Ten ranked items. Ten minutes.

Agents that hold one shared context for a whole pod beat per-person assistants on consistency, and the surface is the product, not the model.

Why ranked · Seven independent placements in eight weeks, three from firm-side people who had not seen each other's drops; x-multiplayer-surface is already proposed against it. Velocity is the firm's own usage, not the literature, which is the strongest election signal the pool has.

Contractual AU data residency for frontier-model inference becomes a procurement gate in banking and government before the frontier labs actually deliver it.

Why ranked · Dropped by Claire from a banking RFP and by Aisha from a government panel refresh in the same fortnight, each asking whether the lab had a position; it does not, beyond pos-sovereign-inference's footnote. Band-3 demand is running ahead of band-1 delivery, which is the Engel pattern the ranker weights highest.

Regulatory obligations (CPS 230, APPs, ISM controls) can be expressed as executable eval cases and run on every model or prompt change, turning assurance from a document into a canary.

Why ranked · Suggested by Lena after the July CPS 230 review accepted the routing table as evidence; Aisha seconded it from a government assurance conversation. Cross-band: two band-1 repos encoding regulatory controls as test suites appeared in August, and no existing field explains them.

Quantium's right to play in pharma is through trial-operations analytics and real-world evidence, not discovery; the AI wedge is document-heavy regulatory work.

Why ranked · The Q3 survey run flagged orphan density around clinical-trial document automation, and x-drug-dev-adjacency is proposed to map it. Hana dropped a health-client question that points the same way; demand is not-measurable-here because Engel does not see pharma, which is why the ranker holds it below the served-vertical items.

Provider deprecation cycles are shortening to under twelve months, and every client pattern pinned to a model version carries an unbudgeted migration and recalibration cost.

Why ranked · Argus flagged four deprecation notices in August across three providers, two with under ninety days' notice; the judge-calibration drift result makes each one an eval cost, not just a config change. Priya dropped it from the gateway work and an anonymous drop flagged a client pattern that broke on a silent default roll.

Trace formats for agent runs are converging on an OpenTelemetry extension, and whoever's schema wins determines what an audit log of an agent looks like for a regulator.

Why ranked · Three lab and firm engineers logged findings about incompatible trace formats from separate sessions in the same month; that is the ten-sightings pattern at a third of the way. Band-1 velocity is high on two competing repos, and the auth-broker work needs an answer to it before the audit trail can be shown to a client.

State education departments will procure AI tutoring at scale within three years, and the measurement problem — does it move outcomes — is the engagement, not the model.

Why ranked · Aisha dropped a state government market-sounding and Rohan flagged a board-level conversation in the same week; two sightings from the exec side is unusual and the ranker weights it. Held back because the field's decisional test is weak: no current Quantium engagement hinges on it inside its horizon.

Synthetic training and eval data is entering client pipelines without lineage, and the first regulator to ask 'what was this model evaluated on' will find nobody can answer.

Why ranked · Survey-run orphan: nine band-1 papers and two band-2 repos on synthetic-data watermarking and lineage that fit no current field. Jun dropped it from the Nightingale governance work and Hana from a health-data conversation; both framed it as a measurement question, which is why it ranks above the pure-research candidates.

Packaged agent skills will be distributed and paid for like packages, and the supply-chain security problem that npm took a decade to face arrives in months.

Why ranked · Naming event at band 2 — 'skill' displacing 'plugin' and 'tool' across three harness ecosystems — with one band-1 registry launch. Sam dropped it from the SDLC work; no firm-side sighting yet, and the ranker discounts candidates with no demand-side signal.