Ethical Intelligence
Methodology

Ethical Intelligence Principles

EcoIQ is built on six core principles drawn from systems thinking in industrial stewardship, long-term resilience theory, and evidence-based governance frameworks. These principles govern how scores are weighted, how harm is penalised, and how improvement trajectories are recognised.

Improvement, Not Punishment

EcoIQ is not an ESG policing tool. It is a transition intelligence system designed to help industrial companies understand where they stand, what holds them back, and how to unlock the financing and credibility that responsible modernization enables.

6-Pillar Scoring Framework

EcoIQ scores are calculated across six weighted pillars, each decomposed into measurable sub-dimensions. Weights reflect the relative importance of each dimension for long-term industrial resilience and public benefit creation.

Scoring Formula

EcoIQ Total Score is a weighted composite of six pillar scores, reduced by a harm penalty. All pillar scores are independently calculated on a 0–100 scale from their respective sub-dimension averages.

Harm Penalty System

Systemic harm is penalised by direct score reduction — not merely flagged. The Harm Penalty system creates an explicit incentive structure for harm reduction, as improvements in each signal directly increase the total score.

Score Interpretation Tiers

EcoIQ scores map to six institutional tiers, each reflecting a distinct level of industrial stewardship, transparency, and transition readiness. Tiers determine financing eligibility, investor screening status, and recommended engagement type.

Formula Category System

EcoIQ's 33 scoring formulas are organised into seven thematic categories, each mapping to a distinct area of industrial and ethical intelligence. This structure enables modular analysis — individual categories can be weighted differently for sector-specific or programme-specific assessments.

Transition Intelligence

EcoIQ includes a Transition Intelligence layer that converts a company's current score into an actionable improvement roadmap. This layer identifies the highest-leverage opportunities, models the score impact of each action, and maps companies to relevant financing mechanisms.

Institutional Use Cases

EcoIQ is designed for institutional users who need evidence-based intelligence for climate investment, industrial policy, and transition finance decisions.

Ethical Intelligence Framework

Beyond the 6-pillar EcoIQ score, EcoIQ provides an Ethical Intelligence Analysis layer that compresses the full analytical framework into three master formulas — offering a higher-order view of a company's net societal impact, transition trajectory, and long-term value creation potential.

Internal Formula Architecture — 8 Analytical Categories

The three master formulas are computed from 33 internal sub-formulas organised across eight analytical categories. This architecture allows modular sector-specific weighting while preserving the integrity of the composite output.

Company Discovery, Shariah Screening & 114-KPI Stewardship

A separate, evidence-first research system layered on top of the EcoIQ Intelligence Score above. Discover Companies → lets a user filter and rank real, publicly listed companies across two independent lenses that are never blended into one number.

What this is — and is not

EcoIQ does not provide personalised investment advice. Discover Companies' ranking reflects evidence-backed stewardship relevance under whatever criteria a user selects — it is not expected financial performance, not a price target, and never uses BUY/SELL/HOLD/STRONG BUY/TARGET PRICE/EXPECTED RETURN/UNDERVALUED/OUTPERFORM language anywhere in this system.

Preliminary Shariah Screening

A named, versioned, deterministic methodology (business-activity exclusions + financial-ratio thresholds) — never an invented "Islamic ruling". Every result is labelled "Screened according to [methodology name] v[version]", never "Halal certified", unless an actual third-party certification exists (none does today). Results are PASS / CONDITIONAL / FAIL / INSUFFICIENT_DATA / NOT_SCREENED — a missing required financial input is recorded as missing and excluded from ratio calculation, never silently treated as zero.

The 114-KPI Stewardship Framework

Every company assessment against the 114-item Capital Ethics Compendium is either (a) a real, provenance- tracked evidence link a human reviewer confirmed, or (b) a deterministic status derived from the aggregate of confirmed evidence relationships (SUPPORTS/CONFLICTS/CONTEXT) — never an LLM asked "does this company support KPI Y?" with the answer stored as fact. A deterministic keyword-overlap matcher may PROPOSE a candidate link from newly ingested evidence (a sustainability report, an ESG report, an annual report), but a proposed link can never move a company's KPI status until a staff reviewer explicitly confirms it — see the Evidence Review workflow on each company's page.

Candidate vs Confirmed Evidence

review_state on every KPI evidence link is one of proposed / confirmed / rejected. Only confirmed links count toward a company's displayed KPI status, its discovery ranking, and its Explain Match trace. A proposed link is visible and inspectable, but inert until reviewed — this is what keeps automated candidate matching honest even as more real evidence is ingested.

Discovery Ranking — Components, Not a Mystery Score

Ranking combines five documented components, each shown alongside the composite, never hidden behind it:

• kpi_alignment (default weight 0.40) — supported vs conflicting evidence for the selected KPI(s). Absence of evidence is neutral, never negative: an unassessed KPI contributes the same neutral value as a KPI with only context-only evidence, never a penalty.
• source_authority (0.20) — average real source-quality score of the evidence backing confirmed KPI links, from the harvester verification engine.
• recency (0.15) — average freshness of that evidence.
• corroboration (0.15) — average independent-source corroboration of that evidence.
• data_completeness (0.10) — the latest Shariah screen's own honest completeness percentage.

Weights are configurable (passed explicitly to the ranking service; not hardcoded into a single formula a user cannot inspect). A component that is genuinely unavailable for a company (no harvester-backed evidence, no Shariah screen run) is EXCLUDED from the composite's weighted average — never coerced to zero. A company with no qualifying evidence at all shows "No qualifying evidence", sorts last, and is never displayed as "0 alignment".

Evidence Quality & the Source-Tier Hierarchy

Every source is one of four tiers: Tier 1 — regulatory filings, audited statements, government records (SEC EDGAR, Companies House, FCA filings); Tier 2 — official annual and sustainability reports published directly by the company; Tier 3 — credible independent research (CDP, GRI, SASB, ISSB, multilateral datasets); Tier 4 — marketing and self-reported claims (company website copy, press releases). Discovery's evidence-quality filter lets a user require Tier 1 only, Tier 1–2, or Tier 1–3 — results change based on real evidence availability; a filter with no qualifying evidence honestly returns zero results, never a fabricated match.

Freshness

A Shariah screening is CURRENT, STALE (>180 days since screening or since the underlying financial data's reporting period, whichever is older), or NOT SCREENED. A stale screening displays "Screening Requires Refresh" — it is never silently retained or ranked as equally current.

Conflict Treatment

Conflicting evidence and controversies are never suppressed by a positive-looking ranking elsewhere on the same company. A company can have both strong positive stewardship evidence AND a credible, unresolved controversy — both are shown, side by side, on every company profile and in every discovery result.

Demo vs Real Data

Every company-intelligence record carries an honest is_demo flag. Discover Companies excludes DEMO/illustrative companies by default — a user must explicitly opt in to see them, and they remain clearly labelled throughout. Data origin is computed live from the real flags present, never a static label that could silently drift from what the underlying data actually is.

Evidence Review Workbench — Human Governance

The bridge between "the deterministic matcher proposed this KPI relationship" and "this evidence genuinely supports/conflicts with/merely mentions this KPI" is always a human decision. This is a governance and verification layer — not an investment recommendation engine, not an automatic AI approval system, and not a replacement for Evidence Memory or the 114-KPI framework itself.

Candidate vs Confirmed Evidence

Machine-generated candidate matches are not treated as verified evidence until reviewed. A deterministic keyword-overlap matcher (never an LLM in this codebase) proposes a CompanyKPIEvidenceLink with review_state='proposed' from newly ingested evidence. Only confirmed links count toward a company's displayed KPI status, its Discovery ranking, and its Explain Match trace — a proposed, disputed, rejected, or needs-more-evidence link never does, regardless of how confident the matcher's keyword overlap looked.

The Human Review Workflow

A reviewer's decision separates MATCH VALIDITY (does this evidence really discuss this KPI at all?) from EVIDENCE RELATIONSHIP (what does it conclude?). Seven explicit actions:

• Confirm — Supports / Confirm — Conflicts / Confirm — Context Only / Confirm — Insufficient to Conclude — a reviewer decides not just that the match is valid, but which of four distinct relationships the evidence actually establishes. A valid match is never automatically collapsed into "Supports".
• Reject Match — the evidence does not genuinely relate to this KPI; never counts.
• Needs More Evidence — a real, visible "unresolved" state, distinct from the original proposal — a reviewer looked and could not yet decide.
• Mark Disputed — even a previously CONFIRMED link is not permanently unquestionable; disputing it immediately stops it counting toward the company's KPI status until re-reviewed.

Every action requires an explicit named reviewer and a written reason — there is no code path that moves a link to confirmed without a human argument recording who decided and why. No auto-confirm. No "high confidence therefore verified." No bulk-confirm — only a conservative bulk "mark needs more evidence" is offered, and every affected link still gets its own individually attributed audit row.

Disputes and Re-Review

A confirmed relationship can be disputed at any time. A disputed link honestly stops contributing to the company's KPI assessment, Discovery ranking, and Explain Match the moment it leaves confirmed — no separate propagation logic is needed, because every downstream view already filters strictly on review_state == 'confirmed'. Re-review is simply applying any confirm/reject/needs-more- evidence action again; the full prior history is preserved, never overwritten.

How KPI Assessments Update

After every review decision, the affected CompanyKPIAssessment is recomputed through the existing, unmodified 114-KPI engine — never a hardcoded status flip. Confirmed supporting evidence may move an assessment to Support/Strong Support; confirmed conflicting evidence surfaces as Conflict; rejected/needs-more-evidence/disputed links never contribute. This is the same engine Company Discovery, Explain Match, and the company profile page all read from, so a review decision propagates consistently everywhere without any view-specific special-casing.

What AI Can and Cannot Do

AI assistance in the review workflow (summarising long context, highlighting relevant sentences) is optional and, in this PR, not yet built — the workbench ships with deterministic candidate matching only. If AI assistance is added in a future PR, any suggestion must be visually labelled "AI ASSISTANCE — NOT A REVIEW DECISION", must never itself confirm, reject, or resolve a dispute, and every model call must be recorded as a real ModelInvocation in AI Observatory — never a second, unaudited AI path.

Audit Trail

Every EvidenceReviewAction is an immutable row recording the reviewer, timestamp, previous review state, new review state, the relationship decision (where applicable), and the written reason — never edited or deleted. A link's current state can change; its history never does.

Stewardship Universe — Automated Source Discovery & Evidence Refresh

Turns manual, company-by-company document registration into a governed, repeatable pipeline: discover authoritative sources → register them → fetch and version documents → extract evidence → propose KPI candidates → send to human review → recompute company intelligence → record everything in AI Observatory. Automated source discovery and evidence extraction do not constitute independent verification. KPI relationships become confirmed evidence only after governed human review in the Evidence Review Workbench — this page never replaces that gate, only feeds it more real candidates.

What Makes a Company "Tracked"

A CompanyProfile carries an explicit tracking_status — NOT TRACKED, ACTIVE, REFRESH IN PROGRESS, PAUSED, or ERROR. This is deliberately a coarse, process-level lifecycle only; the richer day-to-day state shown on the Stewardship Universe page (Needs Source Discovery / Review Required / Needs Refresh / Current) is computed live from real conditions every time the page loads — never stored, since a cached copy would go stale the moment anything changed underneath it. A company is never shown as CURRENT while it has zero known sources or unreviewed evidence, no matter how long ago it was last refreshed.

How Sources Are Discovered

Exactly three real, deterministic discovery methods — no generic web search, no guessed domains, no fabricated results:

• SEC EDGAR identity — a company with a mapped CIK gets its regulatory filing history as a Tier 1 source, auto-approved (a government filing index is definitionally authoritative).
• Companies House identity — the same for UK-registered companies, used for identity, status, and filing metadata, never presented as sustainability evidence it does not provide.
• EcoIQ's curated official-domain registry — a small, manually-verified list of exact sustainability-report URLs on companies' own domains, auto-approved as Tier 2.

A fourth, lower-trust path surfaces a company's own pre-existing, staff-entered profile fields (annual report / sustainability report URLs) as CANDIDATES — real URLs, but never independently domain-verified, so they always require explicit staff approval before being registered as a trusted source. A domain is marked VERIFIED only when it matches EcoIQ's own curated registry; otherwise it is PROBABLE and cannot silently become a Tier 1/2 source.

Source Registry & Document Versioning

An approved DiscoveredSource becomes a real, fetchable harvester.Source row — reused verbatim from PR10/11, never a parallel registry. Re-ingesting an unchanged document (same content hash) resolves to the SAME SourceDocument row — a genuine no-op; a changed document creates a new, dated row and the old one is preserved, never overwritten. Rejecting a discovered source never deletes its record — historical provenance of what was found and why it was declined is kept.

Refresh Policy — Why Different Sources Check at Different Intervals

Regulatory filings (SEC EDGAR, Companies House) are checked every 90 days; annual/sustainability/ESG/TCFD reports every 180 days; long-horizon transition plans every 270 days — a documented table, never one universal interval. A source whose last attempt failed is retried after 14 days, sooner than its normal interval, without being hammered every cycle. A source with no successful fetch on record yet is always due immediately.

The Refresh Orchestrator

One service, refresh_company_intelligence(), runs discovery → registration → fetch → dedup → KPI candidate matching → recomputation for one company, wrapped in a single AI Observatory session. It is idempotent (reusing the exact same content-hash/dedup-key primitives PR10/11 already proved), failure-tolerant (one source failing never corrupts or blocks another source's real results), and honest about partial completion — a run is COMPLETE only if every checked source succeeded or was a genuine no-op, PARTIAL if some failed, FAILED if none succeeded. Every run's structured outcome (sources checked, documents new/unchanged, evidence created, KPI candidates proposed, warnings, errors) is preserved as its own CompanyRefreshRun history row, linked to (never duplicating) its AI Observatory session.

Dry Run

A dry run performs genuinely zero database writes — it reports which of a company's EXISTING registered sources are currently due per policy, without discovery, registration, or any network fetch. It is a real preview of existing state, never a simulation of hypothetical future discovery results.

No Auto-Confirmation

New evidence from a refresh produces KPI CANDIDATES only, via the same deterministic matcher and the same review_state='proposed' path PR10/11 established — never an automatically confirmed KPI relationship, no matter how strong the keyword match looks. A confirmed relationship still requires a named human reviewer and a written reason in the Evidence Review Workbench.

Data Health — Transparent, Never a Hidden Score

Data completeness is a documented, five-component checklist, each worth exactly 20%: an official source is known; at least one KPI has confirmed evidence; the Shariah screening is current; there is no disputed evidence; no expected document category (annual/sustainability/ESG/TCFD/transition-plan report) is missing. Every component is shown alongside the percentage — never collapsed into an unexplained number, and never described as an investment quality score.

Scheduling

This repo has no periodic task scheduler wired up today (Celery is installed and used for on-demand background work, but no Celery Beat schedule or platform cron exists). The refresh_stewardship_universe management command (--company, --limit, --due-only, --dry-run) is designed to be invoked by an external cron or platform scheduler — EcoIQ does not claim autonomous continuous refresh beyond what is actually wired up.

Security

Every URL accepted into the Source Registry — whether staff-typed or discovered from a lower-trust candidate — is checked against a URL-safety guard before being fetched: only http/https schemes, no private/loopback/link-local/reserved IP ranges, no internal hostname patterns. A URL failing this check is refused, never silently fetched anyway.

What EcoIQ Does Not Claim

This is a data-supply-chain and governance layer, not an investment recommendation engine. It never shows BUY/SELL/HOLD/target-price/expected-return language, never auto-approves a KPI relationship, and never treats a machine-discovered source or a machine-matched candidate as independently verified. Evidence first; human governance before confirmed claims.

Continuous Stewardship Monitoring — Change Detection & Evidence Alerts

Turns the Stewardship Universe refresh pipeline into a change-aware monitoring system: every refresh deterministically detects what genuinely changed — a new document, an unreachable source, a newly proposed KPI candidate, a possible conflict — and surfaces it as an internal evidence alert. Automated monitoring detects changes in evidence and source state. It does not autonomously determine that a company has improved, deteriorated, become investable, or become unsuitable for investment.

How Scheduling Works

The same refresh_stewardship_universe --due-only command PR13 introduced is the intended scheduled entry point — PR14 adds a --scheduled flag so a cron-triggered run is honestly distinguished in the audit trail (CompanyRefreshRun.triggered_by='scheduled') from an ad-hoc staff CLI invocation. render.yaml documents the exact deployable cron configuration, disabled by default — the same "real cost decision, never a silent side effect" discipline PR13 already applied to the Celery worker block. Until that block (or an equivalent external scheduler) is explicitly enabled, EcoIQ does not claim continuous, unattended monitoring — only what was actually triggered, manually or otherwise, is ever shown as having run.

How Due Dates Work

A scheduled/batch-triggered refresh only re-fetches sources refresh_policy says are actually due — never every active source on every tick, which would hammer providers on a frequent cron schedule. A staff-initiated manual refresh (the "Refresh Company Intelligence" button) still rechecks every active source immediately, a deliberate human override, unchanged from PR13.

How Changes Are Detected

A deterministic StewardshipChangeEvent is created only for a REAL, already-happened change — never for an unchanged or skipped fetch. Document-level: NEW_SOURCE, SOURCE_CHANGED, SOURCE_UNREACHABLE, SOURCE_RECOVERED, NEW_DOCUMENT, DOCUMENT_UPDATED. Evidence-level: NEW_EVIDENCE, EVIDENCE_CHANGED, NEW_KPI_CANDIDATE. Every event records real provenance by reference (the source/document/evidence/KPI-link row it concerns), never a duplicated copy of that row's own data.

Potential Conflict — Never a Confirmed Conflict

Two conservative, explainable signals — never naive keyword-opposite matching treated as authoritative — can flag a newly-proposed KPI candidate as a POTENTIAL_CONFLICT: the new evidence text itself uses an explicit discontinuation/reversal phrase, or it comes from a strictly newer document discussing the same KPI without repeating a currently-confirmed supporting claim. Either way, this system only ever creates POTENTIAL_CONFLICT — only a human reviewer, in the existing Evidence Review Workbench, can confirm an actual CONFLICTS relationship. The distinction is deliberate and load-bearing: match validity (does the evidence relate to this KPI at all) stays separate from evidence relationship (what a human decides it concludes).

How Stale Evidence Is Treated, and How Historical Evidence Is Preserved

Evidence is labelled CURRENT, HISTORICAL, POSSIBLY_SUPERSEDED, DISPUTED, or STALE — computed live, never stored. An older document is never deleted or rewritten when a newer one arrives at the same source; its evidence remains inspectable, simply no longer the latest word, with a newer version linked alongside it. STALE means the evidence's own freshness score has decayed below a documented threshold — a genuine "this may need a fresh look" signal, never a silent deletion or an automatic downgrade to false.

Source Recovery

A source that was previously unreachable and becomes reachable again produces a real SOURCE_RECOVERED event the moment its next fetch succeeds (even an unchanged one) — a company is never left permanently marked degraded once its sources are actually healthy again.

Scheduler Safety

Overlapping refreshes for the same company are prevented by a genuine, backend-portable atomic compare-and-swap (an UPDATE ... WHERE tracking_status != 'refresh_in_progress' — correct on every database backend Django supports, unlike SELECT ... FOR UPDATE, which SQLite cannot honour) — a second concurrent trigger for the same company is refused outright, never silently queued, retried, or double-run.

How Alerts Are Prioritised

Every StewardshipAlert's priority is a plain, documented integer sum of real components — severity weight, change-type weight, source authority, whether human review is required — all stored and shown alongside the total. There is no hidden or tuned "AI priority score" anywhere in this system.

How Human Review Remains Authoritative

Nothing in the monitoring pipeline ever mutates a CompanyKPIEvidenceLink's review state or relationship — it only ever creates change events, alerts, and provenance tags (which refresh run proposed a candidate). Confirming, rejecting, or resolving a conflict remains exclusively the Evidence Review Workbench's job, unchanged from PR12.

Global Stewardship Universe — Real Company Expansion & Evidence-Based Discovery

Answers two separate research questions about a credible, expanding universe of real, publicly listed companies: "Which real companies currently have the strongest verified evidence of alignment with the stewardship principles represented by EcoIQ's 114 KPIs?" and, independently, "Which of those companies currently pass or conditionally pass EcoIQ's documented preliminary Shariah screen?" EcoIQ ranks research evidence, not expected investment performance. A company's presence or position in the Stewardship Universe is not a recommendation to buy, sell, or hold any security. EcoIQ never provides BUY/SELL/HOLD ratings, target prices, or expected-return estimates of any kind.

Real Company Identity

Every company in the Global Stewardship Universe has a defensible, inspectable identity path — a SEC EDGAR CIK, a UK Companies House registration number, or (where genuinely known) a verified official domain. These identifiers previously existed only as internal Python mappings; a CompanyListing row now makes them queryable per company (legal name, ticker/exchange where known, country, sector, regulatory identifier, official domain and its verification status). No company is added to the universe without at least one of these real identifiers — this is never a search-engine-discovered or fabricated list.

Two Strictly Separate Lenses

Unchanged from every earlier PR in this series: Shariah eligibility (PASS / CONDITIONAL / FAIL / INSUFFICIENT DATA / NOT SCREENED) and 114-KPI stewardship evidence (CONFIRMED SUPPORTS / PROPOSED / NEEDS MORE EVIDENCE / CONFLICTS / DISPUTED / STALE / HISTORICAL) are never blended into one score. A company can pass Shariah screening with weak stewardship evidence, or have strong stewardship evidence and fail Shariah screening — both facts are always shown side by side, never collapsed into a single verdict.

The Coverage Matrix — Never Collapsing Missing Data to Zero

Each company's data completeness is shown as an honest, five-value matrix across eight dimensions (identity, regulatory data, financial data, sustainability documents, KPI evidence, human-reviewed evidence, Shariah screening, monitoring status): AVAILABLE, PARTIAL, MISSING, STALE, or NOT APPLICABLE. A dimension with no data on file is always shown as MISSING — never silently treated as a zero, a pass, or a fail.

The Evidence-Backed Stewardship Alignment Indicator

Extends the same weighted-component ranking mechanism Company Discovery already used (PR 11) — this is deliberately NOT a second, parallel scoring engine. Every component is a real, computed number; every component is shown alongside the composite, never hidden behind it; a company missing a given component has that component EXCLUDED from the weighted average (renormalised over what is actually available), never coerced to zero:

• kpi_alignment (0.35) — net confirmed-support vs. confirmed-conflict balance across the selected (or all 114) KPIs. Absence of evidence is neutral, never negative.
• source_authority (0.15) — average source tier backing this company's confirmed evidence (regulatory highest, self-reported lowest).
• recency (0.10) — how current the underlying evidence is.
• corroboration (0.10) — how many independent sources support the same claim.
• data_completeness (0.10) — the latest Shariah screen's own honest completeness percentage, where a screen exists.
• human_review_quality (0.10) — the fraction of this company's linked evidence that has actually been human-reviewed and confirmed, rather than merely machine-proposed.
• evidence_diversity (0.10) — how many of the 10 canonical stewardship categories this company has at least one confirmed supporting KPI in, rewarding evidence BREADTH over a narrow pile of claims in a single category.

This composite is always labelled the "Evidence-Backed Stewardship Alignment Indicator" — never an "ESG score" and never implying better financial returns, lower risk, better management quality, or guaranteed sustainability outcomes. Weights are configurable and always shown alongside the ranking, never hidden.

Evidence Quality Differentiation & Greenwashing Resistance

Confirmed evidence is classified as regulatory (e.g. SEC/Companies House filings), independent (e.g. third-party ratings, NGOs, press), or self-reported (the company's own annual/sustainability reports, investor relations pages, and press releases) — a company with many self-published statements does not automatically outrank a company with fewer but independently-verified claims, since human_review_quality and source_authority weight verified, reviewed evidence over raw volume. Two conservative, honestly-labelled signals surface concentration risk without overclaiming: an "Evidence concentration warning" when an unusually large share of a company's confirmed KPI links all trace back to the same single document, and a "High reliance on company self-reporting" note when all of a company's confirmed evidence is self-reported with no independent or regulatory source yet confirmed. EcoIQ does not claim to "detect greenwashing" — these are transparency signals for a human reader to weigh, not an automated accusation.

Discovery, Strongest Alignment, and Comparison

Discover Companies adds filters for minimum confirmed-KPI coverage, minimum human-reviewed evidence share, monitoring health, and the presence of disputes or open potential conflicts — every filter is an explicit opt-in, never a hidden ranking factor. The separate Strongest Alignment view applies the same ranking, restricted to companies with at least one confirmed KPI (an entirely unevidenced company is excluded from this specific view, since it would otherwise appear with a misleadingly neutral composite) — and is labelled, in the product itself: "Research ranking based on currently available and reviewed stewardship evidence. This is not investment advice or a prediction of financial performance." Company Comparison adds the same real signals side by side (confirmed KPIs, human-reviewed coverage, evidence breadth, disputes, open potential conflicts, stale evidence, monitoring health, last refresh) — never expected returns or price targets.

Scaling Safely

Universe expansion is bounded, not unlimited scraping: a per-domain minimum request interval, a per-run cap on how many sources of a given type (e.g. SEC EDGAR, Companies House) get fetched, and a default batch-size bound on how many companies a single run processes unless a larger --limit is explicitly given. Any bound that drops companies from a run reports exactly how many were dropped — never a silent truncation presented as complete coverage. The expand_stewardship_universe management command seeds a CompanyProfile and syncs real identity fields for every company with a known SEC EDGAR CIK or Companies House number; it never marks a company ACTIVE by itself — that status change happens only as the honest side effect of that company's first real, successful refresh.