Website, public-data, translation, research, and agent-interface changes, recorded separately from mobile-app releases.
This changelog tracks changes to the public CBT Cards website, learning library, printable browser worksheets, schemas, language data, research datasets, toolkit pages, structured data, feeds, and agent interfaces. A website, data, translation, research, or agent entry does not imply that an Android or iOS app version shipped on the same date.
Reviewed practice, agent evaluation, and trust hardening
Consolidated the situation-first practice system and the trust boundaries needed for people and AI assistants to use it without treating every raw prompt, benchmark fixture, store claim, or Git commit as equally authoritative.
Published 11 CBT Cards-owned reviewed practices with ontology, evidence provenance, relations, recommendation/no-match contract, and stable RAG chunks.
Added a record-complete audit and freshness registry around the pinned 115-record source toolkit instead of automatically promoting raw records.
Advanced the portable Agent Skill to v1.8.0 and added 41 practice-semantic cases, blinded human-review tooling, and a frozen-context real-provider execution path.
Optimized runtime imagery/fonts and added byte budgets; reconciled privacy, mobile/store locale, legacy-host, search-distribution, and licensing boundaries without claiming external actions that have not occurred.
Added a separately sourced mobile release history; website/data releases remain distinct from Android/iOS releases.
No hosted-model benchmark result or clinical validation claim is introduced by this release. Current machine-readable reuse remains under CC BY-NC-SA 4.0 unless a later explicit publisher-approved license release says otherwise.
· Public data/research
Provenance-first model-run protocol v1
Added a two-stage path for real model evaluations: generate responses before benchmark answers are visible, then score the captured routing envelope with a deterministic evaluator.
Added model-response and model-run schemas for provider/model/runtime provenance, prompt and dataset hashes, raw answers, and per-case routing results.
Added evaluate_model_run.py, which scores route, target IDs, locale behavior, and boundary routing after generation is complete.
CI runs a synthetic contract fixture through the scorer to test the protocol plumbing.
The fixture is not a model result. The v1 evaluator also does not claim automatic scoring of prose quality, expected_checks, or prohibited claims. No LLM or hosted-assistant score is published by this release.
· Public data/research
Held-out paraphrase and adversarial eval challenge
Added 12 separately authored challenge cases to see whether the friendly starter score survives different wording. It mostly does not.
The unchanged deterministic-contract-router-v1 drops from 24/24 starter routes to 1/12 challenge routes.
On the challenge it also gets 0/10 target selections, 0/1 locale behaviors, and 1/3 boundary routes.
The challenge dataset is pinned by SHA-256 f20bbc8562315de235fe6b935b17be8eeaa41d604b222c91904618658b2c0407.
If a system is tuned on these 12 cases, this generation is no longer held out for that system.
The drop is published as evidence that the keyword baseline is brittle, not as a failure of any LLM. No model was evaluated in this entry.
· Agent interface
Portable Agent Skills release v1.7.0
Published a stricter portable skill distribution without rewriting historical skill releases or breaking the existing immutable URL convention.
Added a CI portability check that requires the mutable alias, compatibility mirror, and strict portable distribution to contain identical skill content.
Runtime-specific install details remain outside the portable skill core. This is an agent-interface release and does not imply a new Android or iOS app version.
· Public data/research
Reproducible agent eval baselines and citation metadata
Added a versioned run format and reproducible deterministic baselines so evaluation numbers have an inspectable source instead of materializing from the traditional scientific instrument known as a slide deck.
Every run pins the SHA-256 of the exact evaluation-case dataset bytes used and records both aggregate and per-case results.
null-route-v1 is an intentionally weak floor: 1/24 correct routes, 0/19 targets, 0/4 locale behaviors, and 1/7 boundary routes.
deterministic-contract-router-v1 uses only the user message and reaches 24/24 routes, 19/19 targets, 4/4 locale behaviors, and 7/7 boundary routes on the current starter set.
CI regenerates both deterministic runs and rejects any checked-in result that differs from the runner output.
Added root CITATION.cff using Citation File Format 1.2.0.
The perfect deterministic-router score is a contract/harness baseline, not an LLM benchmark or evidence of general assistant quality. No DOI or project-wide semantic version is claimed before one is actually assigned or deliberately released.
· Public data/research
Public agent routing and boundary evals
Published a starter evaluation set for checking whether assistants use CBT Cards resources and metadata without collapsing every request into the same cheerful card-shaped answer.
Each case records an expected route, stable resource or raw source IDs, semantic checks, prohibited claims, and a rationale rather than one exact reference answer.
Added a versioned agent-eval case schema and CI checks against the live catalog, review overlay, locale registry, translation state, and raw source indexes.
No model scores are claimed in this release. The dataset defines evaluation targets; future recorded runs should identify the evaluated model, version, and date separately.
· Website/localization
Localization review and generated publication pipeline
Completed Russian machine-draft coverage for the current curated knowledge set and added a deterministic publication path that keeps drafts separate from reviewed public localizations.
Expanded Russian pilot data to all 12 records in the current curated English knowledge set. All 12 remain machine drafts; none is currently human-reviewed or published.
Added deterministic generation for language hubs and localized resource pages at /<locale>/resources/<resource_id>/.
Machine drafts never produce localized public HTML. Generated resource pages require human_reviewed, reviewed_for_publication, and published.
The generator also keeps localization sitemap entries synchronized with the data.
The locale-level public_html flag supports gradual rollout: a locale may contain reviewed published records and unpublished machine drafts at the same time.
· Agent interface
Localization-aware agent skill v1.6.0
Published a new agent-skill version that understands CBT Cards language and translation status instead of assuming that any available translation is official published content.
Published immutable skill v1.6.0 and advanced the version manifest and latest alias.
Added the locale registry and translation overlay to the normal source-selection flow.
Defined machine drafts as development data rather than official CBT Cards localizations.
When no reviewed localization exists, assistants may translate the canonical source for a user but should identify that output as their own translation.
This is an agent-interface release. It does not imply a new Android or iOS app version.
· Public data
Locale registry and translation overlay foundation
Added a language model that keeps source content, translation drafts, review state, and localized publication state separate rather than cloning English pages and hoping they stay aligned.
Added data/locales.json with English as the source locale, Russian as a pilot locale, and German as planned.
Added data/translations.jsonl for machine-readable localized overlays keyed by stable resource ID.
Added versioned schemas for the locale registry and one translation record.
Added CI checks for stable resource IDs, locale state, source-review snapshots, key-point alignment, and publication boundaries.
Russian machine-draft records are explicitly machine_draft, unreviewed, and not_published. They have no canonical localized pages. A source update also makes an older translation fail validation until its source snapshot is updated.
· Agent interface
Public reflection resource skill v1.5.0
Reframed the CBT Cards agent skill around using the public reflection library rather than treating the mobile app as the primary resource surface.
Published immutable skill v1.5.0 and updated the version manifest.
Made published cards, worksheets, and reviewed learning pages first-class choices for human-facing reflection tasks.
Added guidance to keep human answers in ordinary language and expose machine details only when they are useful.
Kept the mobile app as the canonical source for app-specific behavior, privacy, support, and download questions.
This is an agent-interface release. It does not imply a new Android or iOS app version.
· Website
Indexed-page crawl graph quality gate
Added direct homepage links to the major public-resource hubs and a CI rule that checks the sitemap as an actual internal-link graph rather than merely a list of files.
Requires every indexed page except the homepage to receive at least one link from another indexed page.
Requires every indexed page to be reachable from the homepage in at most three internal-link hops.
Requires Learn, Worksheets, Toolkit, Languages, About, Changelog, and Agents to be linked directly from the homepage.
Adds a dedicated homepage public-resources section for worksheets, release history, agent/data interfaces, and language status.
· Public data
Versioned public JSON Schema contracts
Published JSON Schema draft 2020-12 contracts for CBT Cards-owned machine-readable formats and made them discoverable from the public resource catalog.
Added a schema manifest mapping committed instances to versioned contracts and allowing interface-only contracts before a real public instance exists.
Added contracts for the resource catalog, changelog, worksheets, toolkit review overlay, toolkit source manifest, curated knowledge-record lines, locale registry, translation records, evaluation cases/runs, model-run interfaces, and agent skill manifest.
Added schema_url links to applicable public catalog resources.
· Website/public data
Toolkit review and publication overlay
Added a CBT Cards-owned review layer that separates raw toolkit source records from records explicitly reviewed for standalone website publication.
Defined the safe default for unlisted source records as unreviewed and source_only.
Recorded six currently published source IDs and published agent skill v1.4.0.
reviewed_for_publication means editorial and safety review for public web use. It is not clinical validation or proof of efficacy.
· Website/public data
Knowledge, worksheets, agent data, and discovery foundation
The standalone site was expanded from product documentation into a broader public knowledge surface that can be read directly, indexed by search engines, or consumed by read-only AI/RAG clients.
Added six sourced learning guides and a learning hub.
Added printable browser-local worksheets for thought records, worry time, and activity planning.
Added a public toolkit landing page, curated card pages, and source-corpus provenance.
Added a stable resource catalog, curated JSONL knowledge records, worksheet schemas, and versioned agent skills.
Added Atom/JSON feeds and a standard security contact endpoint.
· Website
Standalone CBT Cards website
Established the standalone CBT Cards public site at cbt-cards.github.io, with canonical product documentation, features, how-it-works, privacy, terms, support, and FAQ pages.
This records the website launch only. It is not a mobile-app release record.
CBT Cards