# Changelog ## 0.11.5 · 2026-11-08 ### Fixed - Automatic voice completion and a participant’s manual review request share one action, preventing duplicate summary requests and a misleading conflict error. ### Verified - Hosted synthetic voice clarified a notification/“Sen.”, resumed saved history and edited drafts, and stayed connected through a 324-second answer. - Final review used one successful summary request; edited-summary submission/reload preserved all 36 stored turns independently. The approved fictional test was removed. - 105 automated checks, types, focused lint and production build passed. No new human voice-quality verdict is claimed. ## 0.11.4 · 2026-11-08 ### Fixed - Interview instructions consistently keep time as pacing guidance. Ending requires a clear request to stop or a completed conversation the participant is ready to leave. ## 0.11.3 · 2026-11-08 ### Fixed - Manual review and submission use the confirmed saved conversation after voice finishes. Repeated clicks wait for that save instead of racing the final transcript. - Voice finalisation has more time to confirm saving before reporting a recoverable error. ## 0.11.2 · 2026-11-08 ### Fixed - Short or unclear speech such as “sen” triggers clarification. An early voice ending after an ambiguous fragment keeps the conversation open for resumption. - Fortsätt prata is available during the conversation and summary review. Resuming retains the saved turns and the participant’s draft summary. - A closed connection can settle its saved text before resuming. ### Changed - Five minutes guides pacing. Voice and written interviews can continue when the participant wants to say more, including after a pause. - Speech filtering includes background voice detection. Technical connection limits retain a resume path. ## 0.11.1 · 2026-11-08 ### Changed - Prata is preselected when opening the follow-up. Participants give voice consent and start the call explicitly, or choose Skriv. ### Removed - Three fictional follow-up test submissions, with owner-authorised deletion and readback verification. ## 0.11.0 · 2026-11-08 ### Changed - Swedish voice and written interviews follow up the 6 November session with three questions, in order: tomorrow's action, other learning and takeaways, and what is still missing. - The session page links to the follow-up at /261106/uppfoljning. Existing preparation links also open the follow-up. - Owner results have separate views for preparation and follow-up. ### Added - A private follow-up campaign with its own consent version, participant cookie and bounded response window through 30 November. - Independent editable summaries and full received conversation text continue through review and submission. ### Verified - Hosted written and synthetic Swedish voice follow-ups completed review, edited-summary submission and reload. Saved conversation text stayed independent of summary edits. - The fresh voice run reached review and private save in 205.835 seconds, with 84 successful checkpoints and zero failed requests. - The participant-question fallback referred an unknown booking date to JJ. Callback disablement is checked against the effective provider destination before every new voice call. - Late provider finalisation now has bounded automatic recovery through the verified session readback. - 91 tests, TypeScript, production build and focused lint pass. All 160 public sitemap routes and anonymous private-access denials are verified. ## 0.10.1 · 2026-11-06 ### Added - The 5 November Grok monitor briefing, with all seven visible links. - Four checked digests on internal redeployment, job quality, Arvika's AI guideline and trust in HR search. ### Verified locally - The corpus contains 118 documents and 751 passages. The previous 113 documents and 733 passage identities, texts, offsets and hashes are preserved. - All 751 passages are available to keyword search. The deployed semantic index covers the previous 733 passages. ## 0.10.0 · 2026-11-05 ### Changed - Four primary destinations: Start, Guides, Library and My workbench. Search, the current session and the AI connection remain easy to reach. - A compact home page and Guides collection connect municipal explainers, cases and reading paths. Existing public URLs keep working. - The workbench follows a question through selected material, a user-written working view, uncertainties and a next step. Its question opens a relevant library search. - Existing questions, selections and notes keep their browser storage keys and are not rewritten on page load. Source removal offers an undo including the note. ### Added - Separate working brief and source-pack downloads, plus a complete, editable-instruction handoff for another AI. Nothing is automatically sent to another AI. - Direct downloads of a digest, briefing, selected passage, reading pack or public guide, retaining source context and absolute links. - Shared-reference and same-document indicators help readers inspect the evidence behind a selection. - Shared desktop and mobile navigation, contextual library views and a compact return to the current working question. ### Preserved - All 113 documents, 733 passages, four public guides and the deployed search index remain unchanged. No interview runtime, provider configuration or participant record changed. ## 0.9.0 · 2026-11-04 ### Added - Three Grok monitor runs from 2–4 November, with all 26 visible source links preserved. - Twelve checked digests connect workflow/workforce redesign, municipal AI precedents, current paid Copilot documentation, Eneo, Helsingborg and Svea. - Swedish municipal AI sections distinguish Copilot Premium from other variants and examine Eneo's shared platform, local service ownership and documented Helsingborg participation. - Reading paths for workflow/workforce redesign and examined AI workflows; all new public pages are included in discovery and the signed read-only MCP. ### Changed - The workbench now contains 113 documents, 733 exact passages, six reading paths and four public guides. Existing 675 passage IDs, text, positions and hashes are preserved. - The Copilot section uses Göteborg's 2025 Chat investigation with its actual scope, and the Eneo section separates current Svea project descriptions from unverified success or failure claims. - Review dates of living project documentation are labelled as checks rather than new publication dates. ### Verified - All 154 canonical sitemap entries are reachable and all public datasets match the prepared source. - The refreshed search index serves the new sources; 16/16 targeted hybrid checks and distinct Premium/Eneo semantic lookups passed. Four public guides and 117 MCP resources are available. - Existing owner access controls remain in place. Interview work is deferred to final preparation; no interview runtime or provider setting changed. ## 0.8.2 · 2026-11-03 ### Security - Explicitly disable inherited provider callback events. A cleared webhook selection alone can inherit workspace defaults. - New voice calls require verified private authentication, no audio storage, no webhook or cloud-storage export events and the intended duration cap. - The live provider readback passed these checks after publication. The written and voice submission tests recorded in 0.8.1 continue to cover the unchanged conversation and review flow. ## 0.8.1 · 2026-11-03 ### Fixed - Live voice checkpoints save the signed interview transcript without waiting for provider post-call metadata. - Final provider transcripts are bound to their own interview through an opaque server-generated reference, with more time for finalisation. Other interviews remain inaccessible. ### Verified - Hosted written and live voice interviews complete review, summary editing, submission and reload. The complete received transcript stays unchanged and separately visible to the owner. - Synthetic voice QA: five spoken answers, 86 successful checkpoints, no detected early replies, automatic review and a verified provider-final transcript. Anonymous owner and invitation requests are denied. ## 0.8.0 · 2026-11-03 ### Changed - The Swedish preparation interview uses the accepted ElevenLabs voice, patient turn-taking and a five-minute maximum with a shorter conversational target. - Participants can ask a brief question back. When time or context is insufficient, the interviewer refers them to JJ. Light time cues respect the participant's time. - The participant edits and submits the written summary separately from the full received conversation transcript. Final voice transcription is checked before the private save is confirmed. ### Added - Private personal links for source-grounded role, organisation and relevant professional background, available only in JJ's owner view. No invitations are sent automatically. - Short-lived server-authorised voice connections, checkpointed speech and recovery of a completed voice call after a page reload. Individual background and responses remain outside public discovery and MCP exports. ## 0.7.0 · 2026-11-02 ### Changed - Activated the owner-approved Swedish interviewer brief in written and voice conversations: toolset, skillset and mindset in personal practice and leadership, with shared learning, recognition and freed capacity. - Six adaptive main questions and at most two short follow-ups replace the provisional themes. Non-use and justified scepticism are respected. - Participant review uses a short first-person summary. Voice directs participants to written review and never claims to submit on their behalf. - Summary edits and submission retain the full attributed transcript independently. ## 0.6.2 · 2026-11-02 ### Changed - Removed the Copilot access checkbox, opening branch and owner badge. Participants begin with a neutral question about AI in their working day; no product access is required or assumed. - Added a short notice that this is JJ's independent preparation experiment, not a ready-made model for municipal AI use. - Review copy makes clear that the full conversation text accompanies the editable summary. Original attributed turns remain separate and are not rewritten by summary edits. - Existing stored legacy access values and transcript data are preserved. The broader question-flow revision remains a review draft. ## 0.6.1 · 2026-11-02 ### Added - Explicit crawler and agent pointers in robots.txt for the real public corpus, learning guides, Markdown and MCP surfaces, with use remaining subject to each source's stated rights. - Host discovery at `/.well-known/host-meta.json` and an experimental ARD v0.91 manifest at `/.well-known/ard.json`, advertised by Agentmap and an HTML discovery link. - ARD capabilities generated from the existing ten read-only MCP tools, with representative questions. No feeds or unimplemented services are advertised. - Both public discovery documents are included in the sitemap, now covering 134 pages and resources. ## 0.6.0 · 2026-11-02 ### Added - Fifteen complete Grok briefings, 18 October–1 November, with 164 preserved visible source links. - Thirteen checked primary-source digests covering municipal HR assistants, recruitment, scheduling, Copilot evaluation, labour research, workforce planning and governed HR agents. - A complete XML sitemap at `/sitemap.xml`, linked from `robots.txt`, and a human-readable directory at `/sitemap`. - Crawlable reading URLs for all 98 documents, ten themes, three perspectives and four reading paths, with readable server-rendered content and page-specific canonical metadata. ### Changed - The shared site/MCP corpus now contains 98 documents and 675 exact passages. All 522 earlier passage IDs, text, offsets and hashes are preserved. - Reading paths now connect the revised Danish study, governed HR applications and the local Eneo assistant. - Internal navigation and signed MCP source provenance use normal URLs; legacy hash links remain usable. The homepage and changes view show the current content date. - Public learning-guide exports use the canonical session URL. Private interview flows, individual answers and owner controls remain excluded from discovery exports. ## 0.5.0 · 2026-10-30 ### Added - A plain-language MCP introduction at /mcp, with a copyable setup prompt, official app guides, examples and a file-based fallback. - Three read-only MCP tools to list, search and read the public Kalmar and Helsingborg learning guides, plus signed guide resources. - MCP compatibility with 2025-11-25 alongside 2026-07-28 and 2025-06-18. Guide content participates in the signed dataset fingerprint. ### Changed - Kalmar and workshop learning material are public without authentication. Previous private review links lead to the public Kalmar page. - The public MCP remains limited to explicitly published source documents and static learning material. Interview responses, recordings, summaries, recovery text, private workbench notes and management controls remain excluded. - Existing corpus passage IDs, original text and semantic index are unchanged. Guide retrieval is lexical; source-corpus semantic retrieval keeps its existing scope. ## 0.4.4 · 2026-10-30 ### Added - Private Swedish Kalmar/IMY review experience with case exploration, four HR scenarios, source-version comparisons and a downloadable pilot brief. - Expiring facilitator review access, checked server-side for the document and every asset. The public workshop remains closed and no participant navigation or interview behaviour changes. - Exact supplied revised SKR report, distinguished from the older public download. Review notes never enter the interview store or public knowledge corpus. ### Verification - Focused interaction, access-denial, expiry and source-byte checks; mobile/desktop review. No paid builder turn. ## 0.4.3 · 2026-10-30 ### Fixed - Full voice interviews no longer fail at 30 speaker turns. Short backchannels and interruptions count as speech segments; transcript storage now uses generous bounded size limits for the full conversation. - Long spoken answers can exceed the written-answer box limit without truncation. - Interrupted speech can be saved as a separately labelled recovery copy before continuing. Existing attributed turns are preserved; recovery requires the same participant session and cannot overwrite an active voice lease. ### Verification - Reproduced the former 31-turn rejection with a rolled-back synthetic database check. After repair, 31-turn checkpoints and a 96-turn final transcript pass; long recovery text is preserved and deleted records cannot be resurrected. - The earlier short voice tests did not establish full interview reliability. Added longer-conversation and recovery regressions; private raw participant data is excluded from tests, source and public evidence. ## 0.4.2 · 2026-10-30 ### Fixed - Voice connection and recovery messages are Swedish and do not expose raw provider diagnostics. - Normal voice shutdown acknowledges the saved transcript without a spurious connection-error alert. - The settings observation is attributed to the assistant who inspected the actual project. ### Verified - Independent live checks confirmed a real audio connection, Swedish speech input and output, and adaptive text follow-ups. Participant review remains required because speech transcription can be imperfect. - Owner sign-in was completed by JJ; anonymous and ordinary authenticated database access remain denied. Working material stays closed until JJ opens it. - Two completed Lovable builder turns used 9.5 + 7.9 = 17.4 credits within the approved 20-credit build cap. Runtime AI and hosting are separate. ## 0.4.1 · 2026-10-30 · Interview reliability - Private voice checkpoints every two seconds, renewable 30-second CAS leases, atomic paid delegation quotas and a ten-minute call limit. - Final-save readback gates switching, review and deletion; received speech remains visible if settlement fails. Transcription requires participant review and is not claimed complete. - Voice starts after onboarding consent/session creation. Text-to-voice consent remains explicit. Swedish errors and bounded 30-second/2400-token text responses. - Worker execution context is explicitly wired; npm lock matches dependencies. Public corpus and MCP source bytes are unchanged; workshop material stays closed. - JJ independently confirmed AI app context and visitor analytics OFF. Fictional Swedish provider follow-up and grounded summary received. Real native session creation, closure and private save acknowledgement observed; synthetic WebRTC did not connect within the grace period, so audible conversation is unverified. - Owner sign-in remains owner-initiated; no email sent. ## 0.4.0 · 2026-10-30 · Closed preparation release ### Added - A Swedish preparation conversation for 6 November: initials, voluntary participation, Copilot access confirmation, skipped questions, participant-corrected summaries and deletion without an account. - Server-authorised individual results for JJ, separate from the public knowledge corpus and MCP. - Owner controls for opening the interview and releasing the session working material independently. ### Changed - The preparation conversation comes first on #/261106, followed by practical details. Working material stays server-side until JJ deliberately opens it. - Interview questions remain provisional and are isolated for a later editorial revision. ### Verification status - The documented native GPT-Live 1 relay and consent-checked private adapter are implemented but release-gated; audible behavior, durable final-transcript boundaries and owner sign-in remain unverified. - Streamed GPT-6 Sol is selected for adaptive text replies. Live provider checks are blocked until project AI app context is verified off; no fictional provider call is claimed. - API, voice release gates and microphone-denial fallback tests pass. Desktop/mobile closed-page checks pass. Email sign-in and the canonical callback allow-list are configured; no email was sent. - Interview accepting remains false and workshop material remains closed. The preparation page may be published without enabling interviews. - No audio recording store is introduced. No participant answers are added to the public corpus. ## 0.3.0 · 2026-10-17 ### Added - Ten Grok briefings from 8–17 October, with rendered browser originals and preserved source links. - Eight source digests: Oyster AI talent, IBM CHRO judgment, Deloitte agentic work redesign, Workday autopilots, OECD public workforce and digital government, AI Sweden skills, and SKR municipal collaboration. - Four guided reading paths: released time and recognition; onboarding an AI colleague; everyday capability; municipal adoption. - A four-source comparison action that preserves existing workbench notes and questions and explains when more space is needed. - Direct links from scoped Grok source checks to relevant source digests. ### Changed - Newly added source digests appear on the overview, while source publication dates stay distinct from ingestion dates. - The October update links new sources into themes, perspectives, recurrence clusters, lexical search and the fixed-source OpenAI index. - The Swedish activity remains #261106; commercial details remain private. - Platform build configuration is synchronized at 2.26.5. The platform preview-session helper handles local session seeding and revocation across preview tabs; production access scope is unchanged. ### Evidence and limits - 70 documents: 41 Grok briefings and 29 editorial source digests; 522 exact passages. - Source checks distinguish IBM's worried subgroup from all employees, August Deloitte research from October coverage, and Workday early-adopter access from planned 2027 availability. - Reading paths are editorial navigation, AI + JJ. They do not establish causal relationships or agreement between sources. Survey findings and product claims are not independently reproduced. ## 0.2.0 · 2026-10-07 ### Changed - English is the main interface, editorial and search language. Source originals remain intact; legacy Swedish links still open their corresponding pages. - The Swedish Helsingborg activity is identified as #261106, with a stable dated route for future sessions. - Primary-source digests and source-library material requests have English editorial text, with scope and limitations retained. ### Added - The complete user-supplied 7 October Grok report, preserved byte-for-byte in the source archive and integrated into topics, passages, recurrence navigation and MCP retrieval. - Seven separately checked source digests: Workday's workforce report, SAP/TechWolf, DNB's restructuring announcement, OECD workplace AI and public-service learning culture, a time-reallocation preprint, and the UK government AI playbook. - Scoped linked-source checks beside the new Grok report, including the distinction between applications per offer and filled roles. - Branded 1200×630 sharing image, Open Graph/Twitter metadata, favicon fallbacks, touch icons and a web manifest. ### Fixed - A research question can be cleared without clearing saved sources or notes. - The working prompt can be edited, copied, cleared and reset to the current default. - Search instructions and examples now match the corpus's English working language. - Company-name constraints use whole words, so “disappear” no longer incorrectly restricts results to SAP. ### Corpus and trust - 52 documents: 31 Grok reports and 21 source digests; 418 exact passages. - The fixed-source OpenAI index and signed MCP corpus are updated together. Signatures attest representation and provenance, not factual truth. - Grok reporting, independently inspected source material and AI + JJ interpretation remain separate. The operative AI Act amendment and case-specific legal application remain an explicit verification limit. ## 0.1.1 · 2026-10-06 ### Changed - Calibrated semantic retrieval to reject unrelated queries while preserving the fourteen checked primary-source cases. - Signed provenance identifies AI + JJ editorial digests as Studio Jardenberg publications, with underlying source publishers kept separately. - Readable source views suppress internal anchor markup and duplicate excerpt headings. ### Verified - Canonical custom domain, server-side OpenAI retrieval and complete 387-passage index. - Official Node and Python live signature verification, signed tool and JSON-RPC errors, and current protocol caching. ## 0.1.0 · 2026-10-06 ### Added - A Swedish editorial AI and HR knowledge workbench, with light, dark and system appearance. - Thirty preserved Grok briefings plus fourteen primary-source digests, 387 exact passages and linked provenance. - Whole-word lexical search, server-side OpenAI semantic retrieval and hybrid ranking, with source-context views. - Ten connected HR topics, personal/team/organisation lenses, date and provenance filters. - A local four-source comparison board, personal notes and complete Markdown export with copyable fallback. - A public read-only MCP and verifiable-response trust layer, server discovery and public keys. - Source recurrence navigation, checked research additions and practical session experiments. ### Security and boundaries - OpenAI credentials and MCP signing keys are server-only. Retrieval uses persistent limits, query-vector caching and an atomic fixed-source index bootstrap. - Signed provenance does not establish factual truth. Grok claims remain unverified unless independently checked; primary digests distinguish evidence, limitations and editorial interpretation. - Third-party originals are linked; the new fourteen sources are concise attributed digests, not full republished reports. - Workbench notes stay in the current browser. Initial publication and independent live verification are recorded separately in the release handover.