpus convergence shows the boundary spread. When voids cross several source-authority pairs, the convergence block names the secondary pairs by count (e.g. “fdic x usgs 23; nih x fdic 9; …”), not only the dominant one.
- The moving-gaps lead separates a temporal lead from a worklist finding. A bridging line explains the trajectory pairs are drawn from the full record population (they co-occur in some snapshots), while the worklist carries only the pairs absent across every snapshot dense enough for an absence to mean something.
2026-06-23 — Three deepenings: cross-corpus institutional read, engine-side divergence, moving gaps
- No-shared-key cross-corpus runs now read as an institutional separation. When two sources share no join key, the engine builds a union graph; its dimensions are now source-prefixed (e.g.
nih:activity_code × usgs:review_status), so a cross-agency void is recognized as an institutional separation — the precondition names the inter-agency join, and the institutional-boundary convergence reads N such voids as one boundary, not a scatter of unexplained gaps. The adjudication stays honestly unvalidated (no joined corpus exists to test co-occurrence), but the structural read is surfaced rather than buried under a generic "no live source."
- The divergence read is now the engine's, not the page's. Where the deterministic verdict and the authored landing disagree, the detection is computed and stored on the finding (joining the artifact + the ledger) instead of re-derived at render time.
- Moving gaps over time. The engine now detects pairs whose co-occurrence is OPENING (emerging) or CLOSING (decaying) across the periods — gaps a whole-dataset scan misses because they co-occur in some years. A "Read across time" section surfaces them with their per-year trajectory.
2026-06-23 — The void anatomy: every finding now reads its own interior
A structural-void detector tells you where the wall is; the engine now tells you what the wall is made of. Each void carries two faces — the outward face (what surrounds it: the covering path, the record that would bridge it, the survival test) and a new inward face that reads the void's interior.
- The precondition profile (the inward edge). Each void names the real-world condition that would have to exist for the missing record to form — a collection mandate, a coding-rule change, a definitional change, or — when a cross-corpus void spans two authorities maintained apart — an institutional separation that no single-agency pull can close, only a data-sharing arrangement (or a common entity identifier) between them. It states the co-fire set the record requires and whether the absence is bridgeable or structurally mandated.
- The typed verdict, surfaced. The engine's read of why a void exists (bridgeable gap / candidate rule / structural rule / definitional exclusion / one field restated) now leads the inward face by name, not just by implication. The institutional-separation read also corrects the next move: not "pull from cross-corpus," but establish the inter-agency join.
- How well each void holds up, on the page. Each void now carries how well it holds up under challenge, beside its verdict: two integrity reads, two questions answered.
- The trajectory leads. Where a corpus carries a time axis, each void leads with its kinetic state — a durable wall, an emerging gap, a closing one — read across the snapshots.
- One institutional boundary, not N findings. When many cross-corpus voids cross the same pair of source authorities, the run reads them as one institutional separation observed many times, not a scatter of independent gaps.
- Where the engine's two tiers disagree, it says so. When the deterministic structural verdict and the domain-grounded authored read diverge, the engine states the case for and against each and hands adjudication to its own survival test — the deterministic verdict is the floor, the deep read the working hypothesis.
2026-06-20 — The membrane reads real columns before it transforms them
The ingestion membrane now runs semantic-first: it reads and understands the real columns of an uploaded dataset before the mechanical step translates high-cardinality fields into structure. Concretely, the engine now carries a per-column meaning (geographic, temporal, currency, person-attribute, identifier, classification-code, …), not just a data type, and that understanding reaches the output.
- A sanity gate resolves ambiguous columns. A column the first pass can't place — say a report number that looks like a plain category — is held, then resolved by deeper evidence (a near-unique column is an identifier, a multi-word near-unique column is free text). Resolved with a stated diagnosis rather than guessed.
- The membrane remembers. A learning profile library accumulates column-meanings across runs and recalls them — recognized once, recognized faster and more confidently every run after. Shared across the substrate, so every run teaches the same library.
- Ambiguous columns with no signal yet are left explicitly unresolved (awaiting a deeper pass) rather than mislabeled.
2026-06-19 — Source-record provenance: every validated finding traces to real rows
A confirmed void now carries exemplar source records — actual rows behind each side of the claim ("value A appears in records like {…}, M in all; value B in {…}, N in all; 0 records carry both") — plus an explicit invitation to recount them in your own upload. A finding is no longer a tag; it's a traceable claim you can verify to the row.
2026-06-19 — Population-complete cross-corpus findings
When a cross-corpus read covers the full population of its join key (e.g. all 50 states), the engine no longer mislabels a clean absence as an "underpowered lead." The absence is the realized population truth: it's reframed as a structural fact, and confirmed when the overlap was expected yet never occurs across the whole population — while genuinely modest gaps stay honestly modest. No manufactured confidence.
2026-06-19 — The association read stops surfacing missing-data artifacts
The Intelligence (association) section now demotes shared-missingness co-movements — two fields that are empty in the same records (unknown codes, or a column the membrane marked empty-here) travel together as an artifact, not a real linkage. The strongest real linkage now leads the section instead.
2026-06-18 — Honest validation, made legible
- Two-corpus provenance. Each run states plainly that the graph maps a bounded sample and every surviving void is then validated against the full uploaded corpus, with the per-pair counts shown — so the validation reads as real, because it is.
- The validation arm only counts what's present. A void is confirmed only when both its values actually appear in the validated corpus; a value that isn't there leaves the void untested, not vacuously "confirmed."
- Memory is context, not a vote. A prior run surfacing the same void is flagged for context but never counts as independent corroboration — the confirmed verdict rests on the live full-corpus check alone.
Foundation — the structural-void engine
- Seven analyst lenses. Every run is read through Intelligence, Investigation, Program Assessment, Policy/SOP, Compliance/Audit, Adjudication, and CQI — the engine surfaces only the lenses that find structure in your data.
- Every void is challenged before it surfaces; what remains has already survived that challenge. Voids that are one field restated, or a gap in coverage, are refuted and taken off the worklist.
- Full-corpus validation. Each surviving void is checked against the full uploaded corpus (well beyond the detection sample), so a sampling artifact is caught and a real absence is confirmed by counted fact.
- Convergence anchor. When most voids route through one axis or entity, the engine foregrounds the single boundary seen many times rather than reporting many separate findings (protected under U.S. Provisional 64/043,294).
- Cross-corpus weave. Select two or more sources and the engine folds them on their shared axis, surfacing where one source carries a structure the other never does — the gap that would have to be collected to close it.
- Corpus-regime honesty. A run with no live source to validate against says so up front — "structural candidates, not validated findings" — so untested is never read as rejected.