Emerging Chokepoint Early Warning

Methodology

The Early Warning feed is a research-grade indicator surface, not a news feed. Its inputs are ownership-change events from a licensed corporate-ownership database, filtered to the AI-compute-relevant universe (only Buyout, Corporate Divestiture, Asset Sale, Public-to-Private, Secondary Buyout, Spin-Off, Merger/Acquisition deal types — not seed rounds, growth equity, or private placements). Each event is mapped to one of the seven chokepoint layers via the rules in src/lib/data/chokepoint_layer_definitions.json. Inclusion rules are deliberately conservative: an organization is classified into a layer only when its database industry hierarchy plus a chokepoint-specific keyword (or a canonical-name match) co-occur with semiconductor or datacenter context. A second deal-target-only fallback widens coverage on the 129 database-linked deal targets to cases where industry-axis evidence is unambiguous but explicit chokepoint keywords are absent — those are tagged MEDIUM confidence and clearly distinguishable in the substrate.

Anomaly threshold: a quarter fires when its event count exceeds the trailing-eight-quarter mean by more than two standard deviations and by at least two events in absolute terms. The secondary gate prevents singleton-quarter firings against a near-zero trailing baseline (the mathematical pathology of z-scores on small counts). Threshold z > 2 is calibrated so roughly 2.5% of normal upper-tail quarterly variation crosses it. When the trailing window is uniform (std = 0), z is reported as null rather than computed — degenerate z-scores are never rendered.

Consolidation cluster: the same canonical acquirer makes ≥3 deals in the same chokepoint layer within any 12-month rolling window. Cross-layer actor: the same canonical acquirer appears in ≥2 chokepoint layers within 12 months. Both are computed against mv_chokepoint_layer_acquirers, which joins pe_deal_investors filtered to party_role = 'ACQUIRER'. Acquirer jurisdiction is drawn from organizations.country_hq and tagged with the four-class coordinated-posture taxonomy (allied-coordinated / allied-partial / non-aligned / restricted) loaded from jurisdictional_regimes.json. The same regime version powers the Allied Coordination Gap surface; the cross-link from any non-allied transaction lands the reader on /coordination-gaps reading the same posture for the same country.

Disclosure-rate caveat: deal counts are complete; dollar aggregates reflect the ~46% of database deals that disclose size. Dollar-weighted comparisons across layers should be read with this asymmetry in mind. For consolidation clusters, the per-cluster disclosed sum is approximate (we redistribute the per-quarter disclosed sum across deals within the quarter; see src/lib/chokepoint/anomaly-engine.ts).

Substrate freshness: this snapshot uses the deal universe of a licensed corporate-ownership database (held under subscription; not redistributed), extracted on 2026-04-24; the database's own lastupdated MAX in that snapshot was 2026-04-08. Data lag is structurally one to three weeks: the licensing platform refreshes weekly, and database ingests vary per record class. The chokepoint-early-warning cron runs Mondays 07:00 UTC and rewrites this surface from the latest substrate.

Known limitations:

• Universe filtering. The database substrate is filtered to entities relevant to AI-compute governance. Pure-play chokepoint vendors that don't intersect this scope (e.g., Ajinomoto's substrate division as a chemicals firm) won't appear; only deals against governance-scoped entities surface here.
• Acquirer canonicalization. ~78% of pe_deal_investors rows currently have a canonical party id. Deals without an identifiable acquirer can't contribute to cluster or cross-layer detection. This will improve as substrate resolution coverage grows.
• Quarter discretization. The activity feed bins deals into calendar quarters. Deals at quarter boundaries land in one bucket or the other; the trailing window is computed over populated quarters (not calendar quarters), which avoids inflating z-scores after long dry spells.
• Conservative rules trade recall for precision. Layer counts in the table below show how many entities each rule currently classifies. Rules can be tuned in chokepoint_layer_definitions.json; an internal classification audit report lists known-entity hits and misses per layer.
• Match-reason transparency. Every classification carries a classification_reasons array (the inclusion rule that fired); the per-layer drill-down below exposes a sample so the placement of any given org is auditable rather than asserted.

Retrospective accuracy:

The cite-this permalink for every anomaly is structurally stable: the row id in chokepoint_anomalies is the citation handle, the substrate is append-only, and re-runs of the anomaly engine update last_seen_at rather than replacing rows. A reader citing "Anomaly X flagged in Q3" and re-loading six months later sees the same id, the same evidence transactions, and the same per-layer thresholds — the retrospective contract is that the substrate the flag was made against does not silently rewrite. Anomalies that subsequently fail to recur transition to the archive surface but retain their original ids; the archive count next to the headline KPI (2 archived against 1 active) is the running tally.

What this surface does notyet publish is a per-anomaly outcome ledger — a structured "flagged Q3 2026; here's what the supply-chain reality of that layer looked like by Q1 2027" column. That requires monthly retrospective annotation against external indicators (BIS designations downstream of the flagged actor; subsequent press coverage; same-acquirer follow-on deals) and a curation cadence Scrutica has not yet committed to. Adding that surface honestly means standing up the curation pipeline first; the page will not synthesise outcomes from the anomaly engine's own recurrence signal because that would be circular by construction (the engine surfaces anomalies that recur — using its own recurrence signal as ground truth would inflate apparent accuracy).

Substrate snapshot: 29 pe_investments rows · 98 classified orgs · 27 (layer, quarter) cells with activity · 1 active anomalies · 2 archived anomalies. Anomaly engine last ran: 2026-06-08 07:00:01 UTC. Classification last ran: 2026-04-25 21:32:04 UTC.