empty-chair

Most detectors hunt the suspicious thing that is present: the nominee with 400 directorships, the offshore address. This one hunts the thing that is deliberately absent. It scores a UK company for how much its public ownership disclosure is shaped like the structures where a hidden beneficial owner was later revealed, in the ICIJ offshore leaks or on the OpenSanctions consolidated list (sanctions regimes plus debarment, crime and PEP-linked watchlists). The empty chair where an owner should be.

It reports a signal, not a verdict. A high rank means the disclosure resembles a known concealment shape, never that the company hid anyone or that any person did anything wrong. Most companies with this shape are legitimate: holding companies, family property firms, dormant shells with nothing to hide.

The result

empty_chair v11, a single tuned LightGBM over the registry and PSC (people-with-significant-control) disclosure features plus tell interactions and a grouped out-of-fold target encoding of the office post area. Held out by formation-address cluster so no company mill straddles train and test. The model carries no sector identity: no SIC code, no SIC section, no sector encoding. Earlier versions scored higher with sector in; the bias audit retired them (see the model cards).

metric (grouped holdout, 21,350 companies)	value
PR-AUC	0.199
PR-AUC lift over base rate	4.2×
ROC-AUC	0.848
precision@100	0.42
precision@1000	0.233
blind investigator rule (PR-AUC)	0.049
demographics-only control (PR-AUC)	0.070
shuffle-label control (PR-AUC)	0.049

The naive investigator rule (flag anything silent, corporate-only, or foreign-corporate) scores at the base rate: useless alone. A demographics-only model (incorporation year, sector, region) barely beats chance. The shuffle-label run collapses to chance, so there is no leak, target encoding included. Of the 100 holdout companies the model flags hardest, 42 are genuine later-revealed cases at a 4.8% base rate.

The number is half of what this model family can score, on purpose. With sector features in, the same recipe reaches PR-AUC 0.377, and scoring the full 5.7M universe showed where that extra skill lives: the top 1% was 99.3% real estate, while real estate is only 13.8% of the training positives, and the most label-enriched sector (financial holding, 3.2x) was absent from the top entirely. Sector features let the model park its confidence on one easy cluster instead of reading each disclosure. The trail from 0.377 to 0.199 is in docs/bias-audit.md and the model cards below.

Caveats

Read these before quoting the number.

• PU-learned, the number is a lower bound. Positives are companies whose hidden interest was later revealed. Hidden owners never revealed sit unlabelled in the clean class, so every metric understates true performance.
• The label is a name match. ICIJ and sanctions names are matched to Companies House by normalized exact match. Single-token generic names are down-weighted, but the match set carries false positives that no one has hand-verified; treat the positive labels as noisy.
• Rank, not probability. The model is uncalibrated by design (a monotone map cannot improve a ranking metric). The app presents a population percentile, never a probability of concealment.
• Disclosure shape, not intent. The model never sees who owns the company. A high rank is a structural resemblance, never proof of concealment.

Architecture

An FTI (feature, training, inference) system on Hopsworks. Feature extraction is one shared, pure function (chair_features.py) so training and serving cannot skew.

flowchart LR
    ch([Companies House bulk]):::ext
    icij([ICIJ leaks + OpenSanctions]):::ext
    subgraph FE[Feature]
        icij --> f1[build_labels] --> lab[(revealed_owner)]:::hops
        ch --> f2[ingest_registry] --> reg[(company_registry)]:::hops
        ch --> f3[ingest_psc] --> psc[(psc_shape)]:::hops
        lab --> f2
    end
    subgraph TR[Training]
        reg --> fv{{empty_chair_fv}}:::hops --> t[train LGBM vote + controls] --> m[(Model Registry)]:::hops
        psc --> fv
    end
    subgraph INF[Inference]
        m --> s[score_universe] --> dos[(concealment_dossiers)]:::hops
        s --> pq[(universe_scores.parquet)]:::hops --> app[emptychair app]
    end
    classDef hops fill:#10b98122,stroke:#34d399,color:#e5e7eb;
    classDef ext fill:none,stroke:#6b7280,color:#9ca3af,stroke-dasharray:4 3;

Loading

The inverse framing is the core design. The features describe the shape of the disclosure, not the owner: whether a natural person is declared at all, whether ownership routes only through corporate entities, whether the company sits silent behind a no-PSC statement, an exemption, or a super-secure protected record, whether it is registered at a formation-mill address. The label is what the disclosure was hiding, revealed after the fact.

The file-by-file map:

chair_features.py     shared, pure: CH row + PSC records -> disclosure features + fired tells
build_labels.py       F1  ICIJ + sanctions -> revealed_owner (label)      (job)
ingest_registry.py    F2  CH bulk -> company_registry (case-control 20:1)  (job)
ingest_psc.py         F3  PSC snapshot -> psc_shape                        (job)
train_chair.py        T   feature view -> empty_chair + honesty controls   (job)
auditor.py            I   load model, score a company, list fired tells
score_universe.py     I1  score all 5.7M -> parquet + concealment_dossiers (job)
build_linkage.py      I2  top 1% -> shared-owner nests (linkage.parquet)   (job)
ingest_links.py       F4  PSC snapshot -> owner_degree FG (tested, not adopted) (job)
explain.py            I   plain-language dossier (Anthropic), signal not verdict
ask.py                I   ask-the-register: tool-use loop over the live data
bias_audit.py         pre-publication confound audit -> docs/bias-audit.md
app/server.py         the review app: audit, chair diagram, webs, ask
app/deploy_app.py     deploy the app

Data

All public and free. Companies House basic company data + PSC snapshot (bulk, Companies House), ICIJ Offshore Leaks database (bulk CSV), and the OpenSanctions consolidated targets export (targets.simple.csv, default dataset: sanctions plus debarment, crime and PEP-linked watchlists; GB entities only). The two heavy captures are kept out of git; build_labels and ingest_registry rebuild every feature group from them.

Reproduce

Clone into a Hopsworks project on the /hopsfs/... FUSE mount. Paths self-derive. The Anthropic key lives in a project secret (ANTHROPIC_API_KEY).

# capture the bulk data into data/ (Companies House, ICIJ, OpenSanctions)
python build_labels.py                       # F1  revealed_owner
python deploy_registry.py && hops job run ingest-registry   # F2  company_registry
python deploy_psc.py && hops job run ingest-psc             # F3  psc_shape
python deploy_train.py && hops job run train-chair          # T   empty_chair
python deploy_score.py && hops job run score-universe       # I1  scores + dossiers
python app/deploy_app.py                     # the review app

The demo

emptychair, a two-pane review. Paste any UK company number or name: the company's disclosure evidence renders on the left (every tell with its population base rate, fired ones dark), and a rank rail on the right shows the percentile stamp, the score distribution with the company pinned on it, and streams a plain-language investigator's note. When the company sits in a scored nest, its ownership web renders below: owners as squares, companies as dots colored by score, shared companies as red bridges.

/network shows the concealment webs: nests that share a company are unioned into connected graphs (the largest joins 100+ owners). All graphs are server-rendered SVG that works without JavaScript; with it, they hydrate into pan/zoom, ego-highlight on hover, tooltips, and draggable nodes over a spring simulation seeded from the server layout. The rank comes from the ML model; the note only explains it.

Every page carries Ask the register, a full-height chat drawer pinned to the right edge. The model (ask.py) answers only through deterministic tools over the live data: company lookup, ownership webs, name and owner search, population stats. No embeddings, no invented numbers; if a tool returns nothing, the answer says so. Tokens stream over a websocket (the platform proxy buffers plain HTTP streaming) with each tool call shown as a status row. The chat is contextual: it knows which page is open, and clicking any owner square in a graph points the conversation at that owner. Replies interpret the figures concretely, percentile arithmetic and tell base rates included, under the same never-accuse constraints. Without JavaScript it degrades to a full-page round trip.

Cost note

Universe scoring with the single-seed model takes about 30 minutes for 5.7M companies; a 10-seed soft-vote takes just over two hours and, on the sector-blind recipe, scored 0.196 to the single seed's 0.199 (v10 vs v11 in the registry). The vote buys top-of-ranking stability across retrains, not accuracy.

Model cards

One registry, empty_chair. All versions train on empty_chair_fv v1 and evaluate on the same grouped holdout (21,350 companies, 4.7% base rate). Scores are ranks, uncalibrated by design.

v11, production

• Single LightGBM: 1500 trees, lr 0.02, 31 leaves, min_child_samples 20, no row or column subsampling. Tell interactions, grouped OOF target encoding of post_area. No sector identity; the register's own activity-type flags (holding, dormant) stay in as tells.
• PR-AUC 0.199 (4.2x base rate), ROC-AUC 0.848, P@100 0.42, P@1000 0.233. Controls: shuffle 0.049, demographics-only 0.070.
• Universe audit: top 1% is 34.7% real estate (3.2x), holding-SIC moves the mean score by nothing (0.112 vs 0.122). docs/bias-audit.md.
• Intended use: anonymous structural triage, an ordering of where to look.

v5, retired

• 10-seed LightGBM soft-vote with tell interactions, target encodings of post_area, sic_section and sic_code, plus raw sic_section. Winner of the jul14 autoresearch round (22 experiments, autoresearch/REPORT.md).
• PR-AUC 0.377 (8.0x), ROC-AUC 0.882, P@100 0.95, P@1000 0.373. All controls passed.
• Retired by the universe audit: top 1% was 99.3% real estate at 9.1x lift and holding-SIC alone moved the mean score from 0.206 to 0.309. A strong ranker that had become a corporate-structure detector.

v6 to v10, the ablation trail

v	change	PR-AUC	top 1% real estate
6	drop sic_code TE (single seed)	0.371	88.3%
7	drop sic_section TE too	0.362	97.5%
8	drop all structure features	0.184	not scored
9	drop only raw sic_section	0.186	not scored
10	v9 recipe, tuned, 10 seeds	0.196	not scored

The TE ablations (v6, v7) barely moved the concentration because the raw sic_section categorical carried the sector the whole time. Removing it (v9) halves the measured PR-AUC; the audit on v11 indicates the half that remains reads the disclosure. An owner-degree feature (companies per named PSC owner, owner_degree FG) was tested against v11 and rejected: it found formation agents, hurt PR-AUC, and never touched the labels.