Changelog

Unreleased

• Keep the synthetic cookbook's plain-text numeric context on complete lines for Agenda 1.16 compatibility, preserving every word and fail-closed checks.

• Preserve quantitative scenario provenance, binding constraints and ranking reversals in EN/RU skills; document external-model handoffs and separate MCP environments without adding model-execution claims.

• Restore training-mode switching and ACH requirements in the experimental compact candidate.

• Add fourteen A–G manual development cases for a no-API review; no model result or adoption claim.

• Make per-claim verification flag serialization explicit in EN/RU and Codex instructions.

• Exercise the supported Agenda 1.x range without requiring one exact patch release.

• Add opt-in compact-skill ablation and separate development cases; preserve frozen holdouts.

• Added a self-contained, developer-facing two-case portfolio cookbook. It builds source-backed review bundles in staging, requires strict verification and bundle-integrity PASS, atomically promotes the release, and emits deterministic ZIP distributions as CI artifacts.

1.8.0 - 2026-09-13

Stable promotion of 1.8.0rc1. The tagged candidate passed the full GitHub build gate and an independent installed-wheel burn-in covering a complete packet, an intact REVIEW REQUIRED packet, six corruption classes, and the missing-path operational error. No runtime or contract semantics changed during promotion.

• Added gtta review and the gtta.review-bundle@1.0 Python interface. A single strict-by-default run now atomically writes the rendered memo, JSON receipt, Markdown and HTML evidence reviews, SARIF, bounded repair guidance, and a hash-bearing manifest without overwriting an earlier review directory.
• Added the Agenda-independent gtta check-review-bundle command and gtta.review-bundle-check@1.0 report. It rejects incomplete or unsafe file sets, recomputes output hashes, validates receipt and SARIF structure, and checks status consistency with stable GTTAB001–GTTAB013 rule families.

1.7.0 - 2026-09-13

Stable promotion of 1.7.0rc1. The release candidate passed the full GitHub build gate; no runtime or contract semantics changed during promotion.

• Added the optional gtta.memo-verification@1.0 composition seam and separate gtta.sources@1.0 source catalog. gtta verify memo.json --strict auto-discovers memo.sources.json, projects source-backed MemoArtifact claims into Agenda Intelligence MD 1.9.x, preserves declared quotes, filters unused sources, emits Markdown/JSON/HTML review output, and fails closed without changing the frozen gtta.memo@1.0 schema or claiming factual or legal verification.
• Added SARIF 2.1.0 output to gtta check-contract, preserving stable rule IDs, severities, line locations, truncation metadata, and existing exit semantics.
• Added a deterministic verification-marker preflight for sanctions, regulatory, legal, date-sensitive, and quantitative source-backed claims; strict verification rejects missing verify: true unless a dated primary basis is declared.
• Froze and packaged the Draft 2020-12 JSON Schema for gtta.sources@1.0; tests reject drift between the shipped contract and its runtime model.
• Added gtta verify --repair-prompt repair.md, which emits bounded, claim-specific repair guidance without source text, input mutation, source invention, or a claim that the resulting memo is factually correct.
• Added a dedicated CI job against the real public agenda-intelligence-md==1.9.0 package and the canonical offline cookbook.
• Added native gtta verify --format sarif output with stable GTTA/Agenda rule mapping, severity, packet metadata, and claim-level MemoArtifact locations.
• Added a portable six-case verification benchmark: two golden variants, four expected negative controls, bounded-repair safety assertions, input hashes, JSON receipts, and SARIF regression output. CI uploads both the artifact and Code Scanning result; expected negative findings remain in the JSON receipt.

1.6.0 - 2026-09-07

First stable release since 1.4.0. Supersedes the 1.6.0rc1-rc3 candidates, which stayed GitHub prereleases while PyPI publication was blocked.

• Expanded SKILL_RU.md from a disclosed 9/45-section subset to a full 45/45-section Russian rendering of the canonical method, including Modes D, F, and G, the complete workflow, stop conditions, evidence-packet handoff, default output template, and self-check. Canonical English heading markers remain visible beside Russian labels for deterministic checker compatibility; structural parity and essential runtime markers now have a release check.

• Made gtta.agent importable without the optional LangChain/LangGraph stack. Agent dependencies now load only when AnalystAgent is constructed, with a focused installation error at that seam. The release gate again tests the minimal test,mcp environment, backed by a subprocess regression test that explicitly blocks the optional imports.

1.6.0rc2 - 2026-09-05

• Corrected the release workflow's test environment to install the optional LangChain adapter required by the full test suite. The v1.6.0rc1 GitHub prerelease exposed the mismatch before distributions were built; no package was published to PyPI and no paid API is called by the corrected gate.

1.6.0rc1 - 2026-09-05

• Published the first preregistered gtta-agent-eval-holdout@1.0.0 run using Antigravity 2.12.2 and Gemini 3.8 Flash (High). Both arms passed 10/10 strict structural checks and 0/10 complete declared-behavior expectations. Missing verification flags dominated. The null result does not reproduce the original suite's positive combined-pass difference; immutable artifacts, commitment verification, and execution caveats are retained.

• Preregistered gtta-agent-eval-holdout@1.0.0, a 10-case / 20-sample broader-domain structured declared-behavior suite that adds Mode F coverage. The public commitment binds the exact private cases, expectations, requests, skill, schema, and seed before generation. prepare-artifact-behavior now accepts a suite version and emits the commitment without exposing scorer expectations to model requests.

• Published the first cross-model-family declared-behavior replication using Claude Code 2.1.246 and Claude Opus 4.6. Structural passes were 11/12 skill versus 12/12 baseline; declared-behavior passes were 3/12 versus 0/12. This repeats the positive direction but not Gemini's magnitude and shows weak absolute adoption plus one source-reference invariant failure. Freshness passed against all seven prior runs; no semantic-quality claim is made.

• Published the first gtta-artifact-behavior-eval@1.0.0 run. Both arms passed 12/12 structural checks; the skill arm passed 8/12 frozen declared- behavior expectations versus 3/12 for baseline, an observed +41.7 point difference in this run. Freshness passed against both earlier structured Gemini runs. The result measures declared method-contract adoption only, not truth, reasoning quality, source support, causality, or usefulness.

• Added gtta-artifact-behavior-eval@1.0.0, a preregistered structured suite that scores case-specific declared claim kinds, provenance, reasoning links, options, indicators, verification flags, unknowns, and change conditions only after schema validation. Structural, behavioral, and combined pass rates remain separate; the protocol explicitly does not score semantic or practitioner quality.

• Published the fresh gtta-artifact-eval@1.1.0 Gemini replication. Both arms passed 12/12 strict artifacts with no findings, and freshness passed against the negative 1.0.0 run. The result confirms the narrow section-key/claim-axis adapter correction in this run but shows a schema-conformance ceiling: it is not a positive skill delta or an analytical-quality result.

• Published the first gtta-artifact-eval@1.0.0 run without suppressing its negative result: baseline passed 10/12 strict artifacts and the skill arm 1/12. Ten skill failures used human-facing headings instead of exact section keys and one confused inference provenance with ClaimKind. Version 1.1.0 now gives both arms the exact Mode-to-key map and separates the two claim axes. The schema and frozen outputs remain unchanged; a fresh replication is required before claiming that the correction works.

• Added the offline paired MemoArtifact path over the frozen 12-case suite. Both arms receive the same JSON Schema and output invariants; only the skill arm receives GTTA. The scorer requires valid gtta.memo@1.0, exact expected Mode and evidence mode, records prompt/scorer schema hashes, and labels artifact counts as descriptive rather than quality measures.

• Added explicit finding-cap telemetry in gtta-method-contract@1.2.3 and the paired-eval report. JSON now identifies truncated rules and per-arm truncated samples. Frozen-output rescores show no truncation in either Gemini run, but 12 baseline / 2 skill samples in the first Claude run and 11 baseline / 3 skill samples in the fresh replication; their raw warning totals are lower bounds.

• Published a fresh Claude Opus 4.6 Antigravity replication at seed 20260902 after tightening the runtime instructions. The skill arm again passed 12/12 versus 0/12 for the baseline and stored 111 warnings versus 359. Freshness passed against all three earlier runs. Three skill samples reached the GTTA010 cap, so the reduction from the prior same-ruleset count of 170 is disclosed as directional rather than causal or complete.

• Refined gtta-method-contract@1.2.2 to exempt the exact mandatory evidence-access disclosure from claim warnings and to distinguish a quoted, explicitly rejected generic-advice phrase from an actual recommendation. Frozen-output rescores remain separate from original reports: the Claude skill count moves from 181 capped warnings to 170, while passes and errors are unchanged. Runtime instructions now state that descriptive labels such as [assumption] cannot replace a canonical Axis A tag and require a visible Mode C trigger section.

• Published a freshness-gated cross-model Antigravity run with Claude Opus 4.6 (Thinking), seed 20260901. The skill arm repeated the 12/12 contract pass result versus 0/12 for the generic baseline, but produced 181 capped warnings versus 360. Exact artifacts, metadata correction, hashes, deterministic score, two-run freshness comparison, regression test, and warning triage are committed. The result strengthens structural M3 evidence while leaving external usefulness at U0.

• Refined gtta-method-contract@1.2.1 to accept Evidence Evaluation Matrix as the Mode G evidence-matrix marker. A frozen-output rescore removes one GTTA008 false positive from the seed-20260831 replication while retaining 12 actionable provenance warnings. The runtime method now explicitly covers colon-ended recommendation lead-ins and red-team premise cells.

• Published a freshness-gated same-model Antigravity replication at seed 20260831. The updated skill arm again passed 12/12 contract checks versus 0/12 for the generic baseline, with 13 versus 254 warnings. Exact inputs, outputs, metadata, hashes, score, freshness report, and limitations are committed; the result remains structural M3 evidence and does not change U0.

• Added a replication freshness gate to the offline paired-eval harness. It rejects exact output reuse and high-confidence cosmetic rewrites using disclosed whole-text and shared-line thresholds; passing remains a heuristic check, not proof of independent generation.

• Refined GTTA010 and mode-marker matching in gtta-method-contract@1.2.0: layout metadata, direct questions, table headers, and identifier columns no longer create claim warnings; genuine untagged table cells now report their column. A frozen-output rescore reduces skill-arm warnings from 72 to 25 while preserving the 12/12 vs. 0/12 pass result and the immutable original 1.1.0 report. Canonical MemoArtifact rendering now labels Decision and Change condition layout fields explicitly.

• Published the first controlled paired structural run: 12 cases and 24 opaque samples generated in Antigravity with Gemini 3.7 Flash (High), then scored by gtta-method-contract@1.1.0. The skill arm passed 12/12 and the generic baseline 0/12; exact inputs, outputs, settings, hashes, report, and strict limitations are committed. Method evidence is now M3; external usefulness remains U0.

• Added a zero-paid-API Antigravity eval path: prepare emits opaque Markdown tasks, import-antigravity validates saved responses and records run hashes, model/app metadata, settings, skill hash, and seed before deterministic scoring. The harness never calls a model API.

• Rewrote the README around the stable reasoning and executable-contract surfaces, moved optional agent/API/UI experiments out of the primary positioning, and aligned package and GitHub discovery metadata.

• Started 1.6.0.dev0 development with the versioned gtta.memo@1.0 MemoArtifact: a strict per-claim provenance ledger, cross-reference and mode validation, canonical Markdown rendering, and shared Python, CLI, and MCP interfaces.

• Added GTTA010 to gtta-method-contract@1.1.0 as a warning for likely untagged Markdown prose and table-cell claims. Exact claim accounting remains the structured artifact's responsibility.

• Replaced the ambiguous single maturity bar with independent release (R), method-evidence (M), and external-usefulness (U) axes. Current evidence is recorded as R2 / M3 / U0; unavailable practitioner access no longer blocks an honest software release, but still blocks usefulness claims.

• Promoted the deterministic checker interface to gtta-method-contract@1.0.0 with explicit compatibility criteria.

• Added a 12-case paired baseline/skill agent-eval harness; it reports structural contract findings only and makes no factual-quality claim.

• Added method-contract enforcement to the canonical example gate.

• Migrated four legacy examples that lacked any provenance marker to explicit provenance, calibrated confidence, and declared memo structure.

• Added a two-job PyPI Trusted Publishing workflow that builds and verifies distributions without elevated permissions before a separate OIDC publish.

• Prepared the 1.5.0rc1 distribution-integrity candidate. The complete English and Russian analytical method now ships as package data and loads through a single fail-closed resource interface used by framework and MCP adapters.

• Added a built-wheel installation smoke test and a repository check that keeps packaged runtime resources synchronized with the canonical root skill files.

• Moved MCP to its supported 2.x interface, made it an optional extra, removed the silent no-op fallback, and added a gtta mcp stdio launcher.

• Raised the Python package floor to 3.10 to match MCP 2.x and moved LangChain, LlamaIndex, and MCP dependencies out of the base installation.

• Removed the unavailable PyPI installation path from README. The developer toolkit is documented as a source-installed pre-release until it is actually published.

• Replaced the misleading gtta ingest interface with gtta parse-pdf, which reports parsing only, and removed the repository-only dark-factory worker from the installed CLI.

• Added Docker build-context exclusions for local environments and aligned the security scope with the repository's experimental runtime surfaces.

• Added the initial gtta-method-contract@0.1.0 deterministic preflight with stable rule IDs, severity, JSON/text reports, CLI exit semantics, and an MCP tool. Its documented scope is method conformance only; Agenda Intelligence MD continues to own evidence-packet checks.

• Corrected the PDF command so it reports parsing only instead of claiming a simulated FAISS ingestion, and removed remaining autonomous/fleet wording from the legacy review-queue worker and local batch UI.

• API and UI bind to loopback by default; protected API routes now fail closed when GTTA_API_KEY is missing, and external CLI binds require a key.

• Replaced shell-based CLI launchers with argument-safe subprocess calls.

• Memo results now expose the critic's actual validation_passed state and retain the critique when the iteration cap is reached.

• Reframed the legacy signal worker as a review-queue draft generator; it no longer claims automatic guardrail approval or publication.

• Replaced invented revenue and gross-margin figures with explicitly labelled heuristic cost estimates.

• Renamed the small static regional registry as illustrative context and added primary-source starting points and freshness warnings.

• Removed unsupported model choices and documented the API/batch surfaces as local, in-process experiments rather than production infrastructure.

• fix(docs,evals,signals): restore the honest-scope layer removed in 592867a. That commit stripped the Disclaimer section from the README and from every published signal, removed the Limitations and "what this skill has not been tested on" sections, took the limitation-note requirement out of templates/memo-blank.md, and cut the honest-scope observations from three agent-eval cases, evals/failure-modes.md, a self-run, and two worked examples. The method itself never changed: it still reports evidence mode, still separates fact from assumption, and still routes to human review. The [2.0.0] Autonomous Compliance Edition CHANGELOG block, which announced definitive legal, compliance, and sanctions determinations, is removed with it — the skill does not do that, and AGENTS.md forbids claiming it.

1.4.0 - 2026-08-25

• Removed the weekly cron from .github/workflows/policy-risk-signal.yml. The signal archive is maintained by hand: every signal in it was written and merged by a person, and the schedule never produced one. Of its 17 runs, four failed in May and the following thirteen reported success while doing nothing, because the branch that handles a missing OPENAI_API_KEY exited zero. A dead channel sat behind thirteen consecutive green checks. The workflow remains available as a manual workflow_dispatch draft.

• A dispatch without OPENAI_API_KEY now fails with an error instead of reporting success, and its pull-request body points at the manual route through signals/TEMPLATE.md.

• signals/README.md now states that the archive is maintained by hand on no fixed cadence, so a reader does not infer a schedule from the dated files.

• scripts/validate_signals.py now checks signals/README.md against the ARCHIVE_HEADER and ARCHIVE_FOOTER constants that rewrite it. The pair had already drifted once, silently reverting a documented convention on every run; the guard replaces remembering to edit both.

• Reconciled the plugin manifests with the release history. plugin.json and .claude-plugin/plugin.json were added during the 1.3.0 development cycle and both declared 1.0.0, three releases behind CHANGELOG.md. Both now declare the released repository version.

• scripts/validate_skill_package.py now checks each manifest against the newest dated release heading in CHANGELOG.md, not only against the other manifest. Agreeing with each other was what let the pair sit at 1.0.0 unnoticed. ## Unreleased is skipped: the manifests describe what is published, so ordinary merges do not force a bump.

• Added scripts/test_signal_pipeline.py and wired it into scripts/check.py: the generator's index, feed, and latest-signal output is now run through scripts/validate_signals.py on a throwaway tree, so a title or filename regression fails locally instead of surfacing in an automated draft pull request. Both fixed defects reproduce as failing tests against the previous behaviour.

• scripts/check_markdown_links.py no longer inspects link syntax inside fenced code blocks, where a link is an illustration rather than a repository target. No currently tracked link is affected; the change removes a false-failure trap for future documentation.

• scripts/validate_examples.py now walks examples/ recursively instead of its top level only, and accepts the **Evidence mode**: form alongside **Evidence mode:**.

• Fixed scripts/generate_policy_risk_signal.py recording an index title that scripts/validate_signals.py then rejects. The title was normalized through strip_text() and truncated, and when the <!-- title: ... --> marker was absent it reconstructed a sentence from the ## Signal paragraph. The validator requires the recorded title to appear verbatim in the signal markdown, so both paths could open a pull request that fails CI. The marker is now required, and the verbatim invariant is checked at generation time.

• Fixed the same script silently overwriting an existing signal when a run landed on a date that already had one. It always wrote signals/YYYY/YYYY-MM-DD.md, although the archive documents YYYY-MM-DD-<topic>.md for shared dates; it now derives a topic slug from the title.

• Fixed the archive footer written by the generator, which dropped the documented same-day filename convention from signals/README.md on every automated run.

• Added generation-time checks for the Date: YYYY-MM-DD header line and the --date argument format, both of which scripts/validate_signals.py only enforces after the file is written.

• Fixed unhandled network errors in the OpenAI request path, which raised a traceback instead of a message.

• Removed the unused README_PATH constant and corrected the weekly workflow, which diffed README.md and asked reviewers to check a README signal index that the generator never writes.

• Fixed scripts/validate_codex_sync.py silently ignoring a single path argument: len(sys.argv) > 2 guarded the read of sys.argv[1].

• scripts/validate_signals.py now fails with a message instead of a traceback when signals/latest.md is missing or the JSON payloads hold unexpected types.

• Synced signals/TEMPLATE.md with the archive. It lacked the <!-- title: ... --> marker that every archived signal carries and that the indexes are built from, used a ## Event heading no recent signal uses, and documented only the single-signal-per-day filename.

• Corrected the README claim that every live-source-backed example was retrieved on 2026-05-08; the Middle Corridor example states 2026-05-15.

• Relaxed the agenda-intelligence-md install pin in the integration doc from ==1.3.0 to >=1.3.0 and normalized the repository URL casing. The repository's synthetic packet lints clean under 1.6.0.

• Fixed the packaged skill identifier so SKILL.md uses the lowercase, hyphenated global-think-tank-analyst name exposed by its discovery directory.

• Added scripts/check.py as the single local/CI check interface and a narrowly scoped package validator for discovery fields, the canonical skill symlink, and synchronized plugin manifests.

• Synchronized public documentation with Mode G (Competing Hypotheses) and distinguished development-time repository checks from memo/output validation.

• Fixed two dead links to the author's site. deal-risk-gate.html and for-analysts.html have returned 404 since the site was restructured; the README also claimed two interactive browser demos that are no longer published. The demo sentence now points at examples/, and the contact section links the live case-study page.

• Extended scripts/check_markdown_links.py to fail on 404/410 for links to the author's own site, which previously went unchecked because every http(s):// target was skipped. Network errors and other statuses are reported without failing, so an offline run stays deterministic; SKIP_SITE_LINK_CHECK=1 skips the network step.

• Conformed the repo to the Agent Plugins 1.0.0 layout: added a root plugin.json with the $schema identifier from https://agent-plugins.org. .claude-plugin/plugin.json is unchanged and still serves the Claude Code install path; the specification ignores that directory. skills/global-think-tank-analyst/SKILL.md already matched the spec's discovery rule. The manifest validates against the published Draft 2020-12 schema.

• Added the Agenda Intelligence v1.3 evidence-packet handoff as the primary composition seam, with a runnable synthetic packet, dependency-free CI validator, and ADR 0001. Reclassified older memo-schema, scoring, analyze, and MCP recipes as compatibility workflows.

• Packaged the repo as an installable Claude Code plugin: added .claude-plugin/plugin.json and skills/global-think-tank-analyst/SKILL.md (a symlink to the canonical root SKILL.md; packaging only, not a runtime overlay). Installable via /plugin marketplace add vassiliylakhonin/agenda-intelligence-md, then /plugin install global-think-tank-analyst@agenda-intelligence. Verified locally with claude plugin install from an isolated config (1 skill discovered via the symlink).

• Clarified the commercial role: this repo remains the horizontal reasoning-method dependency behind Agenda Intelligence MD, not a buyer-facing product surface. Added README/AGENTS guidance to route product positioning and evidence-readiness commercialization through Agenda Intelligence MD.

• Added scripts/validate_codex_sync.py and wired it into CI: fails the build when the shared analytical contract in codex/SKILL.md drifts from the canonical root SKILL.md. Intentional codex-only / divergent sections are allowlisted in the script. Enforces the manual sync discipline declared in codex/SKILL.md's Contract provenance section.

• Resynced codex/SKILL.md shared analytical contract with the canonical root SKILL.md: restored the Stop and request — explicit triggers subsection, restored two dropped self-check items (provenance-language match, conflict-surfacing), and reverted three paraphrased contract paragraphs (retrieved-content trust, conflict handling, linguistic faithfulness) to the canonical root wording. Intentional codex-only sections (Contract provenance, Codex Platform Setup, JSON Output Mode, Pipeline Integration, the Mode F agentic-loop note, and Installation) are unchanged.

• Spawned three signals from existing live-source-backed memos: Hormuz oil-price corridor, EU AI Act simplification redeployment, China critical-minerals procurement window. Hardened the signal-generation script to support YYYY-MM-DD-<topic>.md filenames so multiple signals can share a date.

• Added two more live-source-backed worked examples: ECB rate-hold (2026-04-30, ECB primary publications) and EU CBAM enforcement-phase exposure (2026 definitive phase, DG TAXUD).

• Converted the README example list to a domain × evidence-mode matrix for navigation.

• Added ECB and BIS feeds to .github/policy-risk-signal/sources.json to broaden monetary-policy and financial-stability coverage in the auto-signal pipeline.

1.3.0 - 2026-05-09

• Added docs/integrations/agenda-intelligence-md.md — concrete CLI / MCP recipes for composing this skill with Agenda Intelligence MD (markdown score, JSON brief validate+score, evidence-pack audit, MCP loop).
• Added docs/integrations/agenda-intelligence-md-live-demo.md — end-to-end run with real CLI output (95/100 brief-only → 83/100 with an honest evidence pack), captured against the published Agenda Intelligence MD package on 2026-05-08.
• Added a live-source-backed worked example: examples/live-source-backed-memo.md (memo on the May 1, 2026 OFAC "Operation Economic Fury" Iran shadow-banking action), paired with a real JSON brief projection and evidence pack under examples/agenda-projections/.
• Added three illustrative memos: export-controls exposure, critical-minerals supply-risk, EU energy-transition policy.
• Added a new signal (signals/2026/2026-05-08.md) explaining why an honest evidence pack lowers the score.
• Expanded evals/failure-modes.md with five new patterns specific to the GTTA + Agenda Intelligence MD composition (score gaming via stripped unsupported claims, JSON projection content loss, mode mismatch on raw markdown scoring, live_source_backed claim without an evidence pack, stale retrieved_at).
• Added three additional live-source-backed worked examples (real sources retrieved 2026-05-08): Hormuz disruption / energy prices, EU AI Act simplification (Omnibus VII), and the China critical-minerals export-control suspension.
• Trimmed the README Roadmap to reflect what was delivered.

1.2.0 - 2026-05-08

• Repositioned README, AGENTS.md, llms.txt, SKILL.md, and codex/SKILL.md to frame the project as a strategic-risk analysis skill for AI agents, distinct from the companion infrastructure project Agenda Intelligence MD.
• Replaced AGENTS.md with the canonical project-rules spec: project identity, honesty rules, evidence rules, recommended README structure, naming hierarchy, and definition of done.
• Renamed "Output modes" to "Memo modes" and reordered the README to recommended structure (quick usage → memo modes → before/after → examples → evals → signal archive → limitations → roadmap).
• Added examples/ with four illustrative memos (sanctions exposure, regulatory impact, scenario brief, red-team), all explicitly labeled reasoning-only / illustrative.
• Added evals/ with a human review checklist, failure-modes catalogue, and starter scoring rubric.
• Added signals/TEMPLATE.md and reframed the signal archive as public examples of the skill style, not official intelligence.
• Added explicit Limitations and Roadmap sections.
• Added an integration status table that separates implemented from compatible from not-implemented.
• Tightened naming hierarchy: Product = Global Think Tank Analyst; Method = Policy Risk Memo Architect; Companion = Agenda Intelligence MD.
• Backfilled evidence-mode and confidence headers on the existing 2026-04-28 signal.

1.1.0 - 2026-05-06

• Repositioned the project from an OpenClaw/Codex-centered skill to a universal AI-agent playbook.
• Added AGENTS.md with runtime-neutral instructions for ChatGPT, Claude, Gemini, Perplexity, Cursor, Codex, OpenClaw, MCP agents, RAG systems, and internal copilots.
• Added Decision Briefing Pack mode for team-facing decisions, owners, watchlists, triggers, and review cadence.
• Expanded README, llms.txt, and skill guidance with stronger user-facing integration paths and clearer evidence discipline.

1.0.1 - 2026-04-17

• Reduced package to minimal skill content for ClawHub release hygiene.
• Kept core SKILL.md aligned with Policy Risk Memo Architect framing.

1.0.0 - 2026-04-17

• Major rewrite and repositioning to Policy Risk Memo Architect.
• Added stronger evidence-discipline guardrails.
• Added clearer output modes and decision-oriented default template.