source-credibility-model

name: source-credibility-model description: Source credibility model for Caro's Ledge. Defines the six-element credibility system (type-based tier, bias tags, citation-network credibility, source discovery loop, operator override, recency decay), the bias tag vocabulary, the customer-facing signal sets per surface, and operational criteria for credibility-affecting work. Load on any dispatch that touches sources table, source_citations or brief-to-source edge tables, tier/base_tier/effective_tier/tier_override columns, bias_tag columns, the candidate review surface, Haiku classification endpoints, customer-facing credibility signal rendering, the discovery loop, or citation network scoring + decay + override semantics. Extends environmental-policy-and-innovation's 6-level Source Type Hierarchy. Customer-facing signal sets per surface align with caros-ledge-platform-intent's five-surface model. when_to_load: - "Touches the sources table (read or write)" - "Touches source_citations or intelligence_item_citations edge tables" - "Modifies tier, base_tier, effective_tier, tier_override, override_reason, override_date, or bias_tag* columns on sources" - "Touches the candidate review surface (canonical_source_candidates table, admin canonical-sources review components, /api/admin/canonical-sources/* routes)" - "Modifies the Haiku recommend-classification endpoints (canonical-sources/recommend-classification, sources/recommend-classification)" - "Modifies the verification pipeline (src/lib/sources/verification.ts)" - "Adds or modifies customer-facing credibility signal rendering on any of the seven surfaces (Regulations, Research, Market Intel, Operations, Community, Map, Intelligence Assistant)" - "Changes the discovery loop (citation extraction in src/app/api/agent/run/route.ts, source resolution in any consumer, candidate promotion criteria)" - "Adds or modifies citation network scoring (src/lib/trust.ts), recency decay, or override semantics"

Source Credibility Model

Section 1: Purpose and Scope

This skill defines the source credibility model for Caro's Ledge. It is the canonical source of truth for how source credibility is computed, surfaced, and acted on across the platform.

Owned by this skill:

• The six-element credibility model (type-based tier, bias tags, citation-network credibility, source discovery loop, operator override, recency decay) and how the elements compose
• The bias tag vocabulary (three dimensions, multi-value within each)
• The customer-facing credibility signal sets per platform surface
• The operational criteria for tier classification, candidate promotion, override, and decay
• The discovery-loop autonomy thresholds and audit-trail expectations on override

NOT owned by this skill:

• Brief content rules (owned by environmental-policy-and-innovation)
• Platform surface intent (owned by caros-ledge-platform-intent)
• Dispatch loop closure, OBS handling, DP compliance, sweep discipline (owned by sprint-followups-discipline)

Cross-skill load is additive, not exclusive. A Build 8 (Research) dispatch loads all four: env-policy (brief content), platform-intent (Research surface intent), sprint-followups-discipline (loop closure), source-credibility-model (the Q9 Research signal set, candidate review wiring, effective tier computation).

Section 2: The Six-Element Model as a Coherent System

The source credibility model has six elements that together produce a single credibility signal (effective tier) that customers and agents see by default.

1. Type-based default tier. Every source has a static base_tier (INT 1-7) assigned at classification time based on the source's institutional type per the environmental-policy-and-innovation 6-level Source Type Hierarchy (tier 7 = overflow/uncategorized). Base tier is set once at classification and rarely changes; it preserves the classifier's or operator's original credibility judgment for audit and provenance.

2. Bias tags. Sources carry orthogonal bias tags across three dimensions (Funding/Institutional Affiliation, Methodological Orientation, Stakeholder Position). Bias is independent of tier; a tier-2 regulator can have any bias profile, just as a tier-6 analytical outlet can. Bias is a separate axis from authority weight. Bias tags apply to external publisher sources only.

3. Citation-network credibility. Sources accumulate credibility weight from incoming citations. Citation weight is tier-weighted (citations from higher-tier sources count more) and decayed (recent citations count more than old ones). The aggregated weighted citation sum contributes to a source's effective tier.

4. Source discovery through citations. When briefs cite sources not in the registry, the discovery loop captures them as candidates. High-confidence candidates surface to the operator review queue; low-confidence ones accumulate in provisional storage without consuming operator attention until they cross promotion thresholds.

5. Operator override. The operator can explicitly set tier_override on a source with override_reason and override_date. The override takes precedence over computed values in the effective tier formula. Overrides are auditable via the existing source_trust_events table.

6. Recency decay. Citation weight decays with a half-life curve (18-24 months tunable). Recent citations contribute more than old ones. Decay applies to citation-network contributions only; it does NOT apply to base_tier (structural, not time-sensitive) or to accessibility decay (separate existing logic in src/lib/trust.ts).

System diagram

base_tier (from classification, INT 1-7, static) ────────────┐
                                                              │
citation network (tier-weighted, decayed) ────────────────────┤
                                                              ├──> effective_tier ──> customer-facing signal
operator override (when present)                              │                        (Section 8)
  COALESCE(tier_override, computed_dynamic, base_tier) ───────┘

Formula: effective_tier = COALESCE(tier_override, computed_dynamic_tier, base_tier)

Where computed_dynamic_tier is derived from base_tier plus the tier-weighted decayed citation network sum, computed by daily batch recompute.

Why six elements together, not just static tier. Each element addresses a failure mode the others cannot:

• Type-based tier alone is static; a tier-6 analytical outlet whose work is consistently cited by tier-1 regulators is doing tier-3-quality work, but static tier misses that signal
• Bias tags alone don't address authority; a methodologically-transparent industry-funded source can still be authoritative within its scope, but bias warns the consumer about lens
• Citation network alone overweights popular sources without considering institutional type; high citation count from low-tier sources should not elevate a source's credibility
• Discovery alone produces noise without verification gates; high-confidence routing prevents queue flood
• Override alone is human-only; without computed signals, every classification decision is operator labor
• Decay alone is time-only; without network and base signals, decay has nothing to operate on

Section 3: Type Tier Operational Criteria

The base_tier 1-7 mapping extends the env-policy 6-level Source Type Hierarchy with operational decision criteria:

Tier 7 weight in citation-network scoring

T7 weight = 0. A source classified as T7 contributes no credibility signal when it cites other sources, because T7 is the platform's signal that the source's authority has not been established. Promoting other sources based on T7 citations would amplify uncertainty.

This differs from T6 (weight = 0.15) where the source has a known classification (analytical/opinion) and contributes some signal because its type is known even if its tier is low.

Sub-vertical inheritance

When a sub-vertical regulator (e.g. IMO for maritime, ICAO for aviation, FMCSA for US trucking) acts within its vertical, its tier defaults to the parent vertical's regulator tier (typically T2). When a sub-vertical body acts outside its mandate (e.g. an industry consortium publishing analytical commentary), the work is classified by the act's nature, not the institution's parent type.

The classifier prompt should recognize both the institution AND the work product to assign correctly.

Open sub-decision: per-jurisdiction threshold differentials. Today the classifier uses global thresholds (75/55 relevance/freight). Whether EU vs US vs APAC need different calibration is operator-domain-knowledge work; not yet decided. Tracked in the decisions doc Open Sub-Decisions section.

Tier as referenced in classifier prompts vs customer-facing surfaces

Classifier prompts continue to reference "tier 1-7" as the static type taxonomy. They classify base_tier. Customer-facing surfaces and agents reference effective_tier (the dynamic signal that incorporates network + override + decay). The two are distinct schema fields after Q2 lands; consumer migration is per-consumer (decisions doc Open Sub-Decision).

Per-item-type authority floor (provenance gate)

The authority floor is the minimum grounding tier a CRITICAL/HIGH item's FACT claims may carry, and it is calibrated per item type — a single tier floor across all types is a category error (the regulatory floor would false-quarantine 76–88% of correctly-sourced market/research facts, which legitimately ground at T3–6). A CRITICAL/HIGH item's FACT claims MUST carry source_tier_at_grounding at or above the floor for their item type, enforced by validate_item_provenance (migration 141): regulation-family ≤ T2, research_finding ≤ T4 (preserves the T3 research core — IPCC/OECD/IEA/World Bank — with one tier of industry-analysis margin; T5–6 are relabeled to ANALYSIS by the mechanical prose-safe discipline), technology/innovation/tool ≤ T5 (forward default, calibrated against zero live items, REVISIT when the first land). A FACT below its type's floor either re-grounds to a higher-tier source or relabels to a labeled ANALYSIS claim; it is never forced.

Named exemptions (REVISIT, so neither silently becomes permanent). market_signal and initiative are EXEMPT from a tier floor: their strength is corroboration-count (Section 4 — N independent tier-weighted corroborators), not a single-source tier, and codifying the corroboration gate before it is built would put unbuilt mechanism in the gate. regional_data is EXEMPT: its floor is per-SECTION (feasibility facts ≤ T3; cost-data facts any tier with a registered source), not item-level. Both stay exempt until their gates are built; the absence is registered (invariant SC-8), not silent.

Canonical institutional tier (one tier per institution)

Tier is a property of the institution, not of a row. The registry MUST hold one canonical institutional tier per host group — multiple sources rows for the same institution (different pages, ingestion paths, dates) MUST share one base_tier. The grouping key is the registrable domain (eTLD+1) = the institution by default, with documented super-domain exceptions where a shared government domain hosts distinct bodies: europa.eu subdomains are institution-distinct (eur-lex.europa.eu = T1 legal text vs ec.europa.eu = T2 regulator vs eea.europa.eu = T3 agency), and legislation.gov.uk (T1) is distinct from gov.uk departmental (T2). A single deliberate per-row exception is allowed ONLY via the Section 7 override columns (tier_override + override_reason, default none); absent an override, inconsistent per-row tiers are the duplicate-row defect and MUST be canonicalized. Bulk registration MUST classify (never default a real institution to T7 overflow); the register-as-source path assigns the institution's honest tier at registration.

Per-claim attribution = the source containing the span. A FACT claim is attributed to the source whose fetched content actually contains its verbatim span (resolved via the span's pool row), NOT a hardcoded primary. Its grounding-tier stamp is set so that the stamp equals the canonical institutional tier of the source containing the span (the flagged-override row tier where present; NULL when that host is unregistered). No constant stamps: a constant masquerading as a resolved tier is fake certification. (Durable form: a single institutions table with sources referencing it — tracked follow-on; until then the per-host canonical base_tier carries the institutional tier.)

Section 4: Citation Network Semantics

Edge tables

Two edge tables hold citation relationships:

• source_citations (existing per migration 004): source-to-source edges. When source A cites source B, a row is inserted with source_id (citing), cited_source_id, detected_at, context. Used for tier-weighted credibility scoring.

• intelligence_item_citations (NEW per Q1, schema TBD): brief-to-source edges. When intelligence_item X cites source Y, a row is inserted. Backfilled from intelligence_items.sources_used (UUID[]) for INPUT-source associations at launch. Discovered-source associations (from agent "New Sources Identified" markdown tables) fill organically going forward. Historical markdown parsing is out of scope per decisions doc Flag 3.

Tier-weighted citation sum

weighted_sum(source_id) = SUM(
  tier_weight(citing_source.effective_tier) * decay_factor(detected_at)
) FOR each row in source_citations WHERE cited_source_id = source_id

Tier weights (verbatim from Q7):

• T1 = 1.0
• T2 = 0.85
• T3 = 0.7
• T4 = 0.5
• T5 = 0.3
• T6 = 0.15
• T7 = 0

Independence and syndication (corroboration integrity)

Corroboration STRENGTH counts INDEPENDENT corroborators, not raw citation edges. When N outlets republish a single underlying announcement (one press release, one wire story), that is ONE corroboration, not N — syndication is collapsed to its origin before counting. Two integrity rules follow, both load-bearing for any "how well-corroborated is this signal?" judgment (the Market surface's corroboration-count grounding model in particular):

1. Independent over raw. Count distinct INDEPENDENT corroborators (citers sharing a syndication group collapse to one), never the raw edge count. Raw count rewards a single announcement echoed across an aggregator network, inflating credibility from one source.
2. Tier over volume. A corroborator's weight is its institutional tier (T1 best … T7 = 0), not how often it appears. Ten low-tier echoes do not outweigh one T1/T2 independent confirmation; high citation volume from low-tier sources must not elevate credibility (the Section 1 principle).

Implementation: src/lib/sources/source-growth.ts (aggregateConvergence) collapses each syndication group to one unit at its best tier, then reports independent_citers / confirmation_count / highest_citing_tier; the trust citation component (src/lib/trust.ts) consumes those, not raw edge counts. Distinguish the two measures: the tier-weighted decayed SUM below scores a source's STANDING credibility over time (it can grow from many edges); the independent-citer COUNT scores a single SIGNAL's corroboration strength right now. A signal is "strongly corroborated" only on independent, suitably-tiered confirmation — never on volume alone.

Recency decay

decay_factor(detected_at) is a half-life curve. Starting parameter: 18-24 months tunable; operator decides exact value (open sub-decision in the decisions doc). Formula:

decay_factor(detected_at) = 0.5 ^ ((NOW - detected_at) / half_life)

A citation at now contributes weight 1.0. A citation at now - half_life contributes weight 0.5. A citation at now - 2 * half_life contributes weight 0.25.

Decay applies to citation-network contributions to effective_tier ONLY. Decay does NOT apply to:

• base_tier (structural, not time-sensitive; a 2010 EU regulation is still binding at tier 1 in 2026)
• accessibility decay (separate existing logic in src/lib/trust.ts operating on last_accessible timestamps)
• tier_history audit trail (immutable record)

Promotion threshold

Per Q7: a source is candidate for tier elevation when weighted_sum >= 2.5. The threshold accounts for tier-weighted contributions (a single T1 citation = 1.0, so the threshold requires effectively 2-3 high-tier citations or many more low-tier citations).

Recompute cadence

Daily batch. The recompute job:

1. For each source, sums weighted decayed citation contributions
2. Computes computed_dynamic_tier as a function of base_tier plus the network signal
3. Writes effective_tier = COALESCE(tier_override, computed_dynamic_tier, base_tier) to the sources row
4. Logs to source_trust_events when effective_tier changes

The source_citations.detected_at column (existing per migration 004, currently unused in scoring) wires into the decay computation.

Section 5: Source Discovery Loop

Citation extraction at ingest

The agent at brief generation emits a "New Sources Identified" markdown table per the system prompt contract at src/lib/agent/system-prompt.ts:358-368. Extraction logic at src/app/api/agent/run/route.ts:479-580 parses the table and routes each citation:

• If the URL matches an existing sources row: write a row to source_citations (source-to-source edge); also write to intelligence_item_citations once Q1 lands.
• If the URL doesn't match: write to provisional_sources with the agent's tier estimate, incrementing citation_count and updating citing_source_ids on repeated citations.

URL canonicalization at resolution

Per Q10 (migration 087, helper at src/lib/sources/url-canonicalize.ts), all URL-based source resolution goes through canonicalizeUrl(). This lowercases scheme and host, strips www prefix, trims trailing slash, sorts query params, strips fragments. Resolution sites are documented in the Q10 dispatch report; future resolution code MUST use the helper.

Candidate surfacing threshold

Per Q7: candidates with classifier confidence > 0.65 surface to the operator review queue (canonical_source_candidates table; admin UI at src/components/sources/CanonicalSourceReview.tsx). Below-threshold candidates accumulate in provisional_sources without surfacing.

Citation-frequency promotion threshold

Per Q7: a candidate with 3+ citations and weighted sum >= 2.5 promotes to operator review regardless of classifier confidence. This catches sources the platform learns about through repeated citation even when individual classification confidence is moderate.

Tier-opinion preservation

Per Q3: when the agent's brief generation cites an EXISTING source AND has a tier estimate for it that differs from the source's current base_tier, record the opinion as evidence. Schema TBD: either a new tier_opinions table (citing_source_id, target_source_id, opined_tier, opinion_date, opinion_source) or an extension to source_citations (add opined_tier column).

Aggregated disagreement flag: when 5+ opinions disagree with the database tier on the same source within a 90-day window, surface to operator review as a tier reconsideration prompt.

This makes the discovery loop self-improving on existing sources, not just on new ones.

Expected operator review queue size

Per Q7: 5-15 new candidates per week. Thresholds are tunable starting points; adjust based on observed queue size and review experience.

Section 6: Bias Tag Vocabulary

Three dimensions, multi-value within each. Standard is accuracy and transparency at all times.

Dimension 1: Funding/Institutional Affiliation

• industry-funded
• government-funded
• foundation-funded
• subscription-supported
• academic-institutional
• mixed-funded
• funding-opaque

Dimension 2: Methodological Orientation

• peer-reviewed
• methodologically-transparent
• analytical-synthesis
• editorial-opinion
• advocacy
• factual-reporting
• standards-defining

Dimension 3: Stakeholder Position

• industry-incumbent
• industry-challenger
• regulator-aligned
• environmental-advocate
• independent-research
• customer-perspective
• labor-perspective
• investor-perspective

Assignment

The Haiku recommend-classification prompts propose bias tags as part of classification. Operator confirms on candidate review. Multi-tag per dimension is allowed and expected; most sources carry multiple tags within at least one dimension.

Scope

Bias tags apply to external publisher sources ONLY. User-generated Community content (a freight forwarder posting a market observation, a peer response in a community thread) uses author-identity + workspace-verification credibility signals (Section 8 Community row), NOT bias tags. The sources registry concept does not apply to user-generated content.

A member sharing a FreightWaves article on Community: the source (FreightWaves) carries its bias tags as a sources-registry record; the member's act of sharing carries author-identity signals as a Community-content record. Two different credibility models apply at the two layers.

Worked example

ICCT (International Council on Clean Transportation):

• Dimension 1: foundation-funded
• Dimension 2: methodologically-transparent, analytical-synthesis
• Dimension 3: independent-research, environmental-advocate

A single source carries multiple tags within each dimension where they apply. The combination is informative: foundation-funded + methodologically-transparent + environmental-advocate signals "rigorous third-party research with explicit advocacy framing." Customers reading an ICCT-cited brief can calibrate accordingly.

Section 7: Override Mechanism Rules

Columns

Three columns on sources per Q5:

• tier_override INT NULL CHECK (tier_override BETWEEN 1 AND 7)
• override_reason TEXT NULL
• override_date TIMESTAMPTZ NULL

Effective tier formula

effective_tier = COALESCE(tier_override, computed_dynamic_tier, base_tier)

Override takes precedence when present. Falls back to computed dynamic tier from network signals. Falls back to static base_tier when no network signal available (typical for newly-registered sources before citation accumulation).

When to override

• Classifier-vs-evidence disagreement: classifier scored T4 but operator domain knowledge says T3 is correct given the institutional context
• Network signal misleading: rare case where citation network erroneously elevates a low-quality source (or fails to elevate a high-quality one) and operator judgment supersedes
• Bridging classification gaps: sources that fit T7 (overflow) until a better classification path exists may carry an override to a more accurate tier

How to override

Explicit POST to the override endpoint (route TBD, likely /api/admin/sources/[id]/tier-override). Mandatory override_reason field; the audit trail must record why. Audit log to source_trust_events (existing table) with event_type = 'tier_override', created_by = 'human', payload including before/after tier and reason.

Override does NOT modify base_tier. Base tier preserves the classifier's original judgment for provenance.

Revert

Clear tier_override, override_reason, override_date to NULL. effective_tier resumes computation from network + base. A revert is itself a source_trust_events row with event_type = 'tier_override_revert'.

Section 8: Customer-Facing Signal Sets Per Surface

Per Q9. Per-surface signal sets with consistent vocabulary across surfaces. Asymmetry is deliberate per audience need.

Canonical domain INT-to-surface mapping (1-7): see fsi-app/src/lib/domains.ts for the single source of truth consumed by both the Haiku classifier (via domainForItemType) and the surface filters on /regulations, /market, /operations, /research, and the Dashboard surface-coverage widget. Domain integers MUST NOT be hardcoded outside that file; use the named exports (REGULATIONS_DOMAIN, RESEARCH_DOMAIN, etc.) instead. The migration 101 routing rule is the authoritative branch logic and is mirrored verbatim in domainForItemType.

Per-surface implementation is build-dispatch scope

This section specifies WHAT signals each surface foregrounds. Build dispatches specify HOW (visual treatment, badge style, expand-on-click panel structure, color coding). Per-surface implementation lands per Tier 4 build:

• Build 7 Market Intel: tier + recency + signal-strength
• Build 8 Research: tier + bias tag + citation count + recency
• Build 9 Operations: tier + jurisdiction + applicability
• Build 10 Community: author identity + workspace verification (separate model, see below)
• Build 11 Dashboard: aggregates across surfaces

Vocabulary consistency

Tier badge means the same everywhere (T1-T7 with the same labels and colors). Jurisdiction renders the same way wherever it appears. Bias tags render consistently when present. Customers learn the signal vocabulary once and recognize it across surfaces.

Community surface uses a different model

The Community surface renders user-generated content. The bias tag vocabulary applies to external publisher sources only and does not apply to Community content. Community credibility uses author-identity-shaped signals:

• Author identity (verified member, organization affiliation)
• Workspace verification (member is associated with a verified workspace)
• Posting history (operator-tunable signal weight)
• Future moderation signals (flagged content, peer endorsement, etc.)

When a member shares a FreightWaves article on Community, the FreightWaves source carries its bias tags (separate sources-registry signal); the member's act of sharing carries author-identity signals (Community model). Both render on the Community surface side by side.

Section 9: Anti-Patterns

These behaviors mean the model was not understood or was deliberately ignored:

• Expanding category vocabulary to capture distinctions that belong in source_role. A common temptation: "we need an 'analytical_press' category for the 8 analytical-press sources." Operator rejected this in the analytical-press dispatch (2026-05-19). Press content goes in category='research' with source_role='trade_press' and tier 5-6. Categories are surface routing; roles and tiers are credibility differentiation within a surface.

• Exact-URL matching at source resolution. Source registry lookups must canonicalize URLs first via src/lib/sources/url-canonicalize.ts. Direct .eq("url", rawUrl) calls produce silent duplicates from formatting drift (trailing slashes, www, query-param ordering). Q10 fixed the existing sites; new resolution code must use the helper.

• Discarding tier-opinions when source already exists. When Haiku estimates a tier for a cited source and the source already exists with a different tier, do not discard the opinion. Record it per Q3. Aggregated disagreement at 5+ within 90 days surfaces for operator review.

• Conflating role with bias. source_role (institution type: trade_press, academic_research, regulator) is orthogonal to bias (perspective: industry-incumbent, environmental-advocate, peer-reviewed). A source_role='academic_research' source can carry any bias profile across the three bias dimensions. Treating role values as bias values (e.g., assuming academic_research implies peer-reviewed) loses signal and produces wrong customer-facing presentation.

• Treating tier as fully static when network signals exist. After Q2 schema lands, the dynamic computed effective_tier IS the credibility signal customers and agents should consume. Reading base_tier directly when the consumer wants dynamic credibility silently shows stale signal.

• Adding bias tags to user-generated Community content. The sources registry concept does not apply to user-generated content. Community uses a different credibility model (author identity + workspace verification + posting history). Conflating the two models produces wrong UI and wrong audit semantics.

• Skipping URL canonicalization on new source resolution code. When adding a new source lookup, comparison, or write, use canonicalizeUrl() from the helper. Skipping it recreates the same silent-duplicate failure mode Q10 fixed.

• Reading sources.tier directly when the consumer wants the dynamic value. Per Q2 the schema will have both base_tier (static) and effective_tier (dynamic). Most consumers want effective_tier. A few definitional consumers (the SourceTier type definition, classifier prompts that reference the tier 1-7 type taxonomy) want base_tier. Per-consumer review at the migration point per the decisions doc Open Sub-Decision.

• Auto-merging duplicate sources without operator decision. Q10 surfaced 9 duplicate sets in sources plus 29 cross-table collisions. The dispatch policy was surface, do not merge. Merging requires operator judgment for sets like BREEAM vs BRE Group where the canonical name is non-obvious.

Section 10: Cross-References

• environmental-policy-and-innovation: defines the 6-level Source Type Hierarchy this skill extends in Section 3. Integrity rule applies to bias tag assignment (no invented tags; assignment grounded in source evidence). Load alongside on intelligence_items work.

• caros-ledge-platform-intent: defines the five-surface canonical model that drives the per-surface signal sets in Section 8. Intelligence Assistant is the cross-cutting capability per platform-intent Section 4 that surfaces credibility per the Assistant signal set (inline citations + full provenance via the CitationPanel component in src/components/AskAssistant.tsx).

• sprint-followups-discipline: load-trigger rule (fifth named binding rule, added 2026-05-19 alongside this skill) names which dispatches must load source-credibility-model. Sources-schema-touch precondition applies to dispatches that touch the sources table. The Sweep-discipline rule applies to credibility-model audits and reviews.

• docs/design-principles.md: DP-1 single-pane operator review binds on operator-facing credibility surfaces, particularly the candidate review queue (canonical_source_candidates review at src/components/sources/CanonicalSourceReview.tsx) and the override mechanism UI (TBD).

• docs/sprint-2/source-credibility-model-decisions-2026-05-19.md: source-of-truth document for all 10 architectural decisions encoded in this skill. Captured at commit 2e9175a (merged to master in commit 4943e83). Open sub-decisions tracked there: per-consumer base_tier vs effective_tier migration calls (Q2), exact half-life value (Q6), per-surface implementation specifics during Tier 4 builds (Q9), bias vocabulary iteration if needed (Q4), T7 weight = 0 confirmation (Q7).