The bar

Sourced. Derived.
Or absent. No fourth state.

Every numeric and factual claim Afrimintel publishes lives in exactly one of three categories. We replaced "100% accurate" — which is unattainable — with an operational framework: source it, derive it from sourced inputs, or admit we don't have it. Errors are inevitable. Discipline about errors is what makes a record trustworthy.

v1.0 · Adopted 25 April 2026 SLA: 7 business days material 37 pre-deploy checks block drift

The three-state model

Every numeric or factual claim Afrimintel surfaces is in one of three states:

SOURCED

Sourced. Claim derives from a named, citable, primary or institutional document with a date. The source is recorded in the platform's data layer and surfaced to the user. Examples: a deposit's reserve tonnage cited from an NI 43-101 / JORC / S-K 1300 technical report; a country's Fraser Institute IAI score; a transaction value cited from a primary press release or SEC filing.

DERIVED

Derived. Claim is the output of an Afrimintel calculation or composite whose inputs are themselves Sourced or Derived, and whose methodology is published. Output carries explicit confidence range or tier. Examples: the Afrimintel Score for a province (formula and inputs at /methodology/); the country risk composite (30/25/25/20 weighting, published). The simplest Derived class is contained metal — reserve tonnage × grade — which Afrimintel computes rather than quotes, even where a source also states it; the underlying tonnage-and-grade reserve statement remains Sourced, so a producing-asset reserve row carries both pills (tonnage @ grade Sourced; contained Derived).

ABSENT

Absent. Afrimintel does not hold the data and acknowledges the gap directly. Where a user asks for a fact Afrimintel does not hold, the answer is "we do not hold this" — not a guess, not a placeholder, not a training-data approximation. AI surfaces (e.g. ASK Afrimintel) carry the same discipline: the system prompt instructs the model to refuse fabrication and respond [DATABASE ABSENT] when no relevant record exists.

No fourth state. A claim that does not satisfy Sourced, Derived, or Absent does not appear on the platform. There is no category for "we'll figure this out later" or "approximately right" without a tier tag. If we have not done the work to put a claim in one of the three states, we do not publish the claim.

Editorial Judgment — qualitative claims

The three-state model governs numeric and factual claims. A separate discipline applies to qualitative editorial framing — comparative readings ("EU policy continuity through 2030 is more predictable than U.S. policy continuity across administrations"), characterisations of trajectory ("the corridor's effective country-risk profile is dominated by the highest-risk segment"), and decision-shape framings ("this is post-settlement re-entry, not pre-FID screening"). These are not Sourced, not Derived in the strict computational sense, and not Absent — they are editorial judgments.

The discipline that governs them:

Surfaced as judgment, not as fact. Editorial framings are written so a reader can identify them as editorial calls. We do not present comparative readings, decision-shape framings, or trajectory characterisations as if they were sourced facts.
Basis named when load-bearing. When an editorial judgment carries decision-aid weight, the basis is named — the institutional-vs-bilateral structural difference behind a U.S./EU stability comparison; the trilateral-coordination-cost framing behind a corridor-level country-risk reading; the operating-vs-disputed distinction behind a re-entry decision shape.
Subject to the Correction Velocity SLA when challenged. Editorial judgments are not below the correction discipline. A senior reader who flags an editorial framing as incorrect or unsupported gets a response on the same SLA as a numeric correction — material disagreement addressed within 7 business days, non-material within 30. Either the framing is defended with named basis, or it is revised, or it is withdrawn. The Audit Log records the resolution.
Editorial responsibility owns them. Numeric claims pass through the data quality verification role; editorial judgments are owned at the editorial responsibility level. Both are named at the same accountable owner.

This discipline is added in v1.0.45 in response to a hostile-audit observation that the three-state model governs numbers but not the qualitative framing on which decision-aid value depends. Both layers are now published.

Working-state tags: a v1.0.61 honest amendment

A hostile contrarian audit in the v1.0.55–v1.0.61 build window identified that the platform's substrate had begun rendering two additional state tags — Pending and Inferred-propagated — that are not part of the three-state model as published above. This section reconciles the published standard with the substrate as it actually ships, rather than leaving the contradiction unstated.

The three-state model (Sourced / Derived / Absent) remains the governing target for every claim. The two working-state tags are explicitly transitional, not a fourth and fifth permanent state:

Pending. A claim for which the verification work is scheduled but not yet complete. Pending claims carry an explicit target verification date recorded in the platform's sources-verification-log.json. A Pending claim is on a path to becoming Sourced (if the primary source confirms it) or Absent (if it does not). Pending is not a resting state — a claim that remains Pending past its target date is a Correction Velocity SLA breach.
Inferred-propagated. A claim carried forward from a prior build cycle that was either (a) previously tagged Sourced without an independent re-walk, or (b) an editorial estimate originating in an earlier session. Inferred-propagated claims are flagged for primary-source re-verification at a named audit cycle. The tag exists specifically so that the substrate does not over-claim Sourced status for content that has not been independently verified in the current cycle.

Both working-state tags exist because the honest alternative — tagging unverified content as Sourced, or removing it entirely mid-verification — is worse. Tagging unverified content Sourced is fabrication. Removing it loses the framework substrate that is genuinely useful. The working-state tags let a reader see exactly which claims are verified, which are scheduled for verification, and which are carried forward pending re-walk. Every Pending and Inferred-propagated tag is, by design, a visible admission of incomplete verification rather than a hidden one.

The discipline commitment: working-state tags are transitional and dated. The verification log carries the target dates. The audit log records each transition from Pending or Inferred-propagated to Sourced, Derived, or Absent. If the platform's substrate is mostly Pending and Inferred-propagated at any point, that is itself a visible signal — surfaced honestly — that primary-source verification work is the binding constraint, not a hidden weakness.

Field-level provenance

For records classified as INTELLIGENCE-GRADE (currently 18 deposit dossiers in the platform's database), provenance is captured at field level — not just record level. Each Reserve tonnage, Resource grade, operator name, and reporting basis carries its own tier tag and source citation. A user examining Kamoa-Kakula sees not just "Sourced from Ivanhoe NI 43-101" but the specific document, effective date, and per-field provenance.

Schema fields enforcing this discipline (v1.0.18 schema):

res_mt_reserve / grade_reserve / res_type_reserve — Proven + Probable Reserve, separated from Resource
res_mt_resource / grade_resource / res_type_resource — Measured + Indicated + Inferred Resource
reserve_basis / resource_basis — explicit "100% project basis" vs "attributable to [operator]" disclosure
operator_interest — source-qualified ownership detail
dispute_status — enum for contested assets (Manono pattern)
transparency_flag — enum for opaque or recovery-stage operators
last_reviewed — explicit date stamp; pre-deploy pipeline blocks any IG record older than 180 days

Correction-velocity SLA

Material errors corrected within seven business days. Non-material corrections (typos, formatting, version stamps) within thirty business days. Every correction is logged at /audit-log/ with: what changed, why, the source of the authoritative correction, version marker incremented, regression check, editorial sign-off.

The Audit Log is published, not internal. A correction that has not been logged is not complete.

Pre-deploy gate discipline

Every deploy is gated by an automated pre-deploy pipeline that blocks publication on any of the following findings:

Check	Severity
Permanent exclusions present in deployed surfaces	CRITICAL
Hardcoded secrets in deploy bundle	CRITICAL
Wildcard CORS in serverless functions	CRITICAL
localhost in production CORS allowlists	SIGNIFICANT
Counterparty hallucination patterns	CRITICAL
Forward-looking partnership claims without confirmed counterparty	SIGNIFICANT
Methodology reconciliation lock — every province score reconciles to formula within ±0.05	CRITICAL
Score-prose drift — supporting prose disagrees with score chip	SIGNIFICANT
Opportunity Multiplier range consistency across surfaces	SIGNIFICANT
Round-number drift between marketing copy and data	SIGNIFICANT
AI provenance contract clauses present in shared injection helper	CRITICAL
JS syntax integrity across all bundle files	CRITICAL
HTML structural balance	SIGNIFICANT
AI provenance tier tags surfaced in model responses	SIGNIFICANT
IG dossier last_reviewed within 180-day threshold	SIGNIFICANT
Reserve / Resource basis disclosure on attributable records	SIGNIFICANT
Disputed-asset surface — dispute_status flag on multi-claim records	SIGNIFICANT
Shared injection helper present and consumed by both production and audit infrastructure	CRITICAL

A CRITICAL or SIGNIFICANT finding blocks deploy. The pipeline runs at every build; the script is in the bundle at scripts/audit/pre-deploy-audit.js.

AI provenance contract

Afrimintel's AI feature (ASK Afrimintel) is governed by a published Provenance Contract embedded in the system prompt before every query. The contract instructs the model to:

Answer only from the Afrimintel database injected into the system prompt — not training-data inference
Cite the tier (Sourced / Derived / Absent) for every numeric or factual claim
Refuse fabrication of operator names, reserve tonnages, grades, transaction values, or counterparty relationships
Refuse to claim Afrimintel partnerships, customers, or commercial relationships unless explicitly stated in the data
Respond "Afrimintel's current dataset does not include verified information on [topic]" when no relevant record exists

Compliance with the contract is measured by a 20-query test battery (in-database / absent / adversarial / tier-mix categories). A defence-in-depth runtime audit runs after every model response, surfacing inline flags if the response contains counterparty-relationship language, forward-looking partnership claims, or permanent-exclusion terms.

What this Standard does not claim

This is not a guarantee against error. The Standard is a discipline framework, not a promise of perfection. Errors will occur. The commitment is that errors are caught quickly, corrected transparently, and logged publicly. The platform's flagship credibility is not zero-error operation; it is observable error-correction discipline.

Specific things the Standard does not promise:

That every dossier is exhaustively current — operational events outpace dossier refresh; the last_reviewed field shows the floor of staleness, the 180-day threshold the ceiling
That every Sourced citation has been independently re-walked by Afrimintel against the underlying primary document — the Sourced tier requires that a named, dated, citable primary or institutional document exists and is recorded in the platform's data layer; it does not require that Afrimintel has opened the source PDF and verified each cited field. Verification depth varies by record and is highest on the 18 intelligence-grade dossiers (where field-level provenance is documented and quarterly deep audits are scheduled) and lower on the spatial-reference tier (where the citation chain is recorded but per-claim re-walks are not routine). When Afrimintel does perform an independent re-walk, the work is documented in the audit log under a specific record-level audit entry. The hedge text on intelligence-grade dossiers explicitly flags claims that have been cited but not independently re-walked, distinguishing citation hygiene from independent verification
That tier classification is governed by the published tier definitions, with the platform's data layer as the single source of truth; the public methodology page surfaces the intelligence-grade count (18), and the full per-tier inventory is maintained and reconciled internally. The pre-deploy pipeline's CHECK 22 verifies tier-count bookkeeping consistency, not per-record classification accuracy. The judgement that a specific record meets the intelligence-grade criterion (resource statement + named operator + primary-source citation) sits with editorial discipline at ingestion time and with the quarterly deep audit cycle on rotating provinces, not with an automated classification check. Where a deep audit reclassifies a record, the audit log entry records the reclassification
That AI synthesis is decision-grade — AI responses carry inline audit cards but are not a substitute for human review of cited primary sources
That the Standard substitutes for accreditation — it is a published editorial framework, not a peer-reviewed industry standard backed by a governing body

Editorial responsibility

The Standard is signed and operated by a single named editorial responsibility: Nikesh Patel, Honorary Consul of Rwanda in Mauritius (nikesh@afrimintel.com). Distribution of editorial capacity across additional named reviewers is on the build roadmap, sequenced to the point a commercial event requires it. Until then, every record carries one editorial signature, and that signature is accountable.

Versioning and changes to this Standard

Changes to the Standard itself are logged at /audit-log/ under "Standard amendment." Material changes (e.g. addition or removal of a state, change to the correction SLA, change to the AI contract clauses) require explicit Audit Log entry with reasoning and effective date. Non-material editorial revisions (typos, formatting, link maintenance) are versioned but do not require a separate entry.

Adopted 25 April 2026. Last amendment: 2 May 2026 (v1.0.18 schema fields added, pre-deploy gate checks expanded). Companion documents: Methodology · Audit Log · Privacy Policy · Terms of Service.

Sourced. Derived.Or absent. No fourth state.