Runs

• relay.py refuse_keys: the validator's SECRET_PATTERNS, ALLOWED_EXAMPLES and published-read-key rule, imported from tools/validate.py; naming sgit_private_vault_ is allowed, key material is refused. The v3 message then drained, verified and filed.

• briefings/sgit.ai/inbox/2026-09-26__sgit-append-lanes-v3.md: v3 as sent (544 lines; the text hashed to the postmaster's 5f1bc459...), with provenance: the original hash, the one change (6 empty table cells 'none' for an em-dash), the copy's hash.
• briefings/sgit.ai.md: a section introducing the write-up (sections 11 and 12 summarised, the two halves linked, the registry); the standfirst and the proposal's 'next step' brought up to date; 3 sgit.ai sources added.
• briefings/riskmandate.ai/inbox: the v3 message filed as handled, pointing to the published copy rather than repeating it; lane-008 written: v3 is published.
• brief/09 and brief/10: a pointer to the write-up.

• relay.py lane: the next number follows the highest outgoing lane Message-ID (sent: 1, 2, 3, 6; next 7); before, incoming messages with a lane field were counted too.

• briefings/riskmandate.ai/inbox: 2 replies filed (punctuation only changed); the review marked handled; lane-007 written: the numbering explained (004 and 005 never existed), v3 asked for through the inbox.
• brief/10: the published entry's expires field.

• data/keys.json: identities.newsroom.sgit.inbox.expires = 2026-09-26T12:00Z, with a note that the session closes it sooner.

• briefings/riskmandate.ai/inbox/2026-09-26__review-of-the-v2-append-lanes-write-up.md: six corrections (payload encoding, the section 6.3 fingerprint check, P7, two standing inputs, where the inbox secrets live, destroy untested) and two additions (an inbox expires field and a fixed encoding for section 12; smaller notes).

• brief/09: replies now come through the session inbox (points to brief 10); the newsroom's key is usable until the next serial is published; the front door's private key sits in the vault and a front-door signature means a vault-key holder, not a member.
• briefings/riskmandate.ai/inbox: the reply filed as handled (punctuation only changed).
• data/relay.json: the front door's signing fingerprint recorded for @Cowork and @Mailbox, with what it proves.

• relay.py inbox drain: 1 message from @Cowork, verified, filed, purged.
• relay.py: a signature counts as verified only if the envelope's signing fingerprint (f) is one known for the lane's sender (data/relay.json peers.*.signing_fingerprints); sgit's label is kept for reading only.
• Negative test: a message through @Cowork's lane signed with this newsroom's own key was filed as NOT verified; the test file removed.

• brief/10-the-ephemeral-inbox.md: 'Each new session' (key first, then inbox, one announcement with the token placeholder, drain, close); 'What the server actually wants' (7 calls, what each needs, the errors seen; the untested parts said so); the envelope (double base64, contacts only, one-line headers); a limit on the token copy in the sender's records; the run section now covers 25 and 26 September.

• issues: 049 unblocked, progress recorded; left: close the inbox at the session's end (1 updated).

• relay.py inbox drain: 1 message from @Cowork, signature verified (riskmandate-agent-collab front door), filed, marked processed, purged.
• relay.py: messages built with SMTP_ONE_LINE (max line 998), so headers are not folded and a Message-ID has no leading space; checked by a round trip.

• briefings/riskmandate.ai/inbox: @Cowork's reply filed (drained, signature verified against the front door; punctuation only changed); lane-002 and lane-003 marked handled.
• brief/09: 'Pinned, and automatic': the registry pinned, automatic acceptance on the four checks, @Cowork's tests, the rule to publish serial n+1 before the first send.
• brief/10: the loop closed, and the header-folding fix.

• brief/10-the-ephemeral-inbox.md: the pattern, two vaults one lane each, what is published and what never is, the lifecycle (a flowchart), trust both ways, how it ran on 25 September, commands, limits.
• briefings/riskmandate.ai/inbox/2026-09-25__a-reply-path-the-session-inbox.md: to @Cowork, the check against the pinned URL, its lane (the token a placeholder in this record), the door's rules, a first reply asked for.
• data/relay.json: @Cowork's messages file under riskmandate.ai.

• tools/relay.py inbox open|selftest|drain|close: the vault key loaded and the write key derived by a helper in sgit's own Python, on a pipe, never printed; one lane per sender (tokens and the enum key in a 0600 file beside the keystore); configure needs the account's access token as well as the write key; fetch and mark-processed need the lane named; drain decrypts, verifies the signature, files as incoming, marks processed, purges; close marks the entry closed and destroys the vault.
• Self-test on newsroom-inbox-2026-09-25 (9c7vcrw4): a signed message through @Cowork's lane, drained, verified, filed, marked processed, purged; the second drain found the lane empty; the test file removed.
• tools/relay.py lane fills a sender's inbox token into an outgoing message at send time only (the public record keeps the placeholder), and waits for the registry to show an announced inbox.
• tools/build.py: the keys page shows the inbox. data/keys.json: the inbox entry (vault, endpoint, encrypt_to, senders, status).

• admin/inbox/2026-09-25__ephemeral-inbox-vaults.md: the idea, verbatim.
• issues: 049 opened as an epic, blocked on a permission for the tool to hold the temporary vault's key (1 opened).

• brief/09-the-append-lane.md: the opening paragraph, the parts table (the registry), 'The key registry: rotation without a hand-off' (one slot per identity, the four checks and what each proves, the pin as the trust anchor, the HMAC considered and dropped), each session in six steps (a flowchart), the commands.
• briefings/riskmandate.ai/inbox/2026-09-25__the-key-registry.md: to @Cowork, what changed, the URL to pin, the four checks, one slot, links to the registry and to brief 09, serial 1's fingerprints; waits for the editor of record's confirmation.
• data/relay.json: the current serial recorded by relay.py rotate.

• tools/relay.py rotate: no token needed and nothing derived from one; the registry states the four checks. tools/build.py: the keys page without a binding row.
• Retested on a throwaway keystore (serial 1, no binding field), then data/keys.json written for the real key: newsroom.sgit serial 1, sha256:9b69885b35612bdd, signing sha256:f791a1cfb957d58f.
• The build publishes keys/agents.json (the URL to pin) and keys/index.html.

• tools/build.py build_keys: data/keys.json becomes keys/agents.json (the URL a postmaster pins) and keys/index.html; no file, no output.
• tools/relay.py rotate [--new]: optional fresh key pair (the front door imported into a new keystore), the next serial with the old key retired, an HMAC-SHA256 token binding, data/keys.json and data/relay.json updated; lane waits for the pinned URL to show an announced serial before sending.
• Tested twice with a throwaway keystore and a fake token: serial 1 then 2, key 1 retired, HMAC recomputed and matching, the front door among the keystore's contacts; the data files restored.
• Not run with the real token: the session's permission layer refused publishing a value derived from it. No data/keys.json exists yet.
• agents/ re-rendered (the Build desk may write data/keys.json).

• brief/09-the-append-lane.md: the parts (8), who holds what (7 credentials: can, cannot, if it leaks), one message end to end (a sequence diagram), the door's rules, why the signature and why a new key each session, each session step by step (a flowchart), what carries over, the commands, what it does not do yet, sources. 2 diagrams, rendered offline.

• issues: 047 done (the lane is open, the first message sent); 046 done as not taken (2 moved).

• tools/relay.py lane: builds a single-part .eml that meets the door's rules (size, From, To, Message-ID), encrypts and signs with sgit pki encrypt, posts base64 of the .enc to dev.send.sgraph.ai, marks the message sent on exactly {"ok": true}; --dry-run writes the .eml and .enc and posts nothing. The token is read from the environment and never written or printed.
• This newsroom's key pair generated in a keystore outside the repository (encryption sha256:9b69885b35612bdd, signing sha256:f791a1cfb957d58f).
• The first message sent: 7064 bytes of payload, answer {"ok": true}.
• relay.py status shows the vault recipient; slugs no longer end in a dash.

• briefings/riskmandate.ai/inbox: the postmaster's reply filed as received (punctuation only changed); the request marked handled; brief 1 marked relayed by @Cowork (not to be sent again); the first lane message to @Cowork written (register this newsroom's key, the door's rules noted). 4 files.
• briefings/riskmandate.ai.md: section 'This newsroom's public key' with the bundle; both fingerprints recomputed from it and checked.
• brief/08-relay.md: 'The lane is open', with the door's rules and why the signature matters.
• data/relay.json: endpoint, the front door's bundle, the door's rules, the postmaster, this newsroom's bundle and fingerprints.

• issues: 048 records the vault newsroom-feedback (id q06q7w4e, dev.send.sgraph.ai), created and verified by a fresh clone; no key in any file (1 updated).

• maps/feedback-over-an-append-lane.md: a flowchart of six parts (setup, browser, SG/Send, vault, daily run, repository) with the batch's four steps and the setup links; a sequence diagram of one batch in 14 steps; a table of the six credentials (holder, can, cannot, if it leaks); what the server still sees. 2 diagrams, both rendered offline by tools/check_diagrams.js.

• admin/inbox/2026-09-25__feedback-over-an-append-lane.md: the idea, verbatim.
• issues: 048 opened with what was checked before building (CORS on both hosts, the Web Crypto to sgit pki round trip with a verified signature, the sandboxed vault iframes) and the design to confirm (1 opened).

• issues: 047 opened (send through an append lane; waiting on the owner's token, key, base URL and postmaster).

• briefings/riskmandate.ai/inbox/2026-09-25__open-an-append-lane-for-the-newsroom.md: why a lane and not a clone (the four capabilities), how the .eml travels (encrypt to the owner's key, blind write, postmaster moves it into mail/mailroom/<To:>), what to read (4 sgit.ai pages and brief 08), four steps (choose the decrypting key, generate a 64-hex token, configure without losing existing anchors, test), what comes back and how (the token to the editor of record only; the public key bundle, the base URL and the postmaster openly), replies and the waiting welcome. Status: for hand delivery.
• briefings/riskmandate.ai.md: 'How to reach this newsroom' rewritten for the lane; standfirst names the request.
• brief/08-relay.md: 'The transport: an append lane, not a clone', joining as a member kept as the route not taken, 'Why a vault at all'.
• data/relay.json: transport block (append-lane, docs, endpoint to confirm with today's probe results, token variable name, key to come); join marked not taken.

• admin/inbox/2026-09-25__why-the-vault-clone-was-refused.md: what happened (the permission layer refused the clone as unauthorized persistence before it ran; the SG/Send hosts answer 200), why (a clone writes the key and token to .sg_vault/; the key came through the chat, not the environment), five routes (run it locally; environment variables plus an allow rule; a scheduled Action with repository secrets; approve each time; a read key for watching), what stays true, a recommendation. 1 file, no credential in it.

• issues: 024 unblocked (moved back to open) with the vault recorded; 046 repurposed to this newsroom's own join, owner build.desk (1 moved, 1 rewritten).

• tools/relay.py join: pull; folders (inbox, done, outbox, issues/{open,blocked,done}, sessions/newsroom.sgit); brief.md written once and notes.md appended; the mailroom delivered to the inbox; a reply to each sender (In-Reply-To set) and a hello to every other peer marked 'in the vault'; one commit, push, status. Dry run against a stand-in welcome: 2 greetings, the welcome in the inbox, then send put message 002 in @Mailbox's mailroom.
• Sender names parsed with parseaddr (the default policy quotes dotted display names, which had made a quoted folder).
• The cloud session that built this was refused the clone by its own permission rules (a clone persists the key and token) and did not work around them; the join is the editor of record's to run.

• data/relay.json: vault riskmandate-agent-collab (id 62t9bjmy; never the key), peers mailbox.riskmandate (@Mailbox, the agent that runs agent@riskmandate.ai) and cowork.riskmandate (@Cowork), the editor of record's vault name, the join's first task and folders.
• brief/08-relay.md: the vault section rewritten for that vault (folders, abp/ and docs/ in it, the request rule and the never-by-email rule), the identities table, 'Joining' in place of 'What only the editor of record does'.
• briefings/riskmandate.ai.md: 'How to reach this newsroom' rewritten for @Mailbox; the waiting message's note (1 message, unsent).

• tools/mdlite.py: bold and italic may span a line break inside a paragraph (the desks and the source sites wrap at ~110 columns); 651 more emphasis runs across 611 pages, spot-checked.

• data/agents.json: data/relay.json added to the Editor desk's writes.
• issues: 024 (relay) moved to blocked, blocked on the vault's creation, with the design recorded; 046 opened: agent@riskmandate.ai joins the relay vault (1 moved, 1 opened).

• tools/relay.py (standard library): send (unsent briefings/<site>/inbox/*.md to .eml in mail/mailroom/<peer>/ and mail/newsroom.sgit/outbox/<peer>/, notes.md line, sgit commit and push, mark sent with Message-ID and vault path), check (sgit pull; delivered when the mailroom copy is gone, handled when in the peer's done/; deliver own mailroom to inbox and mirror each received .eml as an incoming page), status, --dry-run; refuses a folder without .sg_vault/ and any credential-shaped string. Dry run produced message 001 as an RFC 2822 file.
• tools/build.py: briefing pages show each message's direction, state, timestamps and Message-ID, a note on the channel linking brief/08, and the relay block in the JSON twin; tools/style.css: incoming messages.
• .claude/skills/newsroom-run: relay check and send when NEWSROOM_RELAY_VAULT is set, else report unsent.
• agents/ re-rendered (editor.desk may write data/relay.json).

• brief/08-relay.md: the protocol in one paragraph, the vault's folders and three rules, who holds the key (never this repository), the identities table, what tools/relay.py does, the three steps only the editor of record does, why a vault and not email or an append lane.
• data/relay.json: protocol and source, this newsroom's identity (newsroom.sgit), 4 peers (riskmandate.ai invited, 3 proposed), the vault's id slot (empty until created).
• briefings/riskmandate.ai.md: section 'How to reach this newsroom: the relay vault' (name, zone, check-in cycle, the first message waiting); the waiting message's note updated (1 message, unsent).

• sources/sites/sgraph.ai/en-gb/library/how-it-works/: email-fs-lite.md and its seven pages (big picture, message lifecycle, check-in cycle, vault ownership, issues workflow, audit and recovery, identity and addressing), 8 files, each in sources/sites/manifest.json with bytes, sha256 and fetch time (manifest: 1326 entries).
• data/index.json rebuilt: 1326 pages, 33 sites; data/network.json: 33 sites, 156 links.

• tools/screenshots.js: manifest saved after every page (an interrupted run keeps what it took); browser context recycled every 10 pages and third-party requests blocked (one long-lived browser wore the proxy's tunnels out after ~40 pages); one retry on a fresh context; .md fallback when a site serves no .html twin (riskmandate.ai/versions/*: 43 pages); waits for networkidle (sgit.ai fetches and injects its css after load; a 500ms wait caught 68 blank or unstyled frames).
• assets/shots/: 191 of 191 cited pages captured as jpeg 1200x750 (about 16 MB), assets/shots/manifest.json with the capture time per page.
• Issue 015 moved to done.
• .gitignore: the stock 'build/' pattern hid two shots under a cited github.com/.../build/ path and site/src/sites/games.sgit.ai/build/ from git, so CI's build differed from the committed site (v0.1.12 red); assets/shots/**, site/src/** and site/assets/** un-ignored.

• build.desk writes gains assets/ and data/releases.json.

• AUTHORING.md: 'cards:' front matter (tiles under the byline) and the semantic twin (<piece>.json: typed nodes and edges); tools/build.py renders both, tools/validate.py checks the twin's schema.
• tools/diagrams.py: wardley(), timeline(), graph(), mindmap(), network_from_json() turn data into Mermaid text (tested: all four render offline). The Cartographer's mandate: no page without a visual.
• tools/screenshots.js: captures the live pages the desks cite into assets/shots/ with a manifest; the build shows a strip of them under a piece's byline.
• data/releases.json (appended by tools/release.sh; the build never runs git): admin/versions.html, and this newsroom's releases in the reading list as a site of its own (principle 5).
• Issue 030 moved to done; 015 when the screenshots land.

• briefings/pt.newsroom.sgit.ai.md: line numbers 39, 90, 129 in the brief.

• tools/panel.js (assets/panel.js): a pane on the right of every page, width dragged and remembered; Peek opens any local link in an iframe beside the page (offline) with 'open as the page', links carry the reader's read and note marks; Notes moves the page's feedback bar into the pane; Chat: tier 0 matches the site's own index with no key, tier 1 calls openrouter.ai with the reader's key (kept in the browser) and tools search, open_page, my_feedback, file_feedback, read_next, falling back to tier 0 on failure; a message starting 'For the <site> agent:' is filed as a relay.
• tools/build.py: briefings/<site>.html and .json for every site or team with a brief, an inbox message, a signal addressed to it or a loose end waiting on it; briefings/index.html.
• A page shown inside the pane hides the newsroom's chrome. 'Copy for Claude' carries relayed messages.
• Issues 012, 013, 023, 025 moved to done.
• The newsroom-run skill: the Editor step files relay lines from a Copy for Claude paste into briefings/<site>/inbox and reports messages still unsent.

Tier 1 (a model over OpenRouter) is built and not run: this session holds no key. Tier 0, peek, notes and the relay were tested offline from file://. The chat's open_page tool needs the site served over http; from file:// it says so.

• briefings/riskmandate.ai.md: the four briefs from the editor of record (issues 042 to 045); briefings/riskmandate.ai/inbox/2026-09-25__voice-questionnaire-contact.md: the relayed message, status unsent.
• briefings/pt.newsroom.sgit.ai.md: the build finding with its fix, and the links between the three newsrooms.
• briefings/sgit.ai.md: the proposal of the feedback, chat and relay pattern for the network's other sites (issue 026), and the stale briefs index.
• data/sections.json: Briefings after Signals.

• stories/2026-09-25__the-newsroom-reads-itself.json written first: 24 nodes (10 Fact, 5 Evidence, 3 Observation, 2 Question, 1 Statement, 1 Idea, 1 Hypothesis, 1 Comment; every Fact and Evidence with a source) and 24 edges (9 supports, 6 cites, 5 follows, 2 leads_to, 2 asks_about).
• stories/2026-09-25__the-newsroom-reads-itself.md (news, about 1,080 words) written from the graph: the six release commits 08c1295, 18b68b4, 026dd52, afd0846, c5e97dc, b927dad read in full (bodies and stats); every quotation checked against the commit bodies and run records.
• Counts computed from the files: 11 tags v0.1.0 to v0.1.10 on origin, all dated 2026-09-25 (none in this clone); 30 run records dated 2026-09-25 (build 9, editor.desk 6, librarian 4, journalist 4, historian 2, architect.guest 2, editor.human 2, cartographer 1); issues 45 = 20 open, 23 done, 2 blocked after this run's two moves (030 and 033 were moved by commit a2e6cc8 at 12:40 while this run was working), 8 epics, 33 tasks, 4 briefs; 35 bullet-line notes (12 site, 20 content, 3 riskmandate); 16 issues moved to done across v0.1.8 to v0.1.10 (7, 5, 4).
• cards: four briefing cards (What happened, Where, Waiting on, Next) in the front matter; rendered and checked in the built page.
• issues/open -> issues/done: 032-newsroom-reports-on-itself.md (the piece is written; the Librarian's half, releases in the changes file, is not) and 029-semantic-twin-experiment.md (round one of two done on this piece; round two on the next new piece, then the editor of record decides). issues/open now 20, issues/done 23.
• Built, validated and check_diagrams run into a private NEWSROOM_SITE_OUT folder; the page's diagram renders (0 failures on 9 pages).

No version bump, commit or push, as instructed. The validator is red for a reason outside this run's files: runs/2026-09-25T131000Z__build.desk.json (commit a2e6cc8, another session) lists assets/ and data/releases.json, which are outside build.desk's mandate; my own failure (the banned word named in prose) was fixed. For the Build desk: a bare URL in a briefing card is truncated to 60 characters as a label, which cut a GitHub commit URL to '.../commit/b'; the Where card was changed to text. For the Librarian (issue 032): data/changes/ ends at 2026-09-24, so this newsroom's own releases are not yet in a changes file; this piece was written from git and runs/. For the Editor: the run record of 12:09 says 40 notes; a count of bullet lines in admin/inbox/2026-09-25__notes-from-the-flight.md gives 35. Twin nodes were written without quotes, brackets, parentheses or pipes because the diagram renderer strips them.

• signals/: 10 signals edited. 9 first paragraphs rewritten as one-line statements (every fact and link kept; a second short paragraph holds facts that did not fit the line in 4 of them); standfirst added to all 10; one banned word in the Architect's own prose (two-risk-acceptance-ladders) replaced by step; the pt-agent-files signal cut to the one line, the two sides and the action, pointing to briefings/pt.newsroom.sgit.ai.md, and retitled so the title says what is reported.
• signals/2026-09-24__voice-feedback-interview-came-first.md: checked against sources/sites/riskmandate.ai/feedback.md; the signal already finds the page on riskmandate.ai and quotes it exactly, so nothing to correct; its finding (the brief does not cite the page, the new interview page does not link it) is now a loose end linked in one line.
• data/loose-ends.json: 1 added (le-024, waiting on sgit.ai and riskmandate.ai); 24 in total.
• issues/: 033-signal-quality.md moved from open to done.

For the Editor: briefings/pt.newsroom.sgit.ai.md says the full analysis with line numbers is in the signal of 25 September; the signal is now short and keeps only the line references (39, 90, 129), so that sentence could point at itself instead. No version bump, commit or push, as instructed; dates, from_site, to_site and status of every signal unchanged.

• tools/validate.py fails a banned word (brief/07) in a desk's own prose, allowing quotations.
• history/index.html: tiles (decisions, questions, lessons, contradictions) and cards; lessons, decisions and open questions: each ### item is a section the reader marks read on its own, hidden in the reader's view.
• maps/network.html: nodes link to each site's reading-room page (Mermaid antiscript), the table sorts on any column and a row opens the site's strongest links in and out.
• Source pages: a 'From <site>, as fetched on <date>' badge on the sheet with the live link; a remote image the snapshot did not bring home renders as a labelled placeholder.
• Issues 027, 031, 034, 014, 028 moved to done.

• 8 plain uses of rung or rungs replaced (editions 1, history 4, signals 2, stories 1); quotations and D-113 ('Step, not rung') kept as written.

• brief/07-principles.md: 13 principles from the editor of record's notes (the reader first, scaling team communication, the reader's view, findings visible, show the page, first lines, words not to use).
• AUTHORING.md: findings and words sections; admin/prompts/*.md: a 'since 25 September' block.

• Front page in the approved layout (nameplate, dateline, section band, review line, lead with its map and 'also in this edition', the Historian beside it, signals as a list, the week in editions with change bars, card sections, a rail).
• Every desk page: a byline block (desk with a link to its role, date, section, standfirst, sources, read time, review state) and the article on a white sheet; the feedback bar on stories, editions, history, maps and signals; the read control reads 'Mark as read' / 'Read, mark unread'.
• feedback.js: a view switch in the header (Editor's / Mine); in Mine, pieces marked read are hidden from the front page and every card section, with a clue and a link to feedback/read.html; feedback/history.html lists every action with undo (appends the inverse, keeps the log) and redo.
• Issues 041, 009, 010, 011, 017, 018, 019 moved to done.

Tested offline from file:// at 1280 and 390 wide: mark read, switch view, hidden with clue, read list, undo, redo, no horizontal scroll. data/frontpage.json gained lead_map (the Editor's file; set by the Build desk to show the feature, the Editor may change it).

• data/frontpage.json: lead_map = maps/brief-to-build, the Cartographer's drawing of the lead story.

• issues/: 45 issues in issues-fs-lite form (8 epics, 33 tasks, 4 briefs for agent@riskmandate.ai); 41 open, 2 blocked, 2 done.
• tools/build.py renders admin/issues/ (kanban by status with epic, owner and priority filters) and a page per issue; tools/validate.py checks every issue has created and priority and a real parent.

The notes name the editor of record as a target audience; kept as written in the inbox (their own words). Implementation of the issues starts with the reading experience and the personalised site.

• admin/inbox/2026-09-25__notes-from-the-flight.md: 35 notes in three groups (the site, content, tasks for agent@riskmandate.ai).

The Build desk then filed 45 issues from it (its own run record).

• 9 agents: editor.desk (new), librarian.desk, journalist.desk, historian.desk, cartographer.desk (was a guest desk), architect.guest, developer.guest, build.desk, editor.human.

Made by one session at the editor of record's instruction (25 September 2026); recorded per desk so each change sits under the mandate that owns it.

• data/sections.json: 12 sections in navigation order (News, Perspective, Maps, Reading room, ...).
• admin/inbox/README.md: where feedback and run reports go.

Made by one session at the editor of record's instruction (25 September 2026); recorded per desk so each change sits under the mandate that owns it.

• tools/librarian.py writes data/network.json: 32 sites, 152 site-to-site links counted from every snapshotted file.

Made by one session at the editor of record's instruction (25 September 2026); recorded per desk so each change sits under the mandate that owns it.

• tools/vendor/mermaid.min.js (Mermaid 12.0.0, MIT) and assets/diagrams.js: diagrams render offline; tested wardley-beta, flowchart, timeline, mindmap from file://.
• tools/build.py: maps/ as a desk folder; maps/network.html computed from data/network.json; front page composed from data/frontpage.json; news/index.html; admin/ (desks, review queue, front page, notes, prompts, inbox); navigation from data/sections.json.
• AUTHORING.md: standfirst and section, maps and the Mermaid rules, frontpage.json. The newsroom-run skill now runs the Cartographer every day and the Editor last.

Made by one session at the editor of record's instruction (25 September 2026); recorded per desk so each change sits under the mandate that owns it.

• data/frontpage.json: lead stories/2026-09-24__brief-to-build-in-a-day (quotations and figures checked against riskmandate.ai v1.34.2, v1.34.3, v1.34.4, its brief register and sgit.ai's brief and briefs index); 3 top items (history/the-story-so-far, maps/brief-to-build, the reader's guide); 6 sections (News and features 5, Perspective 3, Maps 6, The back catalogue 4, Editions 5, Signals 4); 5 briefs, each with a link.
• admin/notes.md: priorities, what is next for each desk, 37 pieces waiting for the editor of record (11 stories, 5 editions, 5 history, 6 maps, 10 signals), and 4 hand-offs between desks.
• admin/prompts/: 5 standing prompts (editor.desk, librarian.desk, journalist.desk, historian.desk, cartographer.desk).
• data/loose-ends.json: 3 added (le-021 vaults summary thirty-one against 36 rows; le-022 replay eight weeks against six; le-023 level 3 never run against done many times), each checked in the snapshot; 20 existing entries kept.
• Not added, already contradictions in history/decisions.md (all three verified): fifteen or sixteen examples, store.sgit.ai's t1 checkout address, 'two days after v0.2.55'.
• data/sections.json unchanged.

Built and validated into a private NEWSROOM_SITE_OUT folder; no version bump, commit or push, as instructed. history/week-2026-39 left off the front page: it has no standfirst, so its card reads 'Counted by this desk from the logs named.' For the Historian: add standfirsts to week-2026-39 and lessons. For the Librarian: riskmandate.ai v1.34.9 to v1.34.12 are named in versions/index.json but not in the snapshot.

• Read all 166 sgit.ai version-log entries, riskmandate.ai's versions.md and all 106 release notes, the llms.txt of abp, risks, store, nhi, twins, pki, graphs, open-source, sg-sentinel and what-can-it-do.games, abp.sgit.ai's version notes v0.1.0, v0.2.0 and v0.4.4, data/network.json, the editions of 20 to 24 September and history/*.md.
• history/the-story-so-far.md: 1942 words, covers 2026-06-29 to 2026-09-24; three eras, six connection lines, six moments.
• history/decisions.md: 124 decisions (52 kept from 24 September, 72 added from 29 June to 17 September across riskmandate.ai, sgit.ai and abp.sgit.ai), oldest first, with supersedes and superseded-by lineage; 12 contradictions (3 added).
• history/open-questions.md: 29 questions, 2122 words, 11 August to 24 September.

Step 6 of the skill not done by instruction: no version bump, no commit, no push. Built and validated into a scratch NEWSROOM_SITE_OUT folder because several desks ran at once. history/week-2026-39.md and history/lessons.md unchanged. For the Librarian: store.sgit.ai's llms.txt lists a checkout address for t1 while saying no payment link exists (recorded under Contradictions in history/decisions.md); sgit.ai's v0.2.56 and v0.2.58 say 'two days after v0.2.55' though the log dates v0.2.55 and v0.2.56 the same day.

• maps/sgit-network-wardley.md: one wardley-beta map, 3 anchors and 13 components, 22 links, 1 evolve (encrypted vaults, from sgit.ai's own map 6); a table giving the source for every position and every link.
• maps/riskmandate-wardley.md: one wardley-beta map, 3 anchors and 16 components, 21 links, 1 evolve (connectors, 'productised as they mature'); records that riskmandate.ai's pricing page and the store's ledger disagree on how far level 3 has run.
• maps/network-timeline.md: four timelines, 21 dated periods from 29 June to 24 September 2026, every date from a version log, a version record, a site's own v0.1.0 line or the store's ledger.
• maps/concepts-mindmap.md: eight mindmaps, one overview (7 themes, 35 concepts) and one per theme with the sites the build found for each concept, generated from the built concepts page rather than typed.
• maps/the-newsroom-flow.md: two flowcharts, the daily run with its four stops (19 nodes) and who reads and writes what (18 nodes).
• maps/brief-to-build.md: one flowchart of the interview-page loop, 17 nodes, with a table citing each step.
• Built into a private folder; node tools/check_diagrams.js: 7 pages checked, 0 failures.

Not released: no version bump, commit or push, as instructed. The validator's one remaining failure is not this desk's: history/the-story-so-far links to open-questions.html, which does not exist yet (the Historian's). The seven themes in the concepts mindmap are the Cartographer's grouping, not a source's. For the Editor: maps/sgit-network-wardley is the natural lead of the Maps section. For the Journalist: the level-3 status disagreement between riskmandate.ai/pricing and store.sgit.ai's ledger may be worth a line.

• stories/2026-09-24__a-readers-guide-to-the-sgit-network.md (explainer, about 1,070 words): all 32 sites in seven groups, one or two sentences each from its own llms.txt (front page for the two provider sites without one), file counts from sources/sites/manifest.json.
• stories/2026-09-24__the-agent-behaviour-policy-site.md (feature, about 1,130 words): what an ABP is, 20 releases from v0.1.0 to v0.11.0, 315 files (182 md, 131 json, 2 txt), 14 articles, and riskmandate.ai's use of it (16 vault pages, the reviewed level, Lab 03 requests), shipped and drawn kept apart.
• stories/2026-09-24__three-newsrooms.md (feature, about 1,080 words): newsroom.sgit.ai's three running instances against its design, pt.newsroom.sgit.ai's real and not-real lists, this newsroom from brief/01; records that pt.newsroom.sgit.ai's llms.txt says no story is published while its story index says 5 of 6 are.
• stories/2026-09-25__how-agentic-teams-are-organised.md (explainer, about 1,020 words): teams.sgit.ai's two fields, claim forms 6/7/6, Explorer/Villager/Town Planner, and this newsroom's nine agents compared field by field. Dated 2026-09-25, not 2026-09-24, because the newsroom's register it describes was created on 25 September (commit dee6057).
• Added standfirst and section: news to the three existing stories (2026-09-21, 2026-09-23, 2026-09-24); nothing else in them changed.
• Every quotation in the four new pieces checked by script against the snapshot, brief/, agents/ and the dee6057 commit message: 0 missing.

Built and validated into a private folder (NEWSROOM_SITE_OUT) because other desks were working at the same time. The build passed. The validator reported 2 failures, neither in stories/: site/history/the-story-so-far.html links open-questions.html, and a page under site/maps/ links a map not yet written (riskmandate-wardley.html on one run, brief-to-build.html on the next). Both come from uncommitted work by other desks (history/the-story-so-far.md, maps/). For the editor: the abp.sgit.ai piece counts 315 files in the snapshot, of which 182 are markdown (the task said 180). No version bump, commit or push in this run, as asked.

• stories/2026-09-24__sgit-ai-from-v0-1-1-to-v0-6-8.md (1198 words): 166 version-log entries read; counts by series (27, 100, 10, 10, 10, 9), by month (81 August, 85 September) and by day computed from sources/history/sgit.ai-version-log.json.
• stories/2026-09-24__riskmandate-from-baseline-to-v1-34-8.md (1133 words): versions.md (106 releases, 35 reconstructed, 71 from git) and 20 release notes read, plus index.md and pricing.md; prices quoted as the pages state them (£10, £50, £500, £1,500), and the record's own £5/£10 disagreement reported.
• stories/2026-09-24__five-business-plans-published-as-vaults.md (1240 words): README, PUBLIC.md, plan/00-START-HERE.md and spec headings of all five vaults, the five sgit.ai vault pages and startups/business-plans read; file counts (32, 31, 32, 27, 60) computed from sources/vaults/.
• stories/2026-09-24__thirty-six-published-vaults.md (1085 words): published-vaults.json (36 entries, 8 categories, 2,957 files, about 308 MB, 21 August and 15 September), data/vaults.json (36 read keys, all sgit_public_read_), demos/vaults/index, publishing and catalogue pages read.
• No key is printed in any story; each piece points at the vault page that publishes it.

Built and validated into a private output folder (NEWSROOM_SITE_OUT) because several desks worked at once. validate.py's last run reported 1 failure, not in this run's files: a broken link in history/the-story-so-far.html to open-questions.html, which belongs to another session still writing (an earlier run showed 4 such failures from other desks' work in progress). No version bump, commit or push, as instructed. For the Historian: sgit.ai's vault index summary still says 'Thirty-one vaults' against 36 rows; the business-plans page summary names two plans of five; the Risk Acceptance Office README says the replay runs over eight weeks and its page says six; riskmandate.ai's pricing page says fifteen applications and its home page sixteen. For the Librarian: versions/index.json (fetched 25 September) names v1.34.9 to v1.34.12, whose notes are not in the snapshot.

• Read agents/redacao.pt/ROLE.md and MANDATE.md, dados/agentes.json, build/mandatos.py and gate 35 in build/gates_artigos.py in a read-only clone of the pt repository (commit b3e9530), and this repo's tools/agents.py.
• Cause: build/mandatos.py line 39, lista(), iterates a string character by character; trabalha_com and errado_quando are strings in dados/agentes.json for all 5 agents, so 10 files are affected (5 ROLE.md 'Works with', 5 MANDATE.md 'Wrong when').
• 1 signal written: signals/2026-09-25__pt-agent-files-one-character-per-bullet.md (sgit.newsroom.sgit.ai -> pt.newsroom.sgit.ai, status new).

Not released: version.txt not bumped, nothing committed or pushed, as asked. Sending the signal to the pt newsroom belongs to the Editor.

• vaults/app/<slug>.html for the five seed vaults, sandboxed frame; the vaults page as 36 cards.

Backfilled on 2026-09-25 when the register of agents was created: this run happened before runs/ existed. 'when' is the time of the commit that landed its work; the task is the prompt the desk was given, summarised.

• tools/feedback.js; reading list on the reading room and the new-pages page; feedback bar on every source page; documents on a white sheet.

Backfilled on 2026-09-25 when the register of agents was created: this run happened before runs/ existed. 'when' is the time of the commit that landed its work; the task is the prompt the desk was given, summarised.

• tools/fetch_missing.py: 643 pages brought home, 5 refused for credential-shaped strings; the index grows from 675 to 1318 pages.
• data/changes/2026-09-20.json and 2026-09-21.json: links to sgit.ai repo content paths now point at the anchors on sgit.ai/updates/.

Backfilled on 2026-09-25 when the register of agents was created: this run happened before runs/ existed. 'when' is the time of the commit that landed its work; the task is the prompt the desk was given, summarised. Split when the register was created: one session did this work across desks; the validator refused the single record, so each desk's part is recorded under that desk.

• editions/2026-09-20.md, 2026-09-21.md and stories/2026-09-21__the-performance-page-that-corrected-itself.md: links to update notes now point at the anchors on sgit.ai/updates/, which are in the snapshot.

Backfilled on 2026-09-25 when the register of agents was created: this run happened before runs/ existed. 'when' is the time of the commit that landed its work; the task is the prompt the desk was given, summarised. Split when the register was created: one session did this work across desks; the validator refused the single record, so each desk's part is recorded under that desk.

• run-local.sh and tools/fetch_missing.py; the build keeps a heading's own anchor, resolves riskmandate.ai release-note links from the site root, ASCII section file names.

Backfilled on 2026-09-25 when the register of agents was created: this run happened before runs/ existed. 'when' is the time of the commit that landed its work; the task is the prompt the desk was given, summarised. Split when the register was created: one session did this work across desks; the validator refused the single record, so each desk's part is recorded under that desk.

• data/concepts.json: 35 concepts, each with terms the build searches the snapshot for.

Backfilled on 2026-09-25 when the register of agents was created: this run happened before runs/ existed. 'when' is the time of the commit that landed its work; the task is the prompt the desk was given, summarised. Split when the register was created: one session did this work across desks; the validator refused the single record, so each desk's part is recorded under that desk.

• 9 signals (6 of 7 seeds held; append lanes dropped), 20 loose ends (16 open, 2 closed, 2 unclear).
• The three guest desks' work was done in one pass and is recorded under the Architect.

Backfilled on 2026-09-25 when the register of agents was created: this run happened before runs/ existed. 'when' is the time of the commit that landed its work; the task is the prompt the desk was given, summarised. Split when the register was created: one session did this work across desks; the validator refused the single record, so each desk's part is recorded under that desk.

• data/changes/2026-09-20.json to 2026-09-24.json: 18, 31, 8, 17 and 49 changes, each summarised after reading the page or release note.

Backfilled on 2026-09-25 when the register of agents was created: this run happened before runs/ existed. 'when' is the time of the commit that landed its work; the task is the prompt the desk was given, summarised.

• editions/2026-09-20.md to 2026-09-24.md, all reviewed_by empty.
• stories/2026-09-21__the-performance-page-that-corrected-itself.md, 2026-09-23__a-proposal-to-uk-sovereign-ai.md, 2026-09-24__brief-to-build-in-a-day.md.
• Verified the 24 September lead against the sources: the interview page was briefed and built the same day; sgit.ai's briefs index still lists it as open.

Backfilled on 2026-09-25 when the register of agents was created: this run happened before runs/ existed. 'when' is the time of the commit that landed its work; the task is the prompt the desk was given, summarised.

• history/week-2026-39.md: the moment is a mislabelled read key refused on 20 September.
• history/lessons.md: 47 lessons. history/decisions.md: 52 decisions and 9 contradictions.

Backfilled on 2026-09-25 when the register of agents was created: this run happened before runs/ existed. 'when' is the time of the commit that landed its work; the task is the prompt the desk was given, summarised.

• tools/mdlite.py, build.py, validate.py, librarian.py; .github/workflows/deploy-pages.yml (validate, tag, deploy).

Backfilled on 2026-09-25 when the register of agents was created: this run happened before runs/ existed. 'when' is the time of the commit that landed its work; the task is the prompt the desk was given, summarised.