Live mirror of CHANGELOG.md from the public MCP server repository. Refreshed hourly.

Changelog

All notable changes to HEORAgent MCP Server.

v1.13.0 (2026-05-28) — Feature: AI Transparency Disclosure (ISPOR ELEVATE-GenAI aligned)

Adds a structured AI-assistance disclosure block to tool outputs, aligned with the ISPOR ELEVATE-GenAI reporting guidelines (Fleurence RL et al., Value Health 2025;28(11):1611–1625).

New

• ai_disclosure_level parameter on 16 tools: "off" | "standard" (default for most) | "submission" (default for HTA/regulatory tools). Controls whether and how the disclosure block is appended to tool output.
• buildDisclosure(audit, opts) in src/formatters/disclosure.ts: renders a formatted AI Assistance Disclosure section from the audit record. Derives data sources from the existing sources_queried: SourceAudit[] field — no schema duplication.
• extractDisclosureLevel(args, default): safe extraction from raw args before Zod parsing, allowing per-call override without modifying 30 Zod schemas.
• AI_DISCLOSURE_LEVEL_SCHEMA_PROPERTY: reusable JSON Schema fragment published in inputSchema.properties for all 16 wired tools.
• addToolCall(record, trace) in src/audit/builder.ts: appends a ToolCallTrace to audit.tools_called (immutable append).
• ToolCallTrace interface in src/audit/types.ts: { name, ms, outcome, output_size_bytes? }.
• Persona-driven defaults in web/lib/systemPrompt.ts: payer/HTA-reviewer personas default to "submission"; analyst personas default to "standard"; scratchpad intent → "off".
• 3 new homepage example cards in Chat.tsx demonstrating submission-ready disclosure, payer dossier full disclosure, and disclosure-off scratchpad workflows.

Wiring by tool tier

ISPOR citation

Fleurence RL, Dawoud D, Bian J, Higashi MK, Wang X, Xu H, Chhatwal J, Ayer T; ISPOR Working Group on Generative AI. ELEVATE-GenAI: Reporting Guidelines for the Use of Large Language Models in Health Economics and Outcomes Research: An ISPOR Working Group Report. Value Health. 2025;28(11):1611–1625. doi:10.1016/j.jval.2025.06.018

v1.11.3 (2026-05-22) — Fix: expose run_owsa and study_types in MCP schemas

run_owsa was accepted by the models.cost_effectiveness handler (if (params.run_owsa !== false)) but absent from the JSON inputSchema — MCP clients couldn't discover or disable one-way sensitivity analysis. study_types was defined as a Zod enum in literature.search but similarly missing from the JSON schema.

Both fields are now published in their respective inputSchema objects with full type and description. tests/schemas/mcpToolSchemas.test.ts extended with 2 new drift-guard assertions (total: 13).

v1.11.2 (2026-05-22) — Fix: expose 6 hidden hta_dossier fields in MCP schema

heterogeneity_per_outcome, upgrading_per_outcome, severity_modifier, health_inequalities, pv_classification, and regulatory_landscape were all accepted and used by the hta.dossier handler but absent from the published MCP JSON inputSchema. External MCP clients (Claude Desktop, Smithery, etc.) could not discover or pass these fields, silently breaking the pipe workflows that pv_classify and regulatory_status_check were designed to feed into hta.dossier.

All 6 fields are now fully documented in the inputSchema with types and descriptions matching the Zod schemas. tests/schemas/mcpToolSchemas.test.ts extended with 6 new drift-guard assertions.

v1.11.1 (2026-05-22) — Bug fixes: MFN schema exposure, PartSA MFN runner, telemetry

Fixed: MFN fields missing from MCP-published tool schemas

mfn_sensitivity was implemented in the models.cost_effectiveness Zod schema and handler but absent from the exported costEffectivenessModelToolSchema JSON — external MCP clients (Claude Desktop, Smithery, etc.) could not discover the field. Likewise mfn_context was missing from the hta.dossier MCP inputSchema. Both fields are now present in the published schemas with full descriptions and required-field lists.

Added tests/schemas/mcpToolSchemas.test.ts as a permanent drift guard so Zod and MCP schemas can't diverge silently again.

Fixed: MFN sensitivity always used Markov runner even for PartSA models

runMfnSensitivity was called with runMarkovAndComputeICER regardless of model_type. When the base model was partsa, the MFN curve was computed from a Markov run, producing mixed-method output (PartSA base case + Markov MFN curve). The callback now dispatches to runPartSA when model_type="partsa", producing a consistent single-method result.

Fixed (web): hta_body enum in web/lib/tools.ts missing "gvd"

The web-tier tool definition exposed ["nice", "ema", "fda", "iqwig", "has", "jca"] — "gvd" was present in the MCP server schema but not in the Claude web UI tool definition. GVD dossiers were silently inaccessible from the web UI.

Fixed (web): MCP tool errors tracked as status=ok in PostHog

McpSession.dispatch() catches all errors and returns "Error: ..." strings (by design — so Claude receives the error text). The chat route's trackToolCall call sat immediately after dispatch() and always emitted status: "ok". The route now checks for the "Error: " prefix and emits status: "error" with error_class: "McpError" and the message body.

Also fixed: the PostHog distinctId was hardcoded to "chatgpt_adapter" for all surfaces. Claude web UI calls now use distinctId: "anon_claude_web" so the two surfaces are distinguishable in analytics.

v1.11.0 (2026-05-09) — MFN-aware tooling: basket data, dossier section, CE price sweep

Implements the Most-Favored-Nation pricing layer across three tool surfaces. Triggered by CMS proposed GUARD (Part D) and GLOBE (Part B) payment models, which anchor US drug prices to a 19-country OECD basket minimum — a structural shift that makes the gap between US net price and the MFN ceiling a first-order market-access input. Design log #27.

New: src/data/mfnBasket.ts — 19-country basket data + ceiling math

• MFN_BASKET_2026 — canonical 19-country ISO-2 list (AT BE CZ DK FR DE IE IT NL NO ES SE CH GB AU JP KR CA IL) per CMS GUARD/GLOBE proposed rule, revision 2026-03.
• computeMfnCeiling(basket_prices, opts?) — returns { ceiling: number|null, contributing_countries, missing_countries }. Returns null when no basket prices supplied; never fabricates a ceiling from memory.
• Rejects negative / NaN / Infinity prices with TypeError.
• 16 unit tests in tests/data/mfnBasket.test.ts.

New: src/models/mfnSensitivity.ts — deterministic ICER price sweep

runMfnSensitivity(baseParams, inputs, runModel) sweeps drug price from min_basket to current_us_price at N discrete points (default 11) and returns:

• curve — ICER at each price point.
• crossovers — price at which ICER crosses each WTP threshold (linear interpolation); null when the curve never crosses.
• icer_at_ceiling / icer_at_current — convenience aliases for first/last curve points.

Why deterministic sweep instead of another PSA? MFN is an exogenous price shock, not statistical uncertainty. 11 Markov runs vs 1000+ PSA runs; output is payer-readable ("ICER drops from $X to $Y; WTP crossover at price $Z").

13 unit tests in tests/models/mfnSensitivity.test.ts.

Extended: models.cost_effectiveness — mfn_sensitivity input field

When caller supplies mfn_sensitivity: { min_basket, current_us_price, n_points?, wtp_thresholds? }:

• Runs the deterministic price sweep after the base-case model.
• JSON output gains mfn_sensitivity: { range, curve, crossovers, icer_at_ceiling, icer_at_current }.
• Text output gains a ### MFN Price Sensitivity section with the price table and WTP crossover bullets.
• Zod schema enforces min_basket ≥ 0, current_us_price ≥ 0, n_points 2–101.

6 integration tests in tests/tools/costEffectivenessModelMfn.test.ts.

Extended: hta.dossier — mfn_context input field

When caller supplies mfn_context: { basket_prices, us_current_net_price?, basket_revision?, excluded_countries? } and the HTA body is NICE / EMA / FDA / IQWiG / HAS / GVD:

• Renders an MFN Exposure section with the full 19-country basket table, computed MFN ceiling, gap-to-US %, and 4 mitigation strategy recommendations (evidence-package investment, managed-entry agreements, launch sequencing, confidential rebate structures).
• The section is opt-in (requires basket_prices to be non-empty). No auto-render body in v1.11.0 — AMCP is pending design log #24.
• GVD dossier early-return branch also includes the MFN section (mirrors design log #26 Regulatory Landscape pattern).

15 integration tests in tests/tools/htaDossierMfn.test.ts.

Extended: src/server.ts — MFN telemetry flags

trackToolCall on success now includes:

• mfn_sensitivity_invoked: true when models.cost_effectiveness is called with mfn_sensitivity.
• mfn_context_emitted: true + mfn_basket_countries: N when hta.dossier is called with mfn_context.

Enables PostHog HogQL queries to measure MFN feature adoption without schema changes.

Extended: web tier — SYSTEM_PROMPT + tool schema

• web/lib/claude.ts SYSTEM_PROMPT: new MFN (MOST-FAVORED-NATION) PRICING & GLOBAL ACCESS STRATEGY block. Covers 3 market archetypes (evidence-constrained / IRP-influenced / structural), evidence-anchor strategy, when to call mfn_sensitivity vs mfn_context, and a hard rule against fabricating basket prices.
• web/lib/tools.ts: cost_effectiveness_model schema gains mfn_sensitivity object; hta_dossier schema gains mfn_context object. Claude can now pass both without schema errors.

12 web tests in web/__tests__/mfnPhase4.test.ts.

Full test suite

1133 tests passing (1121 MCP + 12 new web). 0 failures.

v1.10.2 (2026-05-12) — stop reusing the 500-char telemetry cap as the client response

A 0-CRITICAL hygiene release that fixes a quiet bug discovered while debugging a real ChatGPT failure on 2026-05-12 09:15:24 (user 1be263 called hta.dossier with no payload).

The bug

classifyToolError in src/analytics.ts returned a single error_message field, truncated to 500 chars. That truncation was intended for telemetry hygiene — PostHog event properties have a size limit, and a multi-issue ZodError dump on a heavy schema can run 1-2KB.

src/server.ts:475 then used that same truncated string as the client-facing response content (the text of the text content block returned to whoever called the tool). Multi-issue ZodErrors arrived at clients (ChatGPT Custom GPT especially) with the JSON cut mid-key — "received": "string", "rece — and were unparseable. ChatGPT bounced instead of retrying with the missing fields. Five real-user errors followed this pattern in the 14 days before discovery.

The fix

Split the field. classifyToolError now returns:

{
  error_class: string;       // unchanged — Error subclass name for dashboards
  error_message: string;     // FULL text — for the client response
  telemetry_message: string; // capped at 500 chars — for PostHog
}

server.ts:472 now passes telemetry_message to trackToolCall (PostHog hygiene preserved) and uses error_message for the text content (full message reaches the client).

Why it isn't strictly redundant with the web-tier fix (deployed 2026-05-12)

The web tier (web/lib/zodErrorFormatter.ts) already reformats raw ZodError JSON arriving from MCP into one-line "field.path: Required" text, AND salvages complete issues from truncated arrays. So clients calling via the web/ChatGPT adapter are protected today even on v1.10.1.

But:

1. Direct MCP clients (Claude Desktop, Cursor, the npm npx users) don't go through the web tier. They see the raw truncated ZodError straight from Railway. v1.10.2 fixes their experience.
2. The web-tier fix is defensive masking; v1.10.2 is the root-cause fix. Both layers help — defense in depth.

Tests

tests/analytics/errorClassification.test.ts updated to cover the split:

• error_message preserves full text (2000-char input → 2000-char output).
• telemetry_message capped at ≤500.
• ZodError on an 8-required-field schema produces an error_message > 500 chars (regression for the 2026-05-12 hta.dossier failure pattern).

Full suite: 8/8 in errorClassification, no other tests touched.

Non-breaking for consumers

error_class and error_message are still present and PostHog dashboards keep working unchanged (telemetry is now slightly more selective about which message it stores). The only behavioral change visible to a client of the MCP server is: error responses are no longer truncated mid-message. That's a strict improvement.

v1.10.1 (2026-05-10) — auto-wire regulatory.status_check (the "make the right thing easy" follow-up to v1.10.0)

v1.10.0 shipped the primary-source regulatory lookup tool. v1.10.1 closes the loop: the tool now fires automatically inside evidence.unmet_need and a new hta_workflow Phase 3.6, so the model can no longer fabricate a "no approved option" claim by simply forgetting to call it. Design log #26.

evidence.unmet_need — default-on regulatory fan-out

When treatment_landscape.current_soc[] is supplied, the handler now fans out to regulatory.status_check for each molecule across the user-supplied jurisdictions[]. Results are injected as a structured regulatory_context[] array AND rendered as inline label-quote attributions in the treatment-landscape paragraph ("Per FDA/OpenFDA label retrieved 2026-05-10: fremanezumab is approved for the preventive treatment of migraine in adults and in pediatric patients 6 years of age and older [citation N].").

• Default-on; opt out via auto_check_regulatory: false.
• Region mapping: us → us; de/fr/it/es/nl → eu; uk → uk (currently degrades gracefully — no UK source yet); jp → graceful gap.
• Citations auto-numbered into the existing registry.
• Concurrency capped at 8 per request (see autoCheck.ts).
• 24h cache shared with explicit regulatory.status_check calls — repeated drug/region lookups within a workflow are free.

hta_workflow Phase 3.6 — new "regulatory_landscape" phase

Inserted between Phase 3.5 (evidence.unmet_need) and Phase 4 (CE model). Fans out across comparators surfaced in earlier phases, pipes results into hta_dossier as a new regulatory_landscape[] parameter. Always runs when comparators are present, regardless of hta_body. Adds ~5-10s to total workflow time on a typical 4-comparator dossier.

hta_dossier — new "Regulatory Landscape" section

Renders for nice / jca / gvd / amcp bodies. Table format: comparator × region × current approved indication × label-revision date × source URL. Provides auditable provenance for the regulatory claims that downstream payers verify line-by-line.

Graceful degradation — non-negotiable

api_error or current_status: "unknown" from regulatory.status_check never blocks dossier rendering. Instead the failure is appended to gaps[]:

• "regulatory_status check failed for {drug} ({region}) — verify label manually before submission"
• "{drug} not found in {region} regulatory database — primary-source verification needed; did you mean: {suggestion_1}, {suggestion_2}?"

The dossier proceeds with whatever regulatory context is available. This is the same design philosophy as the literature-search degradation: surface gaps explicitly, don't fail the workflow.

Cycle safety

regulatory.status_check's handler is statically prevented from importing evidence.unmet_need (per design log #26 Q10). Tests assert this so a future contributor doesn't create an A→B→A loop.

Rate-limit headroom

OPENFDA_API_KEY env var is now respected by the OpenFDA client (the v1.10.0 implementation accepted it but the wiring shipped here). Anonymous OpenFDA limit is 240 req/min; with key, 120K/day. Production should set the env var.

Tests

29 new regression tests across autoCheck (9), evidence.unmet_need integration (8), hta_workflow Phase 3.6 (6), hta_dossier regulatory-landscape rendering (8). Full suite: 111 suites / 1069 tests (up from 110 / 1037).

Compatibility

• Tool count stays at 28 — no new tool, only auto-wiring of v1.10.0's tool into two existing tools.
• auto_check_regulatory: false preserves the v1.10.0 behavior for callers that want the explicit-call path.
• Existing evidence.unmet_need callers see no breaking change unless they were depending on the absence of regulatory_context[] in the output (unlikely).

v1.10.0 (2026-05-10) — regulatory.status_check tool (#28) — primary-source label lookup

New tool that closes a real-user incident category. Design log #25.

The trigger — fremanezumab/pediatric-migraine, 2026-05-07

Michael's colleagues at work asked evidence.unmet_need for a fremanezumab/pediatric-migraine dossier. The output asserted "CGRP mAbs have no approved pediatric indication" — true at LLM training cutoff, false since FDA approval of AJOVY (fremanezumab-vfrm) for pediatric episodic migraine in August 2025 (sBLA 761089/s031). Same staleness trap is waiting on every drug with recent label changes (Aimovig, lecanemab, donanemab, biosimilars, withdrawals…). Pointing the LLM at orange_book via literature_search returns product index entries, not the current Indications and Usage section. HEORAgent had no canonical regulatory-status lookup. v1.10.0 ships one.

What the tool does

regulatory.status_check({ drug: "fremanezumab", region: "us", indication?: "migraine" })

Returns:

• current_status — approved | pending | withdrawn | unknown (never not_approved on database miss)
• approved_indications[] — verbatim label text + age/weight constraints + approval date
• recent_label_revisions[] — last 12 months of changes
• source_urls[] + data_fetched_at for full auditability
• did_you_mean[] — Levenshtein suggestions on no-match (catches typos before the analyst burns hours)

The CRITICAL invariant

current_status never equals "not_approved". Primary-source absence ≠ proof of non-approval — that's the exact fremanezumab failure inverted. Database miss → unknown + did_you_mean[]. Documented in the tool description, asserted in tests.

Sources

• US: OpenFDA (drug/label endpoint) — primary. Optional OPENFDA_API_KEY for higher rate limits (wiring landed in v1.10.1).
• US: DailyMed — cross-check for verbatim Indications and Usage text.
• EU: EMA EPI FHIR — adapter against the EMA Open Data clinical-data API.
• UK: stub — placeholder for eMC + NICE TA index; v1.7.x lookahead.

Caching

24h TTL, in-memory, shared across MCP sessions. force_refresh: true bypasses. Cache key includes drug-name normalisation so Fremanezumab / fremanezumab-vfrm / AJOVY hit the same entry.

Tests

Live OpenFDA smoke test confirmed primary-source retrieval on real label queries. Full suite 110 suites / 1037 tests at v1.10.0 ship.

Companion fix bundled in this release: Codex review P1+P2+P3

Three correctness fixes that shipped alongside the new tool:

• P1 (CE model): model_type: "partsa" silently fell through to Markov when survival_inputs was missing — Zod stripped the field because it was never in the schema. Added survival_inputs to CEModelSchema + the exported tool schema; the handler now hard-fails when partsa is set without it instead of degrading to the wrong model class.
• P2 (workflow): utility_inputs was being built with only one of two QALY fields, then rejected by CE schema (silent fallthrough). Now requires both fields together.
• P3 (workflow): unmet_need_inputs existed in the internal Zod schema but was missing from the exported MCP tool schema — clients couldn't discover the GVD Phase 3.5 surface. Added to tool inputSchema with description.

7 new regression tests for these (3 PartSA, 2 utility, 2 schema-exposure).

v1.6.3 (2026-05-07) — code-review polish for v1.6.2 + Slack-digest hardening

Two parallel reviewers audited v1.6.2 and the new Slack weekly-digest feature within hours of ship. Combined: 0 CRITICAL, 0 HIGH, 6 MEDIUM, 6 LOW. All real findings addressed; cosmetic items deferred. Total tests 833 → 838.

Fixed (schema, MCP server)

• Whitespace tolerance in caseInsensitiveEnum. val.trim().toLowerCase() before lookup — " NICE " now normalises to "nice" instead of falling through to invalid_enum_value. Test file already comment-promised this; now wired. +4 regression tests.
• risk_of_bias.instrument + risk_of_bias.output_format + hta_dossier.intervention_impact — three enums missed in v1.6.2's class-wide application. LLMs naturally pass "RoB2", "AUTO", "Narrows". Now case-insensitive consistent with the rest of the surface.
• Tool description hints — every case-normalised tool now explicitly advertises case-insensitivity in its top-level description so LLMs reading the JSON Schema (project_create / pv_classify / irb_review / hta_dossier / jca_pico_scope / risk_of_bias) actually learn the schema is permissive.

Fixed (Slack weekly digest)

• PostHog 200-with-error-field check in hogql(). PostHog returns HTTP 200 with {error: "..."} for query-level failures (bad HogQL, quota exceeded, project ID mismatch). Pre-fix, the digest silently posted "0 events, 0 users" without surfacing why. Now throws a typed error.
• runWeeklyDigest Promise.all → Promise.allSettled with per-source fallbacks. Single-source failure (esp. anonymous-rate-limited GitHub) no longer kills the whole digest. Failed sources surface as a "⚠️ Degraded sources this run: ..." prepended insight bullet so a missing GitHub-stars row is self-explanatory. PostHog stays load-bearing — both PostHog calls failing still throws.
• AbortSignal.timeout(8000) on every external fetch (npm, GitHub, Railway health, npm registry, all 7 PostHog HogQL queries). Pre-fix, a single stalled call could eat the cron route's 60s budget and silently miss the Monday digest.
• Optional GITHUB_TOKEN env support. Anonymous GitHub API limit is 60 req/hr per IP; Vercel functions share IP pools, so the limit is hit unexpectedly. A classic PAT (no scopes needed) raises the ceiling to 5,000/hr.

Added — pinning tests

• studies: {} empty-object input — pins the documented degraded-but-non-erroring behavior (singleton-wrap → all-defaults → Unclear-on-all-domains result) so a future schema-strictness change can't silently break it.
• risk_of_bias.instrument case-insensitive regression test.

Skipped (cosmetic)

• Type inference widening (z.ZodEffects → string instead of T[number]) — only matters if we add discriminated-union switches downstream, which we haven't.
• Misleading "constant-time" comment in cron route — !== is fine for our threat model on a 32-byte hex secret used only by Vercel infrastructure; comment removed.
• weekStart doc-comment off-by-one — code is correct, comment was misleading; deferred.
• Engagement-gap heuristic n=1 sample noise — wait until we have data showing it actually fires on noise, then tune.
• Token-in-URL — accepted by design (bookmark UX).

Tests

833 → 838 MCP tests passing (+4 trim regression + 1 studies:{} pinning + 1 instrument case-insensitive). 154/154 web tests still passing (Slack stats fixes are network-dependent paths; covered by type-check + manual audit, no fetch-mocking integration test added in this patch).

Non-breaking

All changes are silent failure-mode hardening + LLM ergonomics. No API surface changes; no migration needed.

v1.9.2 (2026-05-09) — polish: Nelder-Mead early exit + 6 review nits

Six small quality improvements deferred from the v1.9.1 review. None are correctness fixes; these are polish on the v1.7-1.9 work. Two have a measurable performance impact:

Performance — full test suite 215s → 19s (11×)

Two of the changes turn out to dominate overall test runtime:

• Nelder-Mead convergence-tolerance early exit in survivalFitting.ts. Pre-fix, the optimizer ran the full maxIter=800 iterations regardless of convergence — typical fits converge in ~50-200 iterations, so 600+ were wasted compute. Added if spread < 1e-8 (or 1e-6 relative) break after the simplex sort. Real survival-MLE fixtures now converge in 50-150 iterations.
• Log floor 1e-300 → 1e-30 in logLikelihoodFromEvents. The ultra-deep floor created near-flat likelihood surfaces in pathological starting regions; raising the floor lets the optimizer navigate cleanly. Combined with the early-exit, this halves runtime for the harder distributions (Log-normal, Gompertz). R's flexsurv and survival packages use a similar order-of-magnitude floor.

Correctness / hygiene

• Bootstrap RNG seeding for tests. computeEVPPI and bootstrapEVPPICI now accept an optional rng parameter (defaults to Math.random). Tests pass a seeded mulberry32 so the bootstrap CI is reproducible across runs. The "CI tightens at N" test no longer needs its +1 slack — assertion tightened to strict widthLarge < widthSmall.

• Tightened parameter-recovery tolerances in tests/models/survivalFitting.test.ts:

  • Exponential N=500: 20% → 10% (~1.5 SD)
  • Exponential N=1000: 12% → 7%
  • Weibull N=500: 25% → 15% (joint 2-param MLE has higher variance)
  • Log-normal μ: abs<0.4 → abs<0.25; σ: 30% → 18%
  • Heavy 60% censoring: 30% → 20%

  Tighter tolerances catch a 10% systematic bias that the previous 3-4 SD width would have missed. All still pass on the seeded fixtures.

• Documented the magic init values in each survival fitter (Weibull [1.0, scaleInit], log-logistic [m, 1.5], log-normal [muInit, 0.8], Gompertz [0.01, rateInit]). Each now has a comment explaining the empirical reasoning so a future maintainer doesn't over-tune them.

• Removed void splitSentences/tokenizeWords/countSyllables workaround in icfReadabilityCheck.ts. These imports were never directly called in the handler — they were "tree-shaking guards" that silently relied on indirect use through computeStats / computeReadabilityScores. The unit tests already exercise them directly, so the explicit imports + void no-ops were dead code. Cleaned up.

Tests

909/909 still passing. No new tests added — all changes either tighten existing assertions or are pure refactors of correct code.

Performance impact in production

Negligible. The Nelder-Mead early exit speeds up survival_fitting IPD calls by ~3-4× in the typical case (faster convergence) but wall-clock for a 500-patient fit was already <100ms before, so the user-visible difference is "fast" → "very fast". The log floor change is functionally invisible at production parameter values.

Non-breaking

All changes are pure quality improvements. No API surface changes; no observable behavior change at production parameter ranges.

v1.9.1 (2026-05-09) — code-review fixes for v1.7.0 / v1.8.0 / v1.9.0

Two parallel reviewers (math/statistics + ICF formula correctness) audited the v1.7-1.9 work within hours of ship. 0 CRITICAL, 4 HIGH, 8 MEDIUM, 7 LOW. All 4 HIGHs were real correctness bugs in code that produces HTA / CMS / IRB-grade outputs. All addressed in this patch.

Fixed (HIGH)

• EVPPI bootstrap CI upper-bound was systematically downward-biased (src/models/evppi.ts:bootstrapEVPPICI). Pre-fix the bootstrap loop capped each resample at the original sample's totalEVPI, truncating the upper tail of the bootstrap distribution whenever a resample's empirical totalEVPI exceeded the original. The fix removes the in-loop cap; only the FINAL reported percentile bounds are clamped. Decision-makers reading "EVPPI = $1,200 (95% CI $400–$2,800)" now see honest tails — narrower CIs no longer underestimate research value.

• Survival IPD mean_survival_restricted was mislabeled across all 5 distributions (src/models/survivalFitting.ts:fitXFromEvents). Pre-fix the IPD path returned unrestricted means (Exp: 1/λ, Weibull: scale·Γ(1+1/shape), Log-normal: exp(μ+σ²/2)), the median (Log-logistic: α), or — most egregiously — the EXPONENTIAL DISTRIBUTION's mean/median ratio applied to a Gompertz median (median × 1/ln(2) = median × 1.4427). All five fitters now call the existing restrictedMean(kmTable, survFn) helper for proper numerical integration of S(t) over [0, max_observed]. Wrong RMST → wrong QALY → wrong ICER → wrong reimbursement decision. The KM-table path was unaffected (already used restrictedMean() correctly).

• countComplexWord -es suffix stripping under-counted complex words (src/icf/syllables.ts). Pre-fix unconditionally stripped trailing "-es", so 3-syllable plurals like processes (pro-ces-ses) or addresses (ad-dress-es) became 2-syllable process/address and missed Gunning's complexity threshold. Result: Gunning Fog and SMOG scores under-reported ICF difficulty — investigators received better scores than reality and didn't rewrite sentences they should. Fix: only strip -es/-ed/-ing when the syllable count is unchanged after stripping (a non-syllabic morphological inflection).

• "effectiveness" removed from medical-jargon dictionary (src/icf/jargon.ts). The previous entry directly contradicted FDA's "Communicating Risks and Benefits" (2011) and NIH Plain Language guidance, both of which recommend "effectiveness" AS the plain-language replacement for "efficacy". An investigator who had already done the right thing would see it flagged and revert to the harder term. The matching "efficacy" entry remains.

Fixed (MEDIUM)

• Verdict logic OR/AND mismatch documented as AND (src/icf/types.ts). The runtime code uses AND semantics ("FKGL ≤ target+1.5 AND <40% exceed → borderline"); the type comment said OR. Aligned the comment to the code; AND is the patient-safety direction.

• Jargon recommendation now fires for any hits, not only ≥3 (src/tools/icfReadabilityCheck.ts). Pre-fix a passing FKGL with 2 jargon terms would emit no jargon-rewrite recommendation. Threshold dropped to ≥1; output cap remains at 5 terms with a "+N more" suffix.

• worst_sentences filtered to only target-exceeding sentences. Pre-fix the field was top-5-by-FKGL regardless; programmatic consumers could see "worst" sentences within target. Now matches the markdown rendering (which already filtered).

• Pass-with-jargon messaging implicitly fixed by the jargon-threshold drop above. A passing FKGL with jargon hits now correctly emits the jargon recommendation rather than the misleading "✅ No rewrite recommendations" line.

Skipped (cosmetic)

• Bootstrap RNG seeding for the "CI tightens at N" test (theoretically flaky but hasn't bit yet)
• 1e-300 log floor in survival MLE tightening to 1e-30 (hasn't caused convergence issues empirically)
• Nelder-Mead convergence-tolerance early exit (test suite is slow but acceptable)
• Initial-parameter documentation for log-logistic / log-normal / Gompertz fitters
• Tightening parameter-recovery test tolerances (current 3-4 SD width is loose but catches the 50% bugs we care about)
• Hidden-import void workaround in icfReadabilityCheck.ts handler (cosmetic)

Tests

10 new regression tests:

• 1 EVPPI bootstrap CI bound check (no longer artificially capped)
• 5 ICF countComplexWord cases (processes, addresses, cakes, walking, encyclopedia)
• 2 ICF dictionary integrity (effectiveness removed, efficacy retained)
• 1 ICF worst_sentences only-exceeding invariant
• 1 ICF jargon recommendation fires for any hits

899 → 909 MCP tests passing. Web tests still 177/177.

Non-breaking

All changes are bug fixes. No API surface changes. The mean_survival_restricted field semantics are now correct (RMST instead of unrestricted mean) — callers that already used it as RMST per its documented meaning get more accurate values; callers that were treating it as unrestricted mean were getting the wrong field anyway.

v1.9.0 (2026-05-09) — survival_fitting patient-level MLE path (no longer ⚠️ EXPERIMENTAL on the IPD input)

The tool now accepts patient-level event-time data (event_data: Array<{time, event: 0 | 1}>) alongside the legacy km_data step-summary path. Caller picks one (Zod refine enforces). When event_data is supplied, the fit is true right-censored maximum likelihood per Collett (2015) and NICE DSU TSD 14 (Latimer 2013) — no approximation warning. The KM-table path remains supported for back-compat with literature-digitization workflows but emits an explicit "approximation, less reliable" warning and points the caller at event_data.

Added

• event_data input with at least 5 patient-level rows. Each row is {time, event} where event=1 for an observed event and event=0 for right-censoring at time.
• fitSurvivalCurvesFromEventData() model entry point. Five distributions (Exponential, Weibull, Log-logistic, Log-normal, Gompertz) fit via Nelder-Mead simplex on the proper right-censored log-likelihood Σᵢ [δᵢ·log(f(tᵢ)) + (1-δᵢ)·log(S(tᵢ))].
• Kaplan-Meier curve from event data. When event_data is supplied, the markdown report's "KM Observed" column is built from the standard KM estimator on those same rows — no separate input needed.
• Mutually-exclusive Zod validation. Caller passes exactly one of km_data or event_data; passing both or neither raises a clear error.

Changed

• Methodology line branches: event_data path cites Collett 2015 + TSD 14; km_data path is honest that it's an interval-censored approximation.
• Tool description no longer leads with "⚠️ EXPERIMENTAL"; instead positions event_data as the preferred path with km_data as legacy.
• CLAUDE.md ⚠️ EXPERIMENTAL list trimmed from 2 → 1: only population_adjusted_comparison (MAIC/STC) remains.

Tests

14 new IPD-path tests in tests/models/survivalFitting.test.ts:

• Schema invariants (rejects N<5, returns 5 fits, monotonic KM curve)
• Parameter recovery — simulate from known Exponential(λ=0.05) / Weibull(shape=1.5, scale=20) / Log-normal(μ=2.5, σ=0.6) and verify recovered params within 20-30% at N=500-1000 with seeded mulberry32 PRNG (deterministic). Strongest available evidence the MLE is correct.
• Model selection sanity (correct distribution wins by AIC on truly-from-that-distribution data)
• S(0)=1, monotonic decreasing, S(median)≈0.5 invariants
• Heavy 60% censoring still recovers parameters within 30%
• Zero-censoring corner case

5 additional tool-level tests for the new schema paths (event_data path methodology, KM-path "approximation" warning unchanged but no longer says "EXPERIMENTAL", mutual-exclusivity validation).

882 → 901 MCP tests passing.

Non-breaking

• km_data API surface unchanged. Existing callers that pass km_data get the same fit they got before, with a slightly clarified warning ("approximation" instead of "EXPERIMENTAL").
• event_data is purely additive — no migration needed.

v1.8.0 (2026-05-09) — icf_readability_check tool (paired with irb_review)

New tool. Closes the v2 deferral from design log #21: paired ICF readability analyzer that was promised when irb_review v1 shipped.

Added — icf_readability_check (icf.readability_check)

Takes ICF text, returns:

• Readability scores — Flesch-Kincaid Grade Level, Flesch Reading Ease, Gunning Fog Index, SMOG Grade.
• Per-sentence breakdown with worst-5 sentences (FKGL desc) flagged so investigators see exactly which sentences exceed the target grade level.
• Medical-jargon detection — curated dictionary of ~80 high-frequency clinical-trial-consent terms (placebo, randomized, adverse event, pharmacokinetics, comorbidity, etc.) with plain-language alternatives. Case-insensitive whole-word matching with optional plural matching (adverse event → also catches adverse events).
• Pass / borderline / fail verdict vs target grade level (default 8 per FDA/NIH guidance; configurable 4-12).
• Concrete rewrite recommendations when verdict is not pass — targeted at worst sentences + jargon hits + sentence-length / syllables-per-word patterns.

Pure logic, no external API. <300ms on a 50-sentence ICF.

References baked into the methodology

• Kincaid JP et al. (1975) — FKGL formula
• Flesch R. (1948) — Reading Ease
• Gunning R. (1952) — Fog Index
• McLaughlin GH. (1969) — SMOG
• NIH Plain Language Guidelines (clinical-trial consent)
• FDA Communicating Risks and Benefits (2011)

Tests

34 new tests across schema validation, syllable counting (heuristic ±1 of CMU dict), sentence splitting (handles abbreviations: Dr. / Mr. / e.g. / i.e.), word tokenization, FKGL/FRE formula correctness on known reference texts, per-sentence breakdown, jargon detection (case-insensitive, whole-word, capped at 5 occurrences), verdict logic, output structure, performance.

848 → 882 MCP tests passing. Web tool count assertions bumped 26 → 27 across 4 test files.

Tool count

26 → 27. Full tool list: literature.search, literature.screen, evidence.network, evidence.indirect, evidence.population_adjusted, evidence.survival, evidence.risk_of_bias, evidence.itc, evidence.clinical_scale, evidence.unmet_need, models.cost_effectiveness, models.budget_impact, hta.dossier, hta.utility, hta.workflow, utils.validate_links, project.create, knowledge.search, knowledge.read, knowledge.write, examples, workflow.maic, pv.classify, pv.signal_workflow, jca.pico_scope, irb.review, icf.readability_check ← new.

v1.7.0 (2026-05-09) — EVPPI promoted out of ⚠️ EXPERIMENTAL

Three quality fixes to the Strong-2014 binning estimator in cost_effectiveness_model's EVPPI path. Removes the long-standing CLAUDE.md caveat ("non-parametric binning, noisy when total EVPI ~0").

Fixed

• Mathematical cap. EVPPI ≤ totalEVPI by definition (per-parameter information value can't exceed full-uncertainty resolution). Pre-fix the binning estimator could overshoot due to sample noise, producing evppi_proportion > 1.0 in rare cases. Now Math.min(raw, totalEVPI) per parameter.
• Noise-floor guard. When the decision is robust to uncertainty (totalEVPI ~ 0), per-parameter binning still produced fake positive signals from sample noise — leading to a misleading "top 5 parameters worth researching" table built from pure noise. We now compute a noise-floor threshold relative to NMB stddev (0.5% × stddev(NMB)); if totalEVPI falls below it, all per-parameter EVPPIs are suppressed to 0 with below_noise_floor: true. The markdown report surfaces a clear "Decision is robust to all uncertainty" message instead of the empty/noisy table.
• Adaptive bin width via Freedman-Diaconis (h = 2·IQR·N^(-1/3)) replacing Sturges' rule. F-D adapts to actual data spread and handles non-normal distributions better — important for cost / utility parameters which are typically right-skewed. Falls back to Sturges' for constant or tied parameters.
• Constant-parameter short-circuit. When a parameter has zero variance, EVPPI is now hard-coded to 0 (rather than picking up unrelated NMB variance from the sort being arbitrary on equal keys). Fixes a fake-positive that the binning estimator alone couldn't avoid.

Added

• Bootstrap 95% confidence interval per parameter (200 resamples with replacement). Each EVPPIResult now carries evppi_ci_lower, evppi_ci_upper, and evppi_se. Skipped for parameters below the noise floor (no point in CI on noise) and for tiny samples (N < 50). Markdown report shows the CI alongside the point estimate: | Parameter | EVPPI | 95% CI | % of EVPI |.

Tests

13 new EVPPI tests across 6 describe blocks:

• Basic invariants (small-N skip, missing-param skip, sort order, non-negativity)
• Mathematical cap (EVPPI ≤ totalEVPI; proportion ∈ [0,1])
• Noise-floor guard (suppression on robust-decision fixture; doesn't fire on high-uncertainty fixture)
• Bootstrap CI (CI brackets point; CI ≥ 0; CI tightens as N grows; CI omitted below noise floor)
• Constant-parameter handling

848/848 tests passing (was 835).

Non-breaking

EVPPIResult adds 4 optional fields (evppi_ci_lower, evppi_ci_upper, evppi_se, below_noise_floor). The pre-existing evppi, evppi_proportion, parameter fields are unchanged. No migration needed.

Open methodology gaps (deferred)

The binning estimator still doesn't match the gold-standard methods (GAM regression per Strong 2014, Gaussian-process regression per Heath-Manolopoulou-Baio 2018) for accuracy in challenging cases. Adding a real GAM smoother is ~2 weeks of work and is candidate for a future v1.x patch when there's appetite. v1.7.0 makes the binning estimator HONEST (no fake positives, proper uncertainty quantification, mathematical bounds) — appropriate for the current default use case where EVPPI is one of many sensitivity outputs, not the primary model output.

v1.6.2 (2026-05-07) — schema hardening for LLM input shapes

Two LLM-input-shape fixes surfaced by a PostHog audit of project.create and evidence.risk_of_bias errors.

Fixed

• Class-wide case-insensitive enums via shared src/util/caseInsensitive.ts. PostHog showed 5 production failures with hta_targets: ["NICE", "ICER"] — LLM callers naturally pass brand casing instead of canonical lowercase tokens. Vanilla z.enum() rejected; new helper preprocesses to canonical case before validation. Truly unknown values still fail with did-you-mean hints.

  Applied to:

  • project_create — hta_targets
  • pv_classify — study_design, primary_objective, regulatory_context, jurisdictions
  • irb_review — study_design, data_handling, risk_level, funding_source, jurisdictions, exempt_category_hint
  • hta_dossier — hta_body, submission_type, output_format
  • jca_pico_scope — drug_class, line_of_therapy, jurisdictions, regulatory_context

• risk_of_bias singleton studies auto-wrap. PostHog showed real-world calls with studies: {...} (singleton object) instead of studies: [{...}] (array). Pre-process auto-wraps before the array schema runs — preserves min(1) constraint and per-element parsing.

Tests

+9 helper tests (tests/util/caseInsensitive.test.ts) + 7 regression tests across project_create and risk_of_bias. Total 822 → 829 passing.

Non-breaking

Canonical lowercase still works exactly as before. New code only adds tolerance for upper/mixed case. No API surface changes; no migration needed.

v1.6.1 (2026-05-07) — hta_workflow GVD routing + Phase 3.5 unmet-need integration

Wires the new evidence.unmet_need tool from v1.6.0 into the hta_workflow orchestrator as Phase 3.5, between risk-of-bias and cost-effectiveness. Also extends hta_workflow to route GVD-specific section generators when hta_body: "gvd".

Added

• hta_workflow Phase 3.5 — automatically calls evidence.unmet_need and pipes the structured unmet_need_summary into the dossier draft. Default-on for any hta_body that surfaces unmet-need (NICE STA, EU JCA, GVD); skippable via skip_unmet_need: true when running an iteration.
• hta_workflow GVD routing — when hta_body: "gvd", the orchestrator routes through the v1.6.0 GVD section generators (Sections 1-13) instead of the generic skeleton, producing per-market subsections (US/UK/EU5/JP) and the gvd_evidence_pack pipe interface.

Fixed

• Phase 3.5 unmet-need parsing. Initial integration assumed the handler wrapped output in {content: ...}; in practice evidence.unmet_need returns the assessment object directly. Phase 3.5 now reads result.unmet_need_summary directly. Caught immediately post-1.6.0; shipped as 1.6.1 patch.

v1.6.0 (2026-05-07) — evidence.unmet_need tool + Global Value Dossier section generators

Two design-log items shipped together. Tool count 25 → 26.

Added — evidence.unmet_need (design log #23)

New tool: structured 4-dimension unmet-need framework. Inputs: indication + jurisdiction + optional literature_evidence (output from literature_search). Output: markdown report + structured unmet_need_summary JSON object that pipes into hta_dossier({hta_body:"gvd"}) Section 4 and hta_dossier({hta_body:"nice"}) for the NICE Severity & Inequalities section.

Four dimensions:

1. Disease burden — incidence/prevalence, mortality, morbidity, demographics
2. Treatment landscape gap — current SoC limitations, response rates, AE profiles, off-label patterns
3. QoL impact — EQ-5D / disease-specific instruments, work productivity, caregiver burden
4. Economic burden — direct medical, indirect costs, productivity loss, healthcare utilisation

Per-jurisdiction depth (light v1): adds country-specific epidemiology and SoC where the user supplies a jurisdiction code. Citations carry URL with pre-validation. 12+ tests.

Added — Global Value Dossier section generators (design log #22)

Existing hta_dossier({hta_body:"gvd"}) was a 13-section skeleton emitting generic boilerplate. v1.6.0 ships actual section generators that consume literature_search / risk_of_bias / evidence_indirect / cost_effectiveness_model / budget_impact_model / evidence.unmet_need outputs and produce GVD-specific prose:

• Section 1 — Disease background (consumes evidence.unmet_need Dimension 1)
• Section 2 — Treatment landscape (consumes evidence.unmet_need Dimension 2 + literature_search HTA precedent)
• Section 3 — Product profile (drug + indication metadata)
• Section 4 — Unmet need (full evidence.unmet_need output)
• Section 5-7 — Clinical evidence (consumes screened literature + RoB + GRADE)
• Section 8 — Economic evaluation (consumes cost_effectiveness_model results)
• Section 9 — Budget impact (consumes budget_impact_model results)
• Section 10 — Pricing & access (per-market subsections US/UK/EU5/JP)
• Section 11 — Reimbursement landscape per market
• Section 12 — Pharmacovigilance (consumes pv_classification if supplied)
• Section 13 — Patient access programs

Plus a gvd_evidence_pack pipe interface so GVD output can pre-fill country-specific dossiers (NICE / JCA / AMCP). DOCX table styling. AMCP Format 4.1 deliberately deferred to v1.7. 15+ tests.

v1.5.2 (2026-05-07) — live-formula XLSX + neurology clinical scales

Two more design-log items in a single release. Both surfaced gaps from the v1.4.x management benchmark vs Claude.ai (slide 6 "❌ today" → ✅).

Added — evidence.clinical_scale (design log #19)

New umbrella tool covering 6 neurology and cognitive scales:

• UMSARS (MSA — orphan, Phase-2-2028 JCA scope)
• UPDRS + MDS-UPDRS (Parkinson's)
• ADAS-Cog, MoCA, MMSE (Alzheimer's / cognitive)

Per-scale total + subscale scoring, MCID-based responder analysis (Krismer 2017 / Horváth 2015 / Andrews 2019 thresholds), trajectory comparison vs natural-history reference cohorts (NNIPPS / EMSA-SG / PPMI / ADNI summary-level v1). Time-to-milestone integration via survival_fitting.

Three new JCA indication sub-classes added to jca_pico_scope:

• neurology_msa — orphan, Phase 2 (2028) JCA scope
• neurology_pd — Phase 3 (2030) general medicines
• neurology_ad — Phase 3 (2030)

Per-country comparator universes:

• MSA: BSC across all (no DMTs in standard care)
• PD: levodopa / rasagiline / DBS depending on stage
• AD: donepezil / memantine / lecanemab / donanemab depending on stage

17 tests. Tool count 24 → 25.

Changed — live-formula XLSX upgrade (design log #20)

Refactored formatters/xlsx.ts so the XLSX output for cost_effectiveness_model and budget_impact_model emits live Excel formulas instead of pre-computed values:

• New "Markov Trace" sheet — n_cycles rows × 13 formula columns. Each row references the Inputs sheet so editing a transition probability or cost recomputes the trace in-place.
• Transition Matrix cells reference Inputs sheet directly.
• CEAC uses COUNTIFS formulas referencing the PSA sheet — drag the WTP threshold and the curve recalculates.
• Summary uses SUMPRODUCT referencing Markov Trace — ICER updates as inputs change.

PSA per-iteration values are kept as static numbers (audit reproducibility — re-running PSA stochasticity inside Excel would break determinism).

Same treatment for budget_impact_model XLSX (year-by-year SUM formulas referencing the inputs sheet).

15 tests. Closes the v1.4.x management benchmark "partial" rating on Slide 6. Customers can now genuinely edit any input → trace recomputes → ICER updates → CEAC curve shifts.

v1.5.1 (2026-05-06) — irb_review code-review fixes

Three parallel reviewers (regulatory accuracy with WebFetch verification, decision-tree correctness, test-gap analysis) audited v1.5.0 within hours of ship. 3 HIGH regulatory citation errors + 4 HIGH correctness bugs + 6 untested branches identified. All real findings verified against primary sources (eCFR via govinfo.gov / Cornell LII; EU CTR 536/2014 via legislation.gov.uk + European Commission) and patched. Total tests 683 → 708.

Fixed (HIGH — regulatory accuracy)

• Pregnant women consent trigger inverted (rulesets.ts). v1.5.0 wording said "Both parents' consent required when research holds no direct benefit" — that's §46.204(d), which actually requires only the woman's consent. Both-parent consent is §46.204(e) and the trigger is "research holds out the prospect of direct benefit solely to the fetus." An investigator following the v1.5.0 obligation text would have obtained father consent unnecessarily on no-benefit studies, or skipped it on benefit-solely-to-fetus studies. Corrected per eCFR §46.204(d) and (e) verbatim.
• §46.306 prisoner-research subcategories misattributed (rulesets.ts). v1.5.0 said "practices that may improve health/well-being of prisoners as a class is the broadest." That description maps to (a)(2)(iii) (class-level), not the broadest. The most commonly invoked sub-paragraph for therapeutic prisoner research is (a)(2)(iv) which is about the individual subject's health/well-being. Now enumerates all four sub-paragraphs (i-iv) with the correct attribution.
• §46.406 missing "generalizable knowledge" eligibility gate (rulesets.ts). v1.5.0 wording read "§46.406 (minor increase over minimal, no direct benefit)" — but §46.406(c) imposes an additional mandatory IRB finding: the research must be "likely to yield generalizable knowledge about the subjects' disorder or condition." Without this, healthy-child studies could be misclassified to §46.406 when they actually require the stricter §46.407 Secretary-determination pathway.

Fixed (HIGH — decision-tree correctness)

• NIH-funded research reported no COI obligation (decisionTree.ts:computeCoi). Pre-fix: funding_source !== "industry" returned { required: false, framework: "none" } for ALL non-industry funding. Real legal bug — PHS 42 CFR 50 Subpart F (FCOI regulation) applies to all PHS-funded research (NIH, AHRQ, CDC, HRSA, FDA, IHS, SAMHSA), not only industry. Now: phsApplies = onUs && (industry || nih || other_government) triggers PHS framework; EU CTR Annex I §M Point 66 unchanged (industry-only).
• Interventional full-board rationale falsely asserted "greater-than-minimal risk" when input was actually risk_level: "minimal" without marketed_drug hint. Now branches on the actual risk_level value — minimal-risk + no-hint reads "Interventional study at minimal risk but no marketed-drug/device hint supplied" with explicit guidance to set marketed_drug=true for cat 1 expedited.
• risk_level: "unknown" warning was gated to study_design === "interventional" only. Other designs (registry, retrospective_chart_review, non_interventional_prospective) silently fell through to full-board with no advisory. Now emits a tier-specific warning on each branch.
• benign_behavioural=true with non-interventional study_design silently ignored. Hint requires study_design === "interventional" per §46.104(d)(3); now warns when set on a non-interventional design instead of dropping silently.

Fixed (MEDIUM)

• CTR 536/2014 Article 14 → Annex I §M Point 66 for COI. The coi_framework: "eu_ctr_article_14" enum value pointed at the wrong article. CTR Article 14 is "Addition of a Member State" (extending a trial to additional EU Member States), not COI. Verified via legislation.gov.uk: investigator economic-interest disclosure lives in Annex I, Section M, Point 66 ("Suitability of the Investigator"). Breaking change to the structured coi_framework field: eu_ctr_article_14 → eu_ctr_annex_i_point_66. The cover-letter text and dashboard label also updated. Acceptable break — irb_review shipped only hours before this patch; no external consumers expected.
• HIPAA subsection citations promoted to user-visible output. v1.5.0 user output cited only "§164.514" without the (b)(1)/(b)(2) sub-precision. Now: "HIPAA §164.514(b)(2) Safe Harbor (18-identifier removal) or §164.514(b)(1) Expert Determination required prior to data sharing outside the covered entity."
• marketed_drug=true + risk_level: "greater_than_minimal" silently dropped. Now warns that expedited cat 1 requires minimal risk; PSUR SAE framework still applies via marketed_drug.
• computeIcfTier returned "standard" for benign-behavioural exempt-cat-3 studies. Pre-fix: only non-interventional designs unlocked "basic" ICF. Post-fix: isMinimal && (isNonInt || benign_behavioural) → benign-behavioural exempt cat 3 now correctly produces a basic ICF.
• Questionnaire-only cat-7 expedited path now emits §46.104(d)(2) second-prong advisory. Surveys with identifiable responses default to cat 7 expedited, but §46.104(d)(2) has a no-disclosure-risk prong this tool doesn't mechanise. Now warns explicitly so the IRB can apply the second prong manually for non-sensitive surveys.

Fixed (LOW)

• §46.407 wording: "HHS Secretary panel review" → "HHS Secretary determination after expert-panel consultation and public comment period" (the panel consults; the Secretary determines).

Tests

+25 new regression tests across:

• 5 v1.5.1 regulatory citation regressions (pregnant §46.204(e), prisoner §46.306(a)(2)(i-iv), pediatric §46.406(c), Annex I in cover letter, HIPAA subsection cites)
• 11 v1.5.1 decision-tree regressions (NIH/other_government COI, academic-only no-COI, foundation-EU no-COI, NIH-EU-only no-COI, interventional rationale text, 3 unknown-risk warnings on non-interventional designs, benign_behavioural mismatch, marketed_drug+greater warning, ICF basic for benign minimal, questionnaire second-prong advisory)
• 7 untested-branch regressions (retrospective+greater+identifiable, registry+minimal+pseudonymized, registry+greater, non_int default cat 4, hint precedence marketed > noninvasive, secondary_data+identifiable, specimen+identifiable)

683 → 708 tests, 100% pass rate, no regressions.

Process learning

The reviewer-hallucination memory (saved 2026-05-05 after 3 incidents in 48h) saved this patch from introducing fabricated regulatory citations. All 4 regulatory findings were verified against official sources before applying any patch — eCFR via govinfo.gov for §46.204, Cornell LII for §46.306 and §46.406, legislation.gov.uk for CTR 536/2014 Article 14 and Annex I. Each verification quote was checked verbatim against the reviewer's claim. The pattern of "spawn 3 reviewers in parallel, fan out by audit angle, verify regulatory claims via WebFetch before patching" is now the default for every release with regulatory output.

v1.5.0 (2026-05-06) — irb_review tool

New IRB / Ethics Committee submission classifier (design log #21). Pure decision-tree logic, <300ms, no external I/O. Tool count 22 → 23.

Added — irb_review (irb.review)

Classifies a planned study under 45 CFR 46 (US Common Rule) + EU CTR 536/2014 to produce an IRB submission scaffold. Inputs: study_design (7-enum), data_handling (5-enum), risk_level, funding_source, jurisdictions (us_irb / eu_cec), 4 vulnerable-population yes/no flags, optional pv_classification, optional expedited_category_claim, plus 9 hint flags that disambiguate exempt/expedited categories.

Outputs:

• US tier: Exempt §46.104 cat 1-8, expedited §46.110 cat 1-7, full-board §46.108. All 8 + all 7 categories reachable.
• EU tier: national-only / ctr_multi_state (with ~60d / ~45d timelines) / non_interventional_only.
• Vulnerable populations: Subpart B (pregnant) / C (prisoners) / D (children, with v2 age-tier-table commitment) / decisionally-impaired obligations.
• Data Management Plan: GDPR Art. 9 special-category trigger (EU + non-anonymous), HIPAA §164.514 PHI flag (US + identifiable), de-identification method (Safe Harbor / Expert Determination / Pseudonymization / must_implement / not_required), cross-border transfer obligations.
• SAE reporting: CTR 536/2014 Annex III (≤7d fatal, ≤15d other), FDA IND Safety (21 CFR 312.32), Post-marketing PSUR. pv_classification.primary_category="PASS_imposed" overrides to CTR Annex III regardless of jurisdiction.
• ICF complexity tier: complex for full-board OR vulnerable, basic for minimal-risk + non-interventional + no vulnerable, standard otherwise.
• COI framework: PHS 42 CFR 50 Subpart F (US) and/or EU CTR Article 14 (EU); fires when funding_source="industry".
• Cover-letter template: ready-to-paste 200-300-word block with drug, indication, review tier, COI status, SAE framework, vulnerable-population caveats.
• irb_ruleset: "2026-05" stamped on every output for cache-bust correctness.

Sign-off ambiguities resolved (v1)

• A1 (multi-jurisdiction shape): US-only → review_tier_eu: null; EU-only → review_tier_us: null, expedited_categories_us: []. Both populated when both jurisdictions present.
• A2 (expedited_category_claim mismatch): Surface BOTH the investigator's claim and the tool's analysis in advisory_warnings — never silently override.
• A3 (risk_level: "unknown"): Conservative default — interventional + unknown risk → full-board + warning. Preserves user safety over user convenience.
• A4 (ICF complexity tier rule): complex if full-board OR any vulnerable; basic if minimal + non-interventional + no vulnerable; standard otherwise.

Tests

57 new tests (683 total). Each Common Rule exempt category 1-8 reachable; each expedited category 1-7 reachable; full-board path; Subpart B/C/D layered correctly; GDPR Art. 9 fires only on EU + non-anon; HIPAA §164.514 fires only on US + identifiable; CTR/FDA/PSUR SAE frameworks; PASS_imposed override; cover-letter content + word count; irb_ruleset stamp; <300ms perf; A1/A2/A3/A4 regression coverage.

v2 deferrals (committed in design log #21)

• UK HRA/REC IRAS pathway, Japan PMDA + ECRIN, Canada TCPS-2.
• Full per-jurisdiction pediatric Subpart D age-tier table (US state-by-state assent ages, EU member-state variations).
• Paired icf_readability_check tool — Flesch-Kincaid grading + medical-jargon detection on actual ICF text.
• IRB cover-letter PDF export via existing DOCX formatter.

v1.4.2 (2026-05-06) — code-review fixes for v1.3.2 / v1.4.0 / v1.4.1

Three releases (v1.3.2 NICE TA precedents + JCA scope eligibility, v1.4.0 hta_workflow orchestrator, v1.4.1 HFrEF per-country comparator depth) shipped to production without independent review. Three parallel code reviews surfaced 10 HIGH and 8 MEDIUM findings of regulatory consequence. The headline "CRITICAL" turned out to be a reviewer hallucination (TA773 → TA849 swap that would have introduced a real fabrication; verified via webfetch against nice.org.uk that TA773 is in fact correct for empagliflozin HFrEF). All real findings addressed.

Fixed (HIGH)

• JCA scope-eligibility date typo: "12 January 2025" → "13 January 2025" in src/jca/scopeEligibility.ts. Reg 2021/2282 Article 34 specifies 13 January 2025 as Phase 1 start; the wrong date was printing into refusal markdown that customers paste into dossiers.
• is_orphan + force_proceed_out_of_scope now in MCP inputSchema. Zod schema had the fields but the JSON Schema advertised to MCP clients did not — making the safeguard's recovery path invisible to the LLM. Now discoverable.
• extractTaNumber regex handles "TA 679" (space) format. Previously matched only "TA679" without space; missed common NICE prose patterns. New extractAllTaNumbers companion picks up multiple TA citations in one prose block.
• findPrecedents drug match is now token-set equality, not bidirectional substring. Bare "valsartan" (ARB) no longer matches "sacubitril valsartan" (ARNI) precedent — eliminates a class of false-positive TA mismatches as the precedents table grows.
• hta_workflow Phase 2 abstract preservation. screen_abstracts JSON output drops the abstract field; if those records were piped straight to risk_of_bias, RoB inference would silently run on empty abstracts and corrupt GRADE downstream. Phase 2 now extracts the included IDs from the screening output and re-maps them onto the original literature records (which still carry abstracts).
• hta_workflow JCA scope bypass warning. hta_body="jca" runs the standard pipeline without calling jca_pico_scope, so the JCA scope eligibility check (Reg 2021/2282 phased rollout) was bypassed silently. Now emits an explicit audit warning so an out-of-scope indication can't produce a credible-looking JCA dossier draft.
• hta_workflow summary-table honesty. Phase 5 row used to hardcode "NICE STA draft" regardless of whether the dossier phase succeeded. Now correctly reads "FAILED — see audit" when dossierRes.ok === false, eliminating a contradiction between the summary table and the body.
• HFrEF outcome priorities lead with the composite primary endpoint. DAPA-HF and EMPEROR-Reduced both used "CV death OR HF hospitalization" as a single co-primary composite — the prior order split the components and ranked all-cause mortality after them, which inverted the logical relationship and implied a hierarchy regulators reject.
• HFrEF instrument list now includes KCCQ-12. The Kansas City Cardiomyopathy Questionnaire is the disease-specific HRQoL instrument used in DAPA-HF / EMPEROR-Reduced and required by EUnetHTA Annex II for HFrEF; the prior instrumentsFor("cardiovascular_hfref") fell through to EQ-5D-5L only.
• HFrEF population_subgroups includes NYHA class, LVEF stratum, ARNI eligibility, eGFR tier, T2D status. Was falling through to generic ["age strata", "comorbidity status"] despite the country-specific comparator universes assuming these subgroups (especially the NL ARNI-eligible split).
• Bare "heart failure" indication ambiguity warning. When the user passes "heart failure" without an EF qualifier, classifyIndication routes to generic cardiovascular. The handler now emits an explicit advisory warning that HFpEF / HFmrEF / HFrEF have materially different comparator universes and prompts re-running with the specific phenotype.

Fixed (MEDIUM)

• hta_workflow idempotentHint: false (was true — wrong because Phase 4 PSA is stochastic and Phases 1+6 hit live external APIs).
• hta_workflow openWorldHint: true (was false — wrong because literature_search calls PubMed/CT/Cochrane/ICER and validate_links makes HTTP requests).
• hta_workflow phase_timings_ms.cost_effectiveness_model no longer set when skip_ce_model: true — was a small non-zero value that misled programmatic consumers checking the timing as a "did CE run?" proxy and caused intermittent test flakes.
• Comment fix in countryRegistry.ts UK branch: was attributing TA679 to empagliflozin (TA679 is dapagliflozin); now correctly cites TA773 (empagliflozin) and TA679 (dapagliflozin) separately.

Tests

• 626 MCP tests passing (was 609) — +17 behavioural regression tests covering each fix.

Process learning

The HFrEF reviewer's "CRITICAL" — claiming TA773 is ivosidenib for AML and that empagliflozin HFrEF is TA849 — was a confident regulatory hallucination. WebFetch against nice.org.uk confirmed: TA773 IS empagliflozin HFrEF (9 March 2022); TA849 is cabozantinib for HCC. Without verifying we would have introduced a real fabrication while "fixing" a false alarm. Memory note saved: always verify subagent regulatory ID claims against the official public database before editing code.

v1.3.1 (2026-05-05) — pv_classify + pv_signal_workflow code-review fixes

Independent code reviews of both PV tools surfaced 1 CRITICAL + 6 HIGH findings of regulatory consequence. All addressed before redeployment.

pv_signal_workflow fixes

• CRITICAL — Cross-field validation. Zod .refine() rules now reject case_counts that produce negative 2×2 cells (drug_event > event_total, drug_event > drug_total, or grand_total too small for the cells). Previously such inputs produced negative PRR/ROR and could fire refuted_signal for what is actually garbage input.
• HIGH — previously_known_signal ignored event identity. New optional reported_event input; verdict requires case-insensitive substring match between reported_event and one of prior_known_signals. A drug with prior_known_signals: ["lactic acidosis"] reporting a fresh signal for "myocardial infarction" is now correctly classified as a new signal rather than silently suppressed.
• HIGH — IC posterior variance was missing 2 of 4 marginal terms (Norén 2006 simplified form). Truncated formula gave IC025 ≈ 0.06 vs correct ≈ 1.07 for Evans 2001 vector — a 1.0-unit error that flipped IC threshold_met for borderline signals and systematically downgraded confirmed_signal to strengthening_signal.
• HIGH — Chi-squared was labelled "Yates" but not actually Yates-corrected. Now applies (|obs−exp| − 0.5)² / exp per the label. Misrepresentation of the statistic to regulators eliminated.
• MEDIUM — Tool description warns callers about MGPS single-stratum confounding (where it can inflate EBGM/EB05 for sex/age-stratified populations).
• LOW — Dead triggers ≥ 3 branch removed in decideVerdict.

pv_classify fixes

• HIGH — Fabricated ENCePP IDs replaced. ENCePP-PASS-001 style identifiers are not registered ENCePP templates. Field renamed from encepp_protocol_template to encepp_study_category with plain-language category labels (e.g., "PASS — post-authorisation safety study (imposed, GVP Module VIII). Use the ENCePP Code of Conduct checklist for protocol structure"). Markdown output explicitly notes the value is a category label, not a retrievable template reference.
• HIGH — ICH E2E rationale honesty. Pre-authorisation rationale now explicitly states ICH E2E is a standalone ICH guideline, not a GVP module. The Module V reference reflects the downstream RMP that the E2E plan informs at MAA submission, not a direct ICH E2E → GVP V mapping.
• HIGH — Conditional/accelerated approval Specific Obligations warning. When regulatory_context is conditional_approval or accelerated_approval AND imposed_by_authority=false, output emits an advisory warning prompting confirmation of CMA Article 14-a SOB status. Previously fell through silently to PASS_voluntary.
• MEDIUM — rmp_commitment + imposed_by_authority=true precedence reversed. Now classifies as PASS_imposed primary with RMP_Annex_4_study as alternative. Article 107n imposition outranks Annex 4 listing per EMA practice; the prior ordering routed the wrong GVP module + omitted the PRAC pre-review obligation.
• MEDIUM — spontaneous_reports + imposed_by_authority=true warning. Audit + markdown now flag this as a contradictory input (spontaneous reporting is an inherent obligation, not something an authority can impose as a study); the imposed_by_authority flag is no longer silently dropped.
• MEDIUM — CMS IRA legal claim softened. Removed the inaccurate claim that "IRA excludes pharmacovigilance cost data from Medicare drug-price negotiation calculations" (no statutory basis). New language: "PV study costs are typically tracked as regulatory obligations separate from HEOR cost-effectiveness modelling and are not standard inputs to the IRA Maximum Fair Price calculation under current CMS guidance."

Tests

• 577 MCP tests passing (was 558) — +19 behavioural tests covering all the review findings (cross-field validation cases, reported_event matching, full Norén IC variance against Evans 2001 vector, Yates chi² value, CMA SOB warning, ENCePP fabricated-ID guard, IRA-claim wording, rmp_commitment + imposed precedence, spontaneous_reports + imposed warning).

Why this is a patch release

Pure correctness + transparency fixes. No API breaking changes (the encepp_protocol_template field rename is a transparency improvement; the previous IDs were not real ENCePP references, so callers depending on them were depending on a fiction). No new tools.

v1.3.0 (2026-05-05) — pv_signal_workflow tool (EMA GVP Module IX rev 2)

Added

• pv_signal_workflow tool — given drug-AE case counts (from EudraVigilance / FAERS / national PV DB / internal spontaneous reports), computes four disproportionality statistics: PRR (Evans 2001), ROR (van Puijenbroek 2002), IC (Bate 1998 / Norén 2006 BCPNN posterior), and MGPS (DuMouchel 1999, EBGM with EB05/EB95 via gamma-Poisson shrinkage). Decides a signal verdict (no_signal / strengthening_signal / confirmed_signal / previously_known_signal / refuted_signal) and emits canonical RMP signal-section text. Pairs with pv_classify (planned-study classifier).
• GVP Considerations P.III pregnancy follow-up. When pregnancy_exposure: true AND rmp_has_pregnancy_concern: true, output includes structured follow-up timepoints (birth / 3 months / 12 months) per the actual P.III gating logic — not blanket-triggered for any pregnancy exposure.
• outcome_serious: true lowers PRR/ROR/MGPS thresholds (2.0 → 1.5) per accelerated-review convention for serious / fatal / life-threatening AEs.
• Multi-method signal corroboration. Per EMA + Maven 2026 guidance, signals confirmed by ≥2 of 4 methods (with N≥3 + χ²≥4) are classified as confirmed_signal. Single-method triggers are strengthening_signal. Matching prior_known_signals reclassifies as previously_known_signal so no spurious new RMP variations.
• 5th NEW landing-page card "PV Signal Detection" added to the web UI showcase grid (16 examples total).

Why this release

EMA GVP Module IX rev 2 (effective 2026) makes EVDAS integration mandatory for all EU MAHs from 12 February 2026, ending the EudraVigilance signal-detection pilot. EMA's accompanying message: "AI-powered pharmacovigilance is now expected, not optional." This tool absorbs the disproportionality-statistics + workflow-recommendation step into HEORAgent so PV teams stop maintaining ad-hoc Excel signal sheets.

Roadmap committed (not in v1.3.0)

• EVDAS programmatic access (eRMR / ICSR download per Reg. 2025/1466) — v2; v1 takes user-supplied case counts.
• Stratified MGPS (by sex / age band) — v2; v1 uses single-stratum gamma-Poisson shrinkage.
• hta_dossier integration — pipe active signals into the PV plan section — v3+.

Tests

• 558 MCP tests passing (was 535) — +23 pv_signal_workflow tests including math against Evans 2001 / Bate 1998 / DuMouchel 1999 published vectors, all 5 verdicts reachable, P.III gating correctness, threshold-tier behaviour, and <300ms performance.

References

• EMA GVP Module IX rev 2 — Signal management (2026)
• EU Implementing Regulation 2025/1466 — mandatory EVDAS integration
• EMA GVP Considerations P.III — Pregnant and breastfeeding women (effective 2026-02-09)
• Evans SJW et al. 2001 (PRR) · van Puijenbroek 2002 (ROR) · Bate 1998 + Norén 2006 (BCPNN/IC) · DuMouchel 1999 (MGPS/EBGM)

v1.2.2 (2026-05-05) — error telemetry + permissive input validation

Fixed

• Analytics instrumentation gap. Tool-call errors now emit structured error_class (Error subclass name — ZodError, TypeError, etc.) and error_message (truncated to 500 chars) properties to PostHog. Before this fix, every error event had error_class:"(none)" and error_message:"(no message)" because trackToolCall() call sites only attached a generic error field that the dashboards weren't querying. Future production errors are now diagnosable from telemetry alone. New classifyToolError() helper handles Error / TypeError / ZodError / non-Error thrown values uniformly.
• evidence.risk_of_bias 26% error rate fix. PostHog showed LLM clients frequently sent studies without title or abstract (both previously required). The tool already returned "Unclear" for any missing reporting signal — strict validation was adding zero methodological rigour and causing 1 in 4 calls to fail. Both fields now default to safe values (title: "(untitled study)", abstract: ""). Added an explicit wrapper-shape error: when caller passes a single study object instead of {studies:[...]}, the error message hints at the correct shape.
• models.cost_effectiveness 40% error rate fix. Switched from .parse() to .safeParse() with a structured field-path error format so LLM clients can self-correct on the next call. Added an explicit hint when caller flattens efficacy_delta to the top level instead of placing it inside clinical_inputs.

Tests

• 535 MCP tests passing (was 521) — +14 behavioural tests covering the three fixes (error classifier shape, risk_of_bias permissive input, cost_effectiveness error helpfulness).

Why this is a patch release

Pure telemetry + UX improvements. No API changes; no breaking changes; no new features.

v1.2.1 (2026-05-04) — jca_pico_scope code-review fixes

Fixed

• HIGH — Indication classifier overmatch. classifyIndication() matched any indication string containing the substring "uc" — silently routing mucositis, Duchenne muscular dystrophy, and glaucoma indications to IBD-UC biologic comparators (vedolizumab/infliximab/ustekinumab). Now uses a word-boundary regex (^|\s)uc(\s|$). Patient-safety-adjacent in a production JCA tool. 4 new behavioural tests covering the false-positive cases.
• HIGH — Dead CountryProfile.outcome_priority and .outcome_instrument_preferences fields. Set on every profile, never read by buildScope (which calls outcomePriorityForCategory directly). Future contributors adding country-specific overrides would see no effect. Both fields removed from the type and from every profile literal.
• MEDIUM — isOncology proxy check. Was checking outcome_priorities[0] === "OS" (correct only by coincidence). Now PicoMatrix carries indication_category explicitly and the surrogate-endpoint warning checks the category directly. Future non-oncology categories with OS-first priorities won't trigger the PFS/ORR warning incorrectly.
• MEDIUM — NSCLC line-of-therapy gap. Detailed EGFR-mutant comparators only fire for line_of_therapy="second_line"; other lines silently fell through to a generic chemotherapy placeholder. Now emits an audit warning AND a markdown ⚠️ block telling the user to re-run with second_line for the well-modeled case.
• MEDIUM — Heterogeneity threshold transparency. The ≥3-distinct-comparators rule is a tool-level assumption, not a published EUnetHTA threshold. Now stated explicitly in the tool description so LLMs and reviewers know it's a decision rule, not a diagnosis.
• MEDIUM — Round-trip integration test strengthened. Now asserts at least one comparator molecule from pico_matrix.picos appears in the hta_dossier output, not just the PICO IDs (which the dossier could mention for unrelated reasons).
• LOW — flattenComparators dead export removed.

Tests

• 521 MCP tests passing (was 514) — +7 behavioural tests covering all the review fixes.

Why this is a patch release

All v1.2.0 functionality is unchanged for correct inputs. Fixes only affect (a) edge-case indication strings that were silently misclassified, (b) dead-field traps for future contributors, (c) error/warning surfaces for previously silent failure modes. No API changes; no breaking changes.

v1.2.0 (2026-05-04) — EU JCA PICO matrix analyzer

Added

• jca_pico_scope tool — produces the canonical EU Joint Clinical Assessment (JCA) PICO matrix for a drug-indication pair across selected EU jurisdictions. v1 covers DE (G-BA / IQWiG), FR (HAS), IT (AIFA), ES (AEMPS / RedETS), NL (Zorginstituut), and UK (NICE, post-Brexit context). Other 22 EU member states return a "consult national HTA" placeholder. Returns a consolidated PICO list (per Reg. 2021/2282) plus per-country comparator universes, outcome instrument preferences, population subgroup focus, and a heterogeneity warning when ≥3 distinct comparators emerge across jurisdictions. Pipe pico_matrix.picos directly into hta_dossier({hta_body:"jca", picos: ...}). Pure decision logic, hardcoded country profiles, <300ms response.
• JCA_REVISION stamp — output includes jca_revision: "2026-05" for auditability. Bumped when EUnetHTA publishes new methodological guidance.
• Surrogate-endpoint flag — for oncology indications, output explicitly notes that PFS / ORR / biomarker response are accepted as secondary outcomes only and may face JCA scrutiny per Annex II of Implementing Reg. 2024/1381.
• Pre-authorisation anticipatory scope — when called with regulatory_context: "pre_authorisation", output is produced with explicit "anticipatory only, not for actual JCA submission" warning. Useful for protocol-design and pre-MA market access strategy.

Why now

EU JCA has been in force since 12 January 2025 for oncology / ATMPs. 2026 brings high-risk medical devices into scope; orphan drugs join in 2028; all medicines by 2030. Manufacturers have 100 days from the consolidated PICO list to dossier submission — and no tool to scope it. This tool absorbs the 3-week consultancy step into a 200ms call.

Tests

• 514 MCP tests passing (was 491) — +23 jca_pico_scope tests, including a round-trip integration test verifying pico_matrix.picos validates against hta_dossier({hta_body:"jca"}) without errors.

References

• Regulation (EU) 2021/2282 — HTA Regulation
• EU Implementing Regulation 2024/1381 — JCA procedural rules
• EUnetHTA Coordination Group — Methodological Guidance Series
• National HTA bodies: G-BA / IQWiG, HAS, AIFA, AEMPS / RedETS, Zorginstituut Nederland, NICE

v1.1.1 (2026-05-04) — NICE PMG36 update: severity modifier + health inequalities

Added

• NICE severity modifier (PMG36 §4.4) — hta_dossier now accepts severity_modifier: { absolute_qaly_shortfall, proportional_qaly_shortfall } and computes the QALY weight (1.0× / 1.2× / 1.7×) per NICE bands. Replaced the end-of-life modifier in April 2022 in opportunity-cost-neutral form. Output names the severity band (No modifier / Moderate / Severe) and renders an effective £/QALY threshold table (£20-30K → £24-36K → £34-51K).
• NICE health inequalities section (PMG36 May 2025 modular update) — hta_dossier now accepts health_inequalities: { affected_groups, baseline_disparity_evidence, intervention_impact, mitigation_plan }. Output explicitly flags interventions that widen disparity (⚠️) vs narrow (✅) vs neutral (⚪). When omitted on a NICE dossier, a one-line gap-flag note tells the reviewer what's missing.

Why now

NICE published a refreshed PMG36 manual on 31 March 2026 (covering devices/diagnostics/digital alongside medicines per the NHS 10-Year Plan). The May 2025 modular inequalities update is now part of every NICE submission. Both changes were under-reflected in our NICE STA template.

Tests

• 491 MCP tests passing (was 483) — +4 severity modifier tests + 4 health inequalities tests.

References

• NICE Health Technology Evaluations: the manual (PMG36, updated 2026-03-31)
• NICE methods modular update — Health Inequalities (May 2025)

v1.1.0 (2026-05-04) — Pharmacovigilance study classification + HTA dossier PV section

Added

• pv_classify tool — classifies a planned study into its EMA regulatory category (PASS imposed/voluntary, PAES, RMP Annex 4, DUS, active surveillance registry, pregnancy registry, spontaneous reporting, ICH E2E plan). Returns the matching GVP module (V/VI/VIII/VIII Addendum I), ENCePP protocol template ID, RMP implications, FDA analogue, and submission obligations. Pure decision-tree logic per EMA GVP rev 4, EU Regulation 1235/2010 Article 107a, and ICH E2E. Pregnancy populations override the primary verdict; pre-authorisation contexts never yield PASS. Returns in <200ms.
• hta_dossier PV Plan section — when pv_classification (the structured output of pv_classify) is passed to hta_dossier, the dossier output includes a Pharmacovigilance Plan section between RoB and CEA listing the GVP module, ENCePP template, submission obligations, and RMP implications. When omitted, a one-line "PV plan not provided" note flags the gap so reviewers see it.
• CMS IRA flag — when pv_classify is called with jurisdictions: ["us"], the output explicitly notes that CMS IRA price-negotiation calculations exclude PV cost data — track PV obligations in the regulatory budget, not the HEOR cost-effectiveness model.
• FDA mapping (v1 stub) — pv_classify includes an indicative FDA analogue per category (PMR, PMC, REMS, FAERS, Sentinel) with explicit "v1 stub, full FDA in v2" labelling. EMA remains the primary jurisdictional coverage.

Tests

• 483 MCP tests passing (was 453) — +26 pv_classify tests covering all 12 PvCategory leaves, hard rules (pre-auth never PASS, pregnancy override), GVP module mapping (every category resolves to exactly one module), output content (CMS IRA flag, FDA stub note), performance (<200ms) — and 4 hta_dossier tests covering the PV section integration.

References

• EMA Good Pharmacovigilance Practices (GVP) Module VIII — Post-Authorisation Safety Studies (rev 4)
• EMA GVP Module V — Risk Management Systems
• EMA GVP Module VIII Addendum I — Drug Utilisation Studies
• EU Regulation 1235/2010, Article 107a (imposed PASS)
• ICH E2E — Pharmacovigilance Planning
• ENCePP Code of Conduct + study protocol templates
• FDA REMS Guidance for Industry (2019); FDA Sentinel Initiative; 21 CFR 314.81

v1.0.6 (2026-05-04) — MAIC workflow orchestration tool

Added

• workflow.maic orchestration tool — runs the canonical MAIC discovery+screening pipeline in one MCP call: ITC feasibility + parallel literature_search (broad + per-trial) + PICO screen_abstracts + risk_of_bias + evidence_network. Returns a structured 9-section report with explicit Next Steps. Built because ChatGPT-5.3 cannot reliably chain 5+ tool calls in parallel; this absorbs the orchestration burden so the LLM only formulates the question. Stops short of running MAIC/Bucher itself — those still require IPD or trial-level effect estimates the search cannot supply. Phase failures degrade gracefully (one skipped phase doesn't abort the pipeline).

Tests

• 453 MCP tests passing (was 442) — +11 maic_workflow tests.

v1.0.5 (2026-05-04) — ChatGPT MAIC workflow recipe

Added

• maic_workflow_recipe example — examples({tool:"maic_workflow_recipe"}) returns a multi-step prompt template ChatGPT users can paste in sequence, plus a recommendation to use the web UI for one-shot depth. Includes trial-name suggestions by indication (UC: QUASAR/INSPIRE/U-ACHIEVE/TRUE NORTH; CD: ADVANCE/MOTIVATE; T2D: SUSTAIN/SURPASS; obesity: STEP/SURMOUNT; HF: PARADIGM/EMPEROR; oncology: KEYNOTE/CHECKMATE; etc.).

Tests

• 442 MCP tests passing — +4 examples tests for the new recipe.

v1.0.4 (2026-05-02) — Bucher consistency, GRADE upgrading, EQ-5D baseline-utility, ChatGPT support

Added

• Bucher consistency check — evidence_indirect now empirically tests Bucher's consistency assumption when direct head-to-head evidence is also in the network. Severity bands per Cochrane Ch. 11.4.3 / NICE DSU TSD 18: |z|<1.5 no conflict, 1.5–1.96 moderate (⚠️), ≥1.96 substantial (🚨), opposite-direction with both significant → substantial. Conflicts are surfaced in the markdown report and the consistency_check field on each IndirectEstimate.
• GRADE upgrading (Guyatt 2011) — observational evidence with strong indicators can be upgraded from Low. Three criteria via the new upgrading_per_outcome param on hta_dossier: large effect (RR <0.5/>2.0 → +1; <0.2/>5.0 → +2), dose-response gradient (+1), plausible confounding biasing toward null (+1). Capped at +2 steps. Skipped when starting certainty is High (RCTs).
• EQ-5D 5L baseline-utility-aware impact estimator. utility_value_set now accepts baseline_utility (0–1). Biz 2026 reports category-level medians but the magnitude depends strongly on cohort baseline utility — 5L compresses utilities most in the 0.6–0.9 range, so a drug for mild plaque psoriasis (~0.85) sees a much bigger ICER increase than one for severe HS (~0.45). Output explicitly labels the result as an extrapolation beyond Biz 2026.
• ChatGPT Custom GPT support. New OpenAPI 3.1 adapter at /api/openapi (web tier) lets you build a Custom GPT in ~5 minutes. One POST endpoint per tool at /api/v1/{tool_name} — same code path as the Anthropic surface, with ChatGPT-friendly caps (psa_iterations≤1000, runs≤1, max_results≤30) so calls fit the 45s Action timeout. Optional CHATGPT_ADAPTER_TOKEN for auth; built-in 60 req/min/IP rate limiter.
• Surface-tagged analytics. Every tool_call PostHog event now carries a surface property derived from clientInfo.name: claude_anthropic_web, chatgpt_adapter, claude_desktop, smithery, glama, pulsemcp, or direct_mcp. session_start events also include surface + client_name for acquisition reports.

Fixed (code review)

• assessInconsistency: when I² is unknown, return not_assessable (was Moderate with downgrade_steps=0, which silently inflated GRADE certainty).
• bucher.ts toWorkingScale: stripped dead se parameter that was a correctness trap for log-scale measures.
• eq5dImpact.ts: zero-median early return — future indication categories without published medians no longer produce degenerate {0,0,0} ranges.
• mcpSession.ts drift guard: changed module-load throw to a warn + lazy UnmappedToolError at call time. A single drift bug no longer crashes the entire web UI cold-start; only the affected tool fails.
• htaDossierPrep schema: replaced z.any() for rob_results / model_results / evidence_summary with proper Zod schemas.
• Adapter route: rate limit added (60 req/min/IP); available_tools 404 list now uses canonical 17-tool list (was 6); MCP_API_VERSION constant replaces hardcoded "1.0.3".

Tests

• 401 MCP tests / 96 web tests = 497 total passing (was 357 at v1.0.2).

References

Bucher HC et al. J Clin Epidemiol. 1997;50(6):683-691; Cochrane Handbook Ch. 11.4.3; NICE DSU TSD 18; Guyatt GH et al. J Clin Epidemiol. 2011;64(12):1311-1316; Biz, Hernández Alava, Wailoo (2026) Value in Health forthcoming.

v1.0.3 (2026-04-29) — Senior HEOR methodology fixes

Fixed

• GRADE inconsistency now uses I² instead of study count. Single-study comparisons no longer auto-downgraded as "Serious" — they return not_assessable (single study cannot be inconsistent with itself, per Cochrane Handbook 10.10). When I² is supplied via the new heterogeneity_per_outcome param on hta_dossier, GRADE applies Cochrane bands: <50% Low, 50–74% Moderate (1-step downgrade), 75–89% Serious, ≥90% Very Serious (2-step). Rationale cites the actual I² value.
• GRADE upgrading (Guyatt 2011) — observational evidence with strong indicators can now be upgraded from Low. Three criteria via the new upgrading_per_outcome param: large effect (RR <0.5/>2.0 → +1; <0.2/>5.0 → +2), dose-response gradient (+1), plausible confounding biasing toward null (+1). Capped at +2 steps. Skipped when starting certainty is High (RCTs).
• EQ-5D 3L→5L impact estimator now baseline-utility-aware. Biz 2026 reports category-level medians but the magnitude depends on cohort baseline utility — 5L compresses utilities most in the 0.6–0.9 range, so mild plaque psoriasis (baseline ~0.85) sees +77% ICER vs severe HS (baseline ~0.45) at +41%, even though both are non_cancer_qol_only. New baseline_utility param on utility_value_set tool.
• Bucher consistency check — when direct head-to-head A-vs-C evidence exists alongside the indirect A-vs-C estimate, the tool now empirically tests Bucher's consistency assumption: z = (direct − indirect) / SE_diff. Severity bands per Cochrane Ch. 11.4.3 / NICE DSU TSD 18: |z|<1.5 no conflict, 1.5–1.96 moderate (⚠️), ≥1.96 substantial (🚨), opposite-direction with both significant → substantial. Conflicts surfaced in markdown output and warnings.

Added

• New modules: src/grade/inconsistency.ts, src/grade/upgrading.ts, src/grade/eq5dImpact.ts, src/network/consistency.ts
• 41 new tests (4 new test files); total 385/385 passing.

References

Cochrane Handbook for Systematic Reviews of Interventions Ch. 10.10, 11.4.3; GRADE Handbook 5.1; Guyatt GH et al. J Clin Epidemiol. 2011;64(12):1311-1316; Higgins & Thompson Stat Med 2002; Bucher HC et al. J Clin Epidemiol. 1997;50(6):683-691; NICE DSU TSD 18; Biz, Hernández Alava, Wailoo (2026) Value in Health forthcoming.

v1.0.1 (2026-04-28) — Risk of Bias assessment tool

Added

• risk_of_bias tool (17th tool) — Cochrane RoB 2 (RCTs), ROBINS-I (observational), AMSTAR-2 (SRs). Auto-detects instrument from study type, infers domain judgments from abstract text, marks "Unclear" when evidence absent. Output includes per-study RoB table and rob_results object for evidence-based GRADE assessment in hta_dossier_prep.
• htaDossierPrep integration — rob_results parameter now replaces heuristic RoB judgments with structured domain assessments for GRADE tables.

Source

Implements design log 07 — based on Cochrane RoB 2 (Sterne et al. 2019), ROBINS-I (Sterne et al. 2016), AMSTAR-2 (Shea et al. 2017).

v0.9.8 (2026-04-22) — ITC methods, evLYG, CMS IRA context

Added

• Heterogeneity statistics in indirect_comparison NMA output — I² statistic, Cochran Q, degrees of freedom, p-value, τ², and interpretation band (Cochrane Handbook: 0–40% might not be important / 30–60% moderate / 50–90% substantial / 75–100% considerable).
• itc_feasibility tool (17th tool) — walks through the 3 ITC assumptions (exchangeability, homogeneity, consistency) and recommends a method (Bucher / NMA / anchored MAIC / unanchored MAIC / ML-NMR required / infeasible). Cites Cope 2014 (BMC Med), NICE DSU TSD 18 (Phillippo), Signorovitch 2023 (J Dermatol Treatment), Cochrane Handbook Ch 11.
• evLYG (Equal Value Life-Years Gained) as optional summary metric in cost_effectiveness_model — CMS IRA-compatible alternative to QALYs. Controlled via summary_metric parameter: "qaly" (default), "evlyg", or "both".
• System prompt updated with CMS IRA QALY prohibition (§1194(e)(2)) and AHA/ACC 2025 $120K/QALY threshold for cardiovascular interventions.

Security

• .gitignore hardening — added defense-in-depth block patterns for common confidential client filename markers.
• Provider comments sanitised — removed specific client references from enterprise fetcher comments (pharmapendium, citeline, cochrane, cortellis) and generalised to "institutional/enterprise proxy".
• Pre-commit hook installed (.git/hooks/pre-commit) that blocks commits containing confidential client name keywords.

v0.9.7 (2026-04-22) — UK EQ-5D-5L transition

Added

• utility_value_set tool (16th tool) — reference data and impact estimator for the new UK EQ-5D-5L value set (NICE consultation 2026-04-15 to 2026-05-13). Three actions:
  • lookup — full characteristics of UK 3L, England 5L, UK 5L (new 2026), or DSU mapping
  • compare — side-by-side comparison of all four value sets
  • estimate_impact — projects ICER/QALY change per Biz, Hernández Alava, Wailoo (2026) Value in Health (forthcoming).
• OHE and EuroQol data sources (43rd and 44th) — curated pointers to Office of Health Economics publications (ohe.org) and EuroQol Group resources (euroqol.org). Category: other. No API key required.
• htaDossierPrep UK 5L transition warning — when hta_body="nice", dossier draft now appends a "UK EQ-5D-5L Value Set Transition" section flagging consultation dates and Biz et al. 2026 impact estimates by indication type.
• cost_effectiveness_model description updated with value-set-dependency note pointing to utility_value_set.
• 15 new tests covering the utility_value_set tool; 6 for OHE + EuroQol fetchers.

Source

Implements design log 09 — based on public OHE / EuroQol materials + Biz, Hernández Alava, Wailoo (2026). Switching from EQ-5D-3L to EQ-5D-5L in England: the impact in NICE technology appraisals. Value in Health (forthcoming).

v0.9.6 (2026-04-19)

Added

• Wiley Online Library source (42nd data source) — CrossRef-based free access to Wiley HEOR journals: Pharmacoeconomics, Health Economics, Journal of Medical Economics, Value in Health. ~77% abstract coverage for recent articles (Wiley joined I4OA 2022). No API key required. Source aliases: pharmacoeconomics, health economics. Included in default source set.

v0.9.5 (2026-04-16)

Added

• risk_of_bias tool (15th tool) — structured risk of bias assessment using auto-detected Cochrane instruments: RoB 2 for RCTs (5 domains), ROBINS-I for observational studies (7 domains), AMSTAR-2 for systematic reviews (16 items). Instrument selected automatically from study_type; override with instrument param. Returns per-study domain judgments (Low / High / Unclear / Some concerns) plus a GRADE Risk of Bias summary object (rob_judgment, downgrade, rationale, overall_certainty_start).
• hta_dossier_prep GRADE integration — new rob_results parameter accepts output from risk_of_bias. When provided, the GRADE table uses the structured RoB judgment instead of the previous heuristic estimate. GRADE table note now indicates which source was used. Backward-compatible: falls back to heuristic when rob_results is omitted.
• System prompt pipeline rule — Claude now calls risk_of_bias after screen_abstracts and passes rob_results to hta_dossier_prep automatically in the standard HEOR workflow.
• 29 new tests covering risk_of_bias (23) and hta_dossier_prep rob_results integration (6). 289 tests total, 72 suites, all passing.

v0.9.4 (2026-04-16)

Added

• Parameter descriptions audited and filled for all tool schemas — perspective, clinical_inputs, cost_inputs, utility_inputs on cost_effectiveness_model; perspective on budget_impact_model; drug_name, indication, output_format, nested PICO fields on hta_dossier_prep; target.intervention/target.comparator on indirect_comparison. Improves Smithery parameter-descriptions score.

v0.9.3 (2026-04-16)

Fixed (from code review)

• BIM market share forward-fill — missing years now inherit from the most recent DEFINED year before them, not the last-defined-globally (which was inflating early-year budget impacts)
• BIM xlsx perspective crash — fixed TypeError when perspective was undefined in Excel export
• XLSX transition matrix — now derived from actual model params (efficacy_delta, mortality_reduction), no longer hardcoded placeholders
• XLSX "Mean ICER" label — renamed to "ICER of means (E[ΔC] / E[ΔQ])" to reflect the formula accurately; added separate "Mean of per-iteration ICERs" for the alternative interpretation
• HTTP JSON parser — now returns 400 with clear error instead of crashing on malformed request body
• HTA template hardcoded outcomes — "Outcomes (PICO)" section no longer defaults to HbA1c/diabetes regardless of indication
• Link validator 429/503 — now categorized as "rate_limited" (transient) instead of "broken"

Changed

• MAIC/STC descriptions — marked as EXPERIMENTAL with clear warnings that summary-level data produces approximate results only; true MAIC/STC per NICE DSU TSD 18 requires individual patient data
• Survival fitting description — marked as EXPERIMENTAL with warnings that KM-summary fits are approximate; true MLE requires IPD
• Excel export language — changed "editable, re-runnable" to honest "structured report — editing cells does not re-run the model"
• FEATURES.md — restructured into focused tables (was one mega-table that rendered badly on Glama); added "Production vs Experimental" section

Added

• 28 new smoke tests covering budget_impact_model, population_adjusted_comparison, survival_fitting, screen_abstracts, validate_links (72 suites, 272 tests total)

v0.9.1 (2026-04-16)

Added

• MCP tool annotations on all 14 tools (readOnlyHint, destructiveHint, idempotentHint, openWorldHint, title). Improves Smithery quality score and gives MCP clients clearer intent signals for tool use.

v0.9.0 (2026-04-16)

Added

• Excel (XLSX) export for budget_impact_model — multi-tab editable workbook (Summary, Inputs, Year-by-Year, Audit) so local market-access teams can localize pricing
• GVD (Global Value Dossier) template in hta_dossier_prep — new hta_body: "gvd" option with 13 sections (Disease Background, Unmet Need, Clinical Evidence, Comparative Effectiveness, Health Economic Summary, Policy Environment, etc.). Driven by Reddit feedback — GVDs are the upstream cross-market evidence document before country-specific dossiers.
• MCP prompts capability — 5 pre-built HEOR workflow prompts (literature-review, cost-effectiveness-analysis, hta-dossier, budget-impact, indirect-comparison) that appear as slash commands in Claude Desktop
• MCP resources capability — declares resources capability (empty list for now) to satisfy MCP clients

Fixed

• Smithery quality score issues: added resources/list and prompts/list handlers (previously returned "Method not found")

v0.8.0 (2026-04-16)

Added

• Excel (XLSX) export for cost_effectiveness_model — editable multi-tab workbook (Summary, Inputs, Transition Matrix, PSA, CEAC, Audit). Yellow cells mark editable inputs so local market-access teams can localize pricing/prevalence and re-run. Driven by Reddit feedback from an HEOR practitioner.
• Updated server-card.json to reflect all 14 current tools and v0.7.1+ metadata (was stale at v0.1.3)

v0.7.0 (2026-04-16)

Added

• validate_links tool — HTTP HEAD check for URLs before presenting them to users. Categorizes as working/browser_only/broken/timeout. Web UI system prompt now mandates validation of all citation URLs before they appear in responses.

v0.6.0 (2026-04-15)

Added

• screen_abstracts tool — PICO-based abstract screening with relevance scoring, study design classification (Cochrane Handbook Ch. 4), and ranked inclusion/exclusion decisions. Turns raw literature_search results into a screened shortlist with PRISMA flow summary.

v0.5.0 (2026-04-15)

Added

• survival_fitting tool — fit 5 parametric distributions (Exponential, Weibull, Log-logistic, Log-normal, Gompertz) to Kaplan-Meier data. AIC/BIC model selection, extrapolation table, clinical plausibility guidance per NICE DSU TSD 14 (Latimer 2013)
• EVPPI (Expected Value of Partial Perfect Information) — per-parameter VOI analysis in PSA output. Shows which specific parameters are worth further research, using non-parametric binning method (Strong et al. 2014)

v0.4.0 (2026-04-15)

Added

• budget_impact_model tool — ISPOR-compliant budget impact analysis with year-by-year net cost, market share uptake curves, treatment displacement, and population growth (Mauskopf 2007, Sullivan 2014)
• population_adjusted_comparison tool — MAIC (Matching-Adjusted Indirect Comparison) and STC (Simulated Treatment Comparison) for population-adjusted indirect comparisons. Follows NICE DSU TSD 18 (Phillippo 2016). Accepts summary-level statistics — no IPD required
• Scenario analysis on cost_effectiveness_model — new scenarios parameter runs multiple what-if variants in a single call with comparison table output
• GRADE evidence quality assessment on hta_dossier_prep — auto-generated GRADE table (Risk of Bias, Inconsistency, Indirectness, Imprecision, Publication Bias) when literature results are provided
• docs/FEATURES.md — comprehensive feature reference with Feature Name, What, Why, How for all 11 tools

Fixed

• Markov model Dead state — 3-state model (On-Treatment/Off-Treatment/Dead) replaces 2-state model. Absorbing Dead state prevents infinite QALY/LY accumulation
• ICER sign handling — wtpVerdict now correctly distinguishes dominant (lower cost + higher QALY) from dominated (higher cost + lower QALY) using delta signs
• Parallel source fetching — literature_search uses Promise.all instead of sequential loop (major performance improvement with multiple sources)
• DOMPurify security — web UI switches from incomplete FORBID_ATTR blocklist to ALLOWED_ATTR allowlist for SVG sanitization
• MCP server security — bearer token auth (MCP_AUTH_TOKEN), CORS origin restrictions (MCP_CORS_ORIGINS), session limits (max 100, 30min TTL)
• EVPI calculation — uses perspective-appropriate WTP threshold instead of hardcoded $50,000
• knowledge_write validation — Zod schema enforces wiki/ prefix and .md suffix at validation layer
• JSON-RPC ID collisions — web UI uses incrementing counter instead of Date.now()
• Duplicate getTimeHorizonYears function consolidated into modelUtils.ts
• Stale "7 tools" comments updated throughout

v0.3.0 (2026-04-14)

Added

• indirect_comparison tool — Bucher method (single common comparator) and frequentist NMA (full network) for indirect treatment comparisons. Supports MD, OR, RR, HR. Auto-selects method based on network structure
• Stability search — literature_search runs parameter (1-5) performs multiple search runs, deduplicates, and ranks by consistency

v0.2.0 (2026-04-14)

Added

• evidence_network tool — analyzes literature search results to build an evidence network map and assess NMA (network meta-analysis) feasibility. Extracts intervention-comparator pairs, builds treatment comparison graph, identifies evidence gaps
• PostHog analytics — anonymous tool call tracking (tool name, duration, status). No user data collected. Opt-in via POSTHOG_API_KEY env var
• Privacy policy and Terms of Service — required for ChatGPT app directory submission

Fixed

• NICE WTP thresholds updated from £20-30K to £25-35K/QALY (effective April 2026)
• CADTH renamed to CDA-AMC — all references, descriptions, and URLs updated from cadth.ca to cda-amc.ca (renamed May 2024)
• IQWiG General Methods updated from v7.0 to v8.0 (2025)
• ICER VAF label corrected to "2023-2026"
• TLV (Sweden) threshold description updated to severity-tiered system (SEK 250K-1M)
• PBAC (Australia) threshold corrected to ~AUD 50K (no formal threshold)
• Version now read from package.json at runtime instead of hardcoded

v0.1.4 (2026-04-12)

Added

• HTTP transport — server supports both stdio (default) and Streamable HTTP (for hosted deployment and Smithery registry)
• Endpoints: POST/GET/DELETE /mcp, GET /health, GET /.well-known/mcp/server-card.json
• Smithery listing — smithery.yaml for MCP marketplace, server-card.json for discovery
• Railway deployment — hosted at heor-agent-mcp-production.up.railway.app

v0.1.2 (2026-04-10)

Added

• DOCX save-to-disk — output_format="docx" now writes Word documents to ~/.heor-agent/reports/ (or project reports/ dir) and returns the file path instead of inlining base64
• ScienceDirect as 41st data source (uses ELSEVIER_API_KEY, same as Embase)
• Source selection table — every literature_search output includes a transparency table showing all 41 sources with used/not-used and reason

Changed

• README fully rewritten to reflect current capabilities (41 sources, 7 tools, all HTA bodies)

v0.1.0 (2026-04-06)

Added

• literature_search — parallel search across 39 data sources with PRISMA-style audit trail
  • Biomedical: PubMed, ClinicalTrials.gov, bioRxiv/medRxiv, ChEMBL
  • Epidemiology: WHO GHO, World Bank, OECD Health, IHME GBD, All of Us
  • FDA: Orange Book, Purple Book
  • HTA appraisals: NICE TAs, CADTH, ICER, PBAC, G-BA, HAS, IQWiG, AIFA, TLV, INESSS
  • HTA cost references: CMS NADAC, PSSRU, NHS Costs, BNF, PBS Schedule
  • Enterprise: Embase, Cochrane, Citeline, Pharmapendium, Cortellis, Google Scholar
  • LATAM: DATASUS, CONITEC, ANVISA, PAHO, IETS, FONASA
  • APAC: HITAP
  • Other: ISPOR
• cost_effectiveness_model — Markov / PartSA / decision tree models
  • PSA (Monte Carlo, 1K-10K iterations), OWSA (tornado), CEAC, EVPI
  • NICE reference case (3.5% discount), US payer, societal perspectives
  • WTP assessment against NHS (£25-35K), US ($100-150K), societal thresholds
• hta_dossier_prep — draft submissions for NICE STA, EMA, FDA, IQWiG, HAS, EU JCA
  • PICO framework, evidence summary, gap analysis
  • EU JCA support with per-PICO sections (Reg. 2021/2282)
• project_create — persistent project workspaces at ~/.heor-agent/projects/
• knowledge_search / knowledge_read / knowledge_write — project knowledge base with wiki support
• Metabolic profile analysis — auto-extracted from literature search results
• Text, JSON, and DOCX output formats
• Full audit trail (sources queried, inclusions, exclusions, assumptions, warnings)
• Localhost proxy support for enterprise APIs behind corporate VPN

See also Privacy · AI Transparency · Source.