The pilot's team lead is out until the sixth, so the nudge waits: if the thread is still quiet after Wednesday, send it then. The board still wants a fixed-fee option before the statement of work can start, and that choice is the only card in the queue with a due date. A job digest came in by forward yesterday evening and has been read behind the fence.
Rephrased from David’s allowlisted inbox — grouped by topic, summaries not excerpts. The demo carries no links to the originals.
A standards body opened a draft evaluation framework for comment, Standards Weekly reports, with a two-week window. The draft asks deployers to keep incident logs, to name a human approver for each high-impact use, and to keep red-team results as records.
Rampart Institute's weekly brief reads renewed enrichment activity at a declared site as political signalling. Its larger concern is reduced inspector access, which widens every estimate that depends on it. The brief also sets out what imagery can and cannot establish on its own.
The week’s strategic and academic-market signals, synthesized — each with a suggested move. A read, not an instruction; the paper never acts.
Public-sector AI pilots increasingly require a named human approver for each automated output before it is used.
what this suggestsWrite down who approves what in the advisory pilot before the first deliverable, and put it in the statement of work.
Evaluation roles in government and labs now ask for evidence of hands-on red-teaming and incident handling, not only policy writing.
what this suggestsKeep one worked example ready: a finding, the log entry it came from, and the change it produced.
Not printed in the paper: this is the decision queue the masthead links to, as it stood at press time, rendered by the queue's own code from the same synthetic store. The buttons here send nothing.
Separation of duties
An internal control that splits a sensitive task so that no one person can both start it and approve it: whoever requests a payment cannot also release it. Auditors check for it early in any controls review.
Lay of the land — includes items below the actionable bar, with the gap named. Estimates, not endorsements; nothing here is an action.
Small grants for independent AI evaluation workday 2
Awards up to $40,000 for reproducible evaluations of deployed models, with a public methods write-up required. Rolling review.
Applied-AI advisor for a county records officeday 16 · verify still open
Part-time engagement to review an automated records-triage pilot and set its approval and audit rules.
gap — wants prior public-sector delivery and an active clearance
Evaluation program lead at an applied-AI labday 5
Runs the lab's pre-release test program: test plans, red-team scheduling, and the sign-off record for each release.
Trust and safety operations analyst, remoteNEW
Writes detection rules, works the incident queue, and reports weekly on what was caught and what was missed.
gap — strong on detection design; the screen leads with SQL
Research assistant, nuclear-security program at a policy instituteUPDATED
Supports a project on verification and inspector access; literature reviews, data work, and briefing drafts.
gap — prefers a graduate degree in security studies or a technical field
An evaluation consortium released a shared test suite for agentic tool use, with held-out tasks scored by a team that did not build the models under test.
A model developer published its staged-release process: each capability tier waits on an external red-team report and a signed go decision before it ships.
Arms-control analysts flagged new construction at a declared enrichment site. The state has not yet declared its purpose, inspectors have not verified it, and imagery alone cannot settle it.
Answer, then say the two-sentence version a non-technical listener would follow — out loud — before you tap.
In safeguards, what is a 'significant quantity'?
What is the two-person rule?
What does it mean for a check to 'fail closed'?
Explain your own system out loud, pyramid-first (answer → because → detail). Tap to check yourself.
1. In one breath: what is this paper, and why does it exist?
2. A forwarded email arrives. What can the model that reads it do?
You named both routes and a signature for each. The enrichment signature needs care: centrifuge plants can be compact (a few thousand machines fit in a warehouse) and draw little power, so imagery alone rarely finds them. Environmental sampling for uranium particles, the conversion plant upstream, cylinder traffic and the dual-use supply chain give them away more often. Lead with that answer next time, then the reason.
Safeguards start with bookkeeping. A facility is divided into material balance areas, and the operator records every inventory change in each one: receipts, shipments and measured discards. At each physical inventory the operator measures what is actually there by weighing, sampling and assay, and inspectors verify that inventory independently and compare it with the books.
The gap between the book inventory and the physical inventory is material unaccounted for (MUF); US practice calls it inventory difference. In a bulk-handling plant it is almost never zero, because every measurement carries error and some material stays held up in equipment. What matters is whether the gap is larger than the propagated measurement error (sigma-MUF) can explain, and whether it approaches a significant quantity.
Between inventories, seals and cameras maintain continuity of knowledge: assurance that material already verified has not moved through a path the books do not cover, so it need not be verified again. Accountancy measures quantities; containment and surveillance complement it between measurements. (Synthetic demo lesson.)
Explain why a nonzero material-unaccounted-for figure is not by itself evidence of diversion, and what an inspector does when it is larger than expected. Then say it to a smart non-specialist, answer first.
answer today’s lesson →