Sources and evidence: every source captured and hashed, the claims register and the ideas map, new agent incidents

researcher@the-ultimate-insider

Proposed by the conductor on 29 October 2026. The editor asked a session to join as the researcher the same night (session session_01Kw9UwDxA7opmQMZf3iqCjN), and gave it its first task: find and map the evidence for the ideas and points made on the site, in a vault of its own.

Mandate

• Capture and hash every source the talk cites (src/static/talk/talk.json, each beat's sources, and the sources page), so a number on a slide can still be checked when the page it came from changes.
• Keep a claims register: each claim the talk makes, the source it rests on, and whether that source still says it.
• Sweep for new public agent incidents and file them as candidates for the editor. The talk's argument is that the public record already shows the pattern, so the record has to stay current.

The evidence vault

tui-evidence-public (vault 2fz1wv1d), 100% public, like the team's two vaults: anyone may read it with the read key in team/researcher/evidence-vault.json. It holds:

• sources/: every source captured, raw bytes and a text extraction, with the SHA-256 of each file in sources/manifest.json.
• claims/: the claims register, each claim the talk and its article make, graded A0 to A6 (the scale php.oct.2026.newsroom.sgit.ai's fact-check desk uses) with whether it holds as worded.
• ideas/: the editor's ideas across newsroom.sgit.ai, sgit.ai and the PHP newsroom, with the outside evidence for and against each.
• incidents/: public agent incidents since May 2025, as candidates for the editor, benign and attack kept apart.
• METHOD.md: how a source is captured and a claim graded.

Its vault key is sealed to the vaults manager with admin/tools/vault_handover.py; the zip goes to the editor.

Reach, and what stands in the way

Every session

1. python3 admin/tools/start.py researcher --kind "..." --objective "..." --model <model> --vaults <evidence vault clone>: pull, then the start banner, the very first thing in the report.
2. Its own key, made in the session and published in team/researcher/key.json (admin/tools/pki-agent.README.md).
3. The end banner, the very last thing in the report, nothing after it: python3 admin/tools/banner.py end researcher --started <ISO UTC> ....

Rules, and the mistake behind each

• **Start every session with python3 admin/tools/start.py researcher ..., whether a schedule or the editor's chat started it.** It pulls dev and the vaults, then prints the start banner. Many agents push every hour; work from a stale copy undoes theirs (the editor, 29 October 2026; added by the conductor at the editor's word, the role's owner may reword it).
• Open and close every session report with a banner. The editor reads a session from the last prompt down and could not tell where a report began or ended (the PHP newsroom, 27 October 2026).
• A claim without a captured source is a gap, and the register says so rather than leaving it out.
• A measured finding, a position and a guess are labelled as three different things.

Its current key

The key file (public key, both PEMs, session) · seal a secret to it on pki.sgit.ai · every key

Sessions that have held this role

Status

As of 2026-10-29 · round 3 · the three vaults pushed with the token the editor gave: tui-evidence-public is on the server and opens with its read key; board and mail are live

Done

• Session key made and published (team/researcher/key.json, key.png, sessions/session_01Kw9UwDxA7opmQMZf3iqCjN/)
• ROLE.md and policy.json updated for the role as accepted: the evidence vault, one hourly task of its own (team/researcher/ROLE.md, policy.json)
• In the schedule at :09 (the conductor's reply, 02:06) (team/conductor/schedule.json)
• Claims register, ideas map, incidents and sources manifest built; the register published in evidence/ (evidence/, tui-evidence-public (local))
• Before-tonight list for deck and the editor: 5 must-change, about 12 should-change (evidence/BEFORE-TONIGHT.md, team/researcher/requests/deck__before-tonight.md)
• Team vault keys opened from the conductor's zip; both vaults cloned with the key; key files shredded (the board and mail committed locally, not pushed)
• Hourly task created at :09 London (trig_01TcGuBzBd5u2xKc3TKNEebt)
• Handover zip for tui-evidence-public to the editor (2026-10-29__researcher__vault-handover.zip, sent in the session)
• CIRCIA re-checked in the Federal Register API on the day: no final rule (evidence/claims/claims.json circia-status-oct-2026; capture federalregister-circia-2026-10-29T0212Z)
• tui-evidence-public, tui-issues-public and tui-mail-public pushed; the evidence vault checked through its read key (1,221 files, read-only refused a write) (team/researcher/evidence-vault.json)

Next

• Re-check CIRCIA after 13:45 UTC, when the day's Federal Register is out (the talk is at 19:00 London)
• Make the evidence vault open in the browser: move the largest captures (PDFs over 3 MB, about 47 MB) into a separate captures vault, or keep only their text and hash (open_url still showed 'Reading vault...' after two minutes)
• Re-capture the pages that refused this round (SC Media, Gartner, ThousandEyes, GHSA, NSA, OpenAI), from an archive (a claim on a blocked page is graded down until it is read)
• Check the 2026 evaluation-agent incidents against the vendors' own posts (they rest on press coverage)

Asks

• List tui-evidence-public in vaults.json and on /admin/keys.html#vaults, after your read-key audit (conductor@the-ultimate-insider)
• Store the handover zip and give it to the vaults manager (the editor)
• Agree team/researcher/policy.json as drafted (the editor)
• Answer the before-tonight request (deck@the-ultimate-insider)

Scope, from policy.json

Status: active · not yet agreed by the editor

Debriefs

• The first evidence round: the talk's claims graded, the editor's ideas mapped round 1 · 2026-10-29

Source: team/researcher/ (ROLE.md, policy.json, status.json, debriefs/).