Regime Signal | Methodology & Settings

1. What Regime Signal Tracks

Regime Signal is healthcare AI labor intelligence for investors. It tracks where automation intensity is rising across healthcare labor workflows, then links each read to observable records.

The product is built for healthcare investors first. Healthtech founders and operators may use the same evidence, but the core job is investor diligence.

2. Evidence Chain

DataSource -> MetricObservation -> SignalEvent -> IntelligenceBrief

Regime Signal does not treat one record as proof of causality. It links observable evidence into a disciplined read that an investor can inspect, challenge, or use in diligence.

DataSource

A public or curated record, such as a workforce notice, job posting, SEC filing, earnings transcript, or vendor announcement.

MetricObservation

A structured field from that record: date, employer, role text, worker count, function tag, source link, ticker, or keyword hit.

SignalEvent

A pattern across records, such as more patient access roles appearing across reviewed sources over a defined period.

IntelligenceBrief

The written investor read, with links back to the evidence that supports it.

3. What Regime Signal Is Not

In scope

Healthcare labor signal monitoring
Workflow-level evidence review
Capture-profile reads for public companies
Diligence support for healthcare investors

Out of scope

Stock picking
Generic AI dashboards
Causal claims from single records
Automated diligence replacement

4. Sources, Coverage, and Universe

Source	Role	Status
Workforce change records	Named changes by employer, state, date, role text, and worker count when the source provides it.	Evidence layer ready. Source selection not yet approved.
SEC EDGAR	AI, labor, automation, and operating commentary in 10-K, 10-Q, and 8-K filings.	Used for selected examples in Phase 0.
Job postings and occupation data	Broad labor pressure context and occupation exposure mapping.	Existing signal layer.
Curated company attributes	Capture profile, business model, labor intensity, and mapping overrides.	Manual review layer.
Weekly vintages	Point-in-time output bundles for review and future tests.	Active.

The current public-company universe contains 532 tickers. The short-term source of truth is apps/web/ticker-universe.js. The tracked securities database is future work and is not yet the live source of truth.

5. Healthcare Function Taxonomy

Function	Definition
Patient access	Scheduling, registration, intake, eligibility, contact center, and patient communications.
Revenue cycle management	Billing, collections, claims operations, denials, and payment posting.
Medical coding and CDI	Medical coding and clinical documentation improvement work.
Prior authorization	Prior authorization intake, submission, follow-up, and appeals support.
Care navigation and case management	Care navigation, referral coordination, and case management support.
Clinical documentation	Ambient documentation, scribing, transcription, and clinical note support.
Nursing support and virtual nursing	Nursing support, virtual nursing, nurse triage, and care team extension roles.
IT and EHR support	Health system information technology, electronic health record support, help desk, and application support.
Finance and administration	Finance, accounting, human resources, procurement, and general administration.
Pharmacy operations	Pharmacy operations, pharmacy technicians, fulfillment, and medication support.
Imaging operations	Radiology operations, imaging scheduling, modality support, and imaging technician support.
Other healthcare	Healthcare work that does not map cleanly to a tracked function.

6. Source Record Classification Method

Source rows enter the public layer only after they clear a confidence floor. Low-confidence rows remain in review and do not enter public totals.

Candidate rows are selected through healthcare North American Industry Classification System codes, curated healthcare employer names, or healthcare-specific role text.
Role text is classified against the constrained healthcare taxonomy.
The row stores function, confidence score, model name, model version, review status, and source metadata.
Generic titles receive low confidence unless the role text contains a specific healthcare function signal.

The healthcare evidence layer is live, but source selection is paused for Phase 0a. Public workforce-reduction records, including WARN notices, may become one provider inside this layer, but they are not the product.

7. Capture Profiles

Capture profiles translate a workflow signal into a company read. They separate companies that buy labor from companies that sell labor or sell the tools used to reduce it.

Labor buyers

Health systems, payers, and operators with large administrative workforces.

Labor sellers

BPOs, staffing firms, and service providers exposed to healthcare administrative work.

Software enablers

Healthcare information technology vendors and software platforms with automation upside.

Hardware enablers

Device, diagnostics, imaging, and automation hardware names with healthcare AI exposure.

Mixed

Companies with both labor pressure and automation upside.

8. SEC Disclosure Scoring

SEC disclosure scoring looks for healthcare-relevant AI, automation, labor, workforce, scheduling, contact center, and operating-efficiency language in public-company filings.

For Phase 0, the inaugural brief should use selected named examples rather than aggregate SEC coverage statistics. Full healthcare SEC coverage repair is Phase 1.

9. Separate Advisory Practice Firewall

The operator runs a separate healthcare go-to-market advisory practice. Clients of that practice are explicitly excluded from Regime Signal coverage.

Regime Signal copy may reference direct healthcare operating experience. It should not use client-identifiable information from the separate practice.

10. Limits and Review

Regime Signal is useful when it makes evidence easier to inspect. It is weak when it outruns the records. Each brief should state what would change the read, and ambiguous rows should stay visible to reviewers before they affect public totals.

Coverage

Waiting for coverage data

Freshness

Waiting for freshness data

Evidence Strength

Waiting for evidence strength data

Quality Trend

Coverage, freshness, and evidence strength

8-point recent status history.

Sources & Freshness

Freshness pending

Loading live source coverage and timestamps...

Waiting

Source	What it does	Cadence	Latest loaded
Loading live source freshness...

Active Data-Quality Controls

Control	Status
Timestamps	Active
Versioned outputs	Active
Snapshot metadata	Active
Coverage labels	Active
Uncertainty intervals	Active

Active Module Versions

Module	Version	Endpoint
Loading modules...	--	--

Runtime Lineage and Governance

Implemented controls include source timestamps, versioned outputs, snapshot metadata, visible coverage labels, and immutable weekly vintages. Pending governance items include published source-diff reporting, formal benchmark protocols, and external-validity work.

Advanced Diagnostics

Core series observations

Last refresh

Active sources

Clusters detected

Suppressed signals

Regime confidence

Backtest Sharpe

Not published

False positive rate

Validation pending

Calibration

Parameter registry

Visual controls show current calibration values without changing the underlying data pipeline.

Composite Weighting

Weights auto-normalize based on configured domains.

Telemetry (z=1.6)Task Exposure (z=1.2)Hiring Mix (z=0.9)Financials (z=0.4)Churn (z=0.6)Reskilling (z=0.3)Postings vs Employment (z=0.8)Productivity Shift (z=0.5)

Signal State

Composite regime signal (normalized)

1.24Watch

Persistence rules

Watch

Condition true in 2 of last 3 periods.

Warning

Condition true in 3 of last 4 + 2 domains confirm.

Alarm

Condition true in 6 of last 8 + 3 domains confirm.

Thresholds

Set cutoffs for watch, warning, and alarm regimes.

Watch ≥Warning ≥Alarm ≥

Parameter	Current Value	Description	Used In
Regime Smoothing Window	28 days	Smoothing window for the labor-demand signal.	ewi_from_fred.py
Regime Baseline Window	365 days	Trailing baseline for z-score normalization.	ewi_from_fred.py
Z-Score Alert Thresholds	1.0 / 1.5 / 2.0	Watch, Warning, Alarm thresholds.	regime.v2
Automation Intensity Tier Breaks	45 / 60 / 75	Intensity tier cutoffs.	risk_matrix.v2
Lead-Lag Horizons	30 / 60 / 90 / 180	Directional horizon set.	lead_lag.v1
Divergence Latent-Margin Rule	gap < -1.0 and 4w Δ < 0	Latent margin signal trigger.	divergence.v1
Labor Confirmation Thresholds	3m +0.10 / 6m +0.20	Momentum thresholds for unemployment confirmation.	labor_confirmation.v1
False-Alarm Dampener	26 weeks	Downgrades unconfirmed weak-postings signals.	labor_confirmation.v1
SEC Disclosure Lookback	120 days	Maximum filing lookback window.	sec_corporate_disclosure.py
SEC Max Filings per Ticker	2	Latest filings sampled per ticker.	sec_corporate_disclosure.py
SEC Direction Thresholds	+2.0 / -2.0	Rising vs falling disclosure cutoffs.	disclosure.py

Implemented and Test-Covered

regime
weekly insights
automation intensity matrix
lead-lag
divergence
watchlist alerts
implications
portfolio support / capture layer
SEC disclosure
weekly vintages
API smoke testing
versioned outputs

Remaining for Formal Sign-Off

Pre-registered historical benchmark protocol
Calibration and error-rate reporting by horizon
Revision-impact studies with vintaged snapshots
Ablation and sensitivity studies
External-validity testing across regimes

The system is production-grade for directional weekly decision support, but users should treat it as a structured analytical input, not a validated predictive engine.