cavendish

Lab app · Demand

Everything asked of the lab.

Seven doors, one list, at two resolutions: everything asked, and the nineteen the week is actually composed from. Plus the three logs the firm writes into — what it asked, what it wrote on what it read, and who is still holding an answer that has moved.

Seven doors in one list. Pull what you are going to do; defer or let go of the rest with a reason. Nothing here asks permission for work the charter says needs none — the strip at the top is the only thing watching, and it moves as you decide.

Shares of committed work, not of a week — reading, triage and field maintenance leave no record, so the operating model's third allocation is not claimed here. Drift, not a score: pulling work to make a bar green would make the number worthless.

Committed

16 pieces · 7 people · 114d grows as you pull work in
  • MMMichael MenacheT3Let the agent break it8dT3Spec first, then agents13dT3One context, whole pod1dAI-SDLC as a repeatable practice8d
  • DYDylan DesmarchelierT3Small model, back of the store5dT4Inference under seal13dStanding up the model gateway16d
  • MLMichelle LamT3The canary suite9dEvidence behind the multi-platform claim8d
  • TBTravis BoastT3Who signed for that13dModel and tool approval with Procurement4d
  • AWAdam WitanowskiTRThe skill wiki2dGoogle Cloud model access5d
  • DDDavid DiazT3When agents flood the queue1dT3Does it still agree with itself?1d
  • ATAndrew TranT4What the assistant is worth7d
Slice by

61 asks, from every door into the lab — a detector, a question the library missed, engagements asking, a note somebody left, a correction owed, a team wanting the lab beside them, a proposal. Columns are the slice, not a workflow. Weight is computed from independent sources, value, the plan and age; nobody maintains a hand-ranked list of 61 things.

No vertical

15 · 73d · 33%

7 more, lower weight

banking

24 · 70d · 32%
  • Signal11 asking3d
    Legacy comprehension agents

    The largest pipeline value attached to any ask in the record, and the capability side is real — long-context retrieval over whole repositories moved twice this quarter. It ranks fourth in its own lane because the join is fuzzy and uncommitted, which caps convergence at the weaker side. That is the ranking behaving correctly on bad data, not a mis-weighting.

    AHMA42d left

    Folded, not merged. Each keeps its own asker, so taking this tells all of them.

    Where it came from
  • Signal5 asking3d
    Compliance-as-eval

    Source diversity is the highest in the pool — standards bodies, two regulators' consultation papers, three vendor eval releases and an internal assurance review, which is four epistemic communities converging rather than one echoing. Demand is the second-strongest ask in the record and the firm already owns the obligation side.

    TBAH42d left
    Where it came from
  • Signal5 asking3d
    AU data-residency inference

    Dropped by Amber from a banking RFP and by Aadhi 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 — the pattern with the shortest distance between a client question and a defensible answer.

    AHAV42d left

    Folded, not merged. Each keeps its own asker, so taking this tells all of them.

    Where it came from
  • Signal5 asking3d
    Estimating agentic builds

    The research run came back dry: one paper, no repos, three vendor claims with no method behind them. A dry run is a finding rather than a failure — the firm is being asked, four engagements deep, about something the literature has not written up, and that is the strongest case in the pool for original work.

    MA42d left

    Folded, not merged. Each keeps its own asker, so taking this tells all of them.

    Where it came from
  • Signal9 asking3d
    Standardising the model estate

    The most-asked question in the record and the one with the shortest half-life — any answer is wrong within a quarter, which is exactly why it belongs to a standing answer with a re-check date rather than a report. Closed objective B.1 died on this and the notes are still in the pile.

    MLDBDYDD42d left

    Folded, not merged. Each keeps its own asker, so taking this tells all of them.

    Where it came from
  • Signal4 asking3d
    The halved-timeline claim, measured

    A competitor claim we are asked to rebut or match, four engagements running. Our own evidence says the answer is neither, which is a harder sell and a better position. Ranks sixth of six in its lane and two people put it on the list by hand anyway.

    MMMA42d left
    Where it came from
  • Signal4 asking3d
    Delegated credentials for acting agents

    Named as a precondition in two banking RFPs. Answering it badly does not lose the deal — it removes agentic scope from the deal, which costs more. Points at an open field, so the work is a position, not a new field.

    AHTB42d left

    Folded, not merged. Each keeps its own asker, so taking this tells all of them.

    Where it came from
  • Signal6 asking3d
    Agent-native billing and metering

    Naming event: 'agent metering' displaced 'usage-based pricing' in band-1 changelogs from three vendors inside six weeks, and ignited in band 3 when two telco RFPs used the phrase. Independently, three losses were recorded against commercial model rather than price. Both drivers arrived at the same object without seeing each other.

    MAOV42d left

    Folded, not merged. Each keeps its own asker, so taking this tells all of them.

    Where it came from

16 more, lower weight

cross-sector

12 · 47d · 21%
  • Signal5 asking3d
    Estimating agentic builds

    The research run came back dry: one paper, no repos, three vendor claims with no method behind them. A dry run is a finding rather than a failure — the firm is being asked, four engagements deep, about something the literature has not written up, and that is the strongest case in the pool for original work.

    MA42d left

    Folded, not merged. Each keeps its own asker, so taking this tells all of them.

    Where it came from
  • Signal9 asking3d
    Standardising the model estate

    The most-asked question in the record and the one with the shortest half-life — any answer is wrong within a quarter, which is exactly why it belongs to a standing answer with a re-check date rather than a report. Closed objective B.1 died on this and the notes are still in the pile.

    MLDBDYDD42d left

    Folded, not merged. Each keeps its own asker, so taking this tells all of them.

    Where it came from
  • Signal6 asking3d
    Agent-native billing and metering

    Naming event: 'agent metering' displaced 'usage-based pricing' in band-1 changelogs from three vendors inside six weeks, and ignited in band 3 when two telco RFPs used the phrase. Independently, three losses were recorded against commercial model rather than price. Both drivers arrived at the same object without seeing each other.

    MAOV42d left

    Folded, not merged. Each keeps its own asker, so taking this tells all of them.

    Where it came from
  • Signal9 asking3d
    Role recomposition under agentic delivery

    Asked by executives, never by procurement, so it never reaches a scope document and is invisible to the pipeline. The firm has an objective for it and the lab has nothing. Nothing in the corpus matched the ask in the client's words — the candidate exists at all because a research run went and found the vocabulary.

    HB42d left

    Folded, not merged. Each keeps its own asker, so taking this tells all of them.

    Where it came from
  • Signal2 asking3d
    Canary suites that fire on provider releases

    Three silent behaviour changes detected in band 1 this quarter that no vendor changelog mentioned. This is what gives Argus teeth, and it is cheap: the suite already exists for internal evals.

    MLTB42d left
    Where it came from
  • Signal3d
    Shared-context AI surfaces for delivery pods

    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 rather than the literature, which is the strongest election signal the pool has — and no client has asked for it once.

    OVDBAH42d left
    Where it came from
  • Signal3d
    Interpretability tooling as a debugging practice

    Tooling maturity crossed a line this quarter — three open releases aimed at application engineers rather than researchers. If it holds it changes what an assurance artefact can contain; if it does not, it costs a Type R afternoon.

    TBDD42d left
    Where it came from
  • Client demand12d
    Can you run a generative AI proof of concept for us?

    Effectively gone, from fourteen a year ago. Clients have stopped buying exploration and started buying outcomes, which is the single biggest change in the record this year.

    Found by a detector84d left
    Where it came from

4 more, lower weight

retail

14 · 46d · 21%
  • Signal12 asking3d
    Agentic data remediation

    Six engagements, the clearest unmet demand in the record, and squarely inside the firm's existing competence. Industry signal is thin and scattered — the research run found the language before the detectors did, which is the shape of every demand-led item that turns out to matter.

    DBOV42d left

    Folded, not merged. Each keeps its own asker, so taking this tells all of them.

    Where it came from
  • Signal5 asking3d
    Estimating agentic builds

    The research run came back dry: one paper, no repos, three vendor claims with no method behind them. A dry run is a finding rather than a failure — the firm is being asked, four engagements deep, about something the literature has not written up, and that is the strongest case in the pool for original work.

    MA42d left

    Folded, not merged. Each keeps its own asker, so taking this tells all of them.

    Where it came from
  • Signal9 asking3d
    Standardising the model estate

    The most-asked question in the record and the one with the shortest half-life — any answer is wrong within a quarter, which is exactly why it belongs to a standing answer with a re-check date rather than a report. Closed objective B.1 died on this and the notes are still in the pile.

    MLDBDYDD42d left

    Folded, not merged. Each keeps its own asker, so taking this tells all of them.

    Where it came from
  • Signal4 asking3d
    The halved-timeline claim, measured

    A competitor claim we are asked to rebut or match, four engagements running. Our own evidence says the answer is neither, which is a harder sell and a better position. Ranks sixth of six in its lane and two people put it on the list by hand anyway.

    MMMA42d left
    Where it came from
  • Signal9 asking3d
    Role recomposition under agentic delivery

    Asked by executives, never by procurement, so it never reaches a scope document and is invisible to the pipeline. The firm has an objective for it and the lab has nothing. Nothing in the corpus matched the ask in the client's words — the candidate exists at all because a research run went and found the vocabulary.

    HB42d left

    Folded, not merged. Each keeps its own asker, so taking this tells all of them.

    Where it came from
  • Signal3d
    Shared-context AI surfaces for delivery pods

    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 rather than the literature, which is the strongest election signal the pool has — and no client has asked for it once.

    OVDBAH42d left
    Where it came from
  • Question2 asking5d
    Client wants a fixed price on an agent build. What do we tell them?

    Third distinct asker in a fortnight. The pipeline value attached to this is the largest of any unanswered question in the log.

    DBAH55d left

    Folded, not merged. Each keeps its own asker, so taking this tells all of them.

    Where it came from
  • Question2 asking5d
    Will agents clean the warehouse or just make more mess in it?

    Same gap. Two verticals, two vocabularies, no join.

    DBSL50d left

    Folded, not merged. Each keeps its own asker, so taking this tells all of them.

    Where it came from

6 more, lower weight

telco

15 · 41d · 19%
  • Signal11 asking3d
    Legacy comprehension agents

    The largest pipeline value attached to any ask in the record, and the capability side is real — long-context retrieval over whole repositories moved twice this quarter. It ranks fourth in its own lane because the join is fuzzy and uncommitted, which caps convergence at the weaker side. That is the ranking behaving correctly on bad data, not a mis-weighting.

    AHMA42d left

    Folded, not merged. Each keeps its own asker, so taking this tells all of them.

    Where it came from
  • Signal12 asking3d
    Agentic data remediation

    Six engagements, the clearest unmet demand in the record, and squarely inside the firm's existing competence. Industry signal is thin and scattered — the research run found the language before the detectors did, which is the shape of every demand-led item that turns out to matter.

    DBOV42d left

    Folded, not merged. Each keeps its own asker, so taking this tells all of them.

    Where it came from
  • Signal5 asking3d
    Estimating agentic builds

    The research run came back dry: one paper, no repos, three vendor claims with no method behind them. A dry run is a finding rather than a failure — the firm is being asked, four engagements deep, about something the literature has not written up, and that is the strongest case in the pool for original work.

    MA42d left

    Folded, not merged. Each keeps its own asker, so taking this tells all of them.

    Where it came from
  • Signal9 asking3d
    Standardising the model estate

    The most-asked question in the record and the one with the shortest half-life — any answer is wrong within a quarter, which is exactly why it belongs to a standing answer with a re-check date rather than a report. Closed objective B.1 died on this and the notes are still in the pile.

    MLDBDYDD42d left

    Folded, not merged. Each keeps its own asker, so taking this tells all of them.

    Where it came from
  • Signal4 asking3d
    The halved-timeline claim, measured

    A competitor claim we are asked to rebut or match, four engagements running. Our own evidence says the answer is neither, which is a harder sell and a better position. Ranks sixth of six in its lane and two people put it on the list by hand anyway.

    MMMA42d left
    Where it came from
  • Signal6 asking3d
    Agent-native billing and metering

    Naming event: 'agent metering' displaced 'usage-based pricing' in band-1 changelogs from three vendors inside six weeks, and ignited in band 3 when two telco RFPs used the phrase. Independently, three losses were recorded against commercial model rather than price. Both drivers arrived at the same object without seeing each other.

    MAOV42d left

    Folded, not merged. Each keeps its own asker, so taking this tells all of them.

    Where it came from
  • Question2 asking5d
    Which model should we standardise on across the estate?

    Sovereignty constraint on top of the model question — the answer changes if the model must be hosted on-shore, and nothing joins the two.

    AVMA54d left

    Folded, not merged. Each keeps its own asker, so taking this tells all of them.

    Where it came from
  • Correction2 asking1d
    Tell p-rohan — Board pack, AI spend section — quoted as a band, not a point

    Used it as briefing, and it is still in front of somebody.

    HBMA19d left

    Folded, not merged. Each keeps its own asker, so taking this tells all of them.

    Where it came from

7 more, lower weight

government

10 · 34d · 15%
  • Signal11 asking3d
    Legacy comprehension agents

    The largest pipeline value attached to any ask in the record, and the capability side is real — long-context retrieval over whole repositories moved twice this quarter. It ranks fourth in its own lane because the join is fuzzy and uncommitted, which caps convergence at the weaker side. That is the ranking behaving correctly on bad data, not a mis-weighting.

    AHMA42d left

    Folded, not merged. Each keeps its own asker, so taking this tells all of them.

    Where it came from
  • Signal5 asking3d
    Compliance-as-eval

    Source diversity is the highest in the pool — standards bodies, two regulators' consultation papers, three vendor eval releases and an internal assurance review, which is four epistemic communities converging rather than one echoing. Demand is the second-strongest ask in the record and the firm already owns the obligation side.

    TBAH42d left
    Where it came from
  • Signal5 asking3d
    AU data-residency inference

    Dropped by Amber from a banking RFP and by Aadhi 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 — the pattern with the shortest distance between a client question and a defensible answer.

    AHAV42d left

    Folded, not merged. Each keeps its own asker, so taking this tells all of them.

    Where it came from
  • Signal9 asking3d
    Standardising the model estate

    The most-asked question in the record and the one with the shortest half-life — any answer is wrong within a quarter, which is exactly why it belongs to a standing answer with a re-check date rather than a report. Closed objective B.1 died on this and the notes are still in the pile.

    MLDBDYDD42d left

    Folded, not merged. Each keeps its own asker, so taking this tells all of them.

    Where it came from
  • Signal6 asking3d
    Agent-native billing and metering

    Naming event: 'agent metering' displaced 'usage-based pricing' in band-1 changelogs from three vendors inside six weeks, and ignited in band 3 when two telco RFPs used the phrase. Independently, three losses were recorded against commercial model rather than price. Both drivers arrived at the same object without seeing each other.

    MAOV42d left

    Folded, not merged. Each keeps its own asker, so taking this tells all of them.

    Where it came from
  • Signal9 asking3d
    Role recomposition under agentic delivery

    Asked by executives, never by procurement, so it never reaches a scope document and is invisible to the pipeline. The firm has an objective for it and the lab has nothing. Nothing in the corpus matched the ask in the client's words — the candidate exists at all because a research run went and found the vocabulary.

    HB42d left

    Folded, not merged. Each keeps its own asker, so taking this tells all of them.

    Where it came from
  • Question2 asking5d
    Which model should we standardise on across the estate?

    Sovereignty constraint on top of the model question — the answer changes if the model must be hosted on-shore, and nothing joins the two.

    AVMA54d left

    Folded, not merged. Each keeps its own asker, so taking this tells all of them.

    Where it came from
  • Question2 asking5d
    What do we say when a client asks for AU data residency on inference?

    Footnote-level. Two sector owners hit this in a fortnight and the field has not caught up.

    MAAV49d left

    Folded, not merged. Each keeps its own asker, so taking this tells all of them.

    Where it came from

2 more, lower weight

insurance

8 · 26d · 12%
  • Signal12 asking3d
    Agentic data remediation

    Six engagements, the clearest unmet demand in the record, and squarely inside the firm's existing competence. Industry signal is thin and scattered — the research run found the language before the detectors did, which is the shape of every demand-led item that turns out to matter.

    DBOV42d left

    Folded, not merged. Each keeps its own asker, so taking this tells all of them.

    Where it came from
  • Signal5 asking3d
    Compliance-as-eval

    Source diversity is the highest in the pool — standards bodies, two regulators' consultation papers, three vendor eval releases and an internal assurance review, which is four epistemic communities converging rather than one echoing. Demand is the second-strongest ask in the record and the firm already owns the obligation side.

    TBAH42d left
    Where it came from
  • Signal4 asking3d
    Delegated credentials for acting agents

    Named as a precondition in two banking RFPs. Answering it badly does not lose the deal — it removes agentic scope from the deal, which costs more. Points at an open field, so the work is a position, not a new field.

    AHTB42d left

    Folded, not merged. Each keeps its own asker, so taking this tells all of them.

    Where it came from
  • Signal2 asking3d
    Canary suites that fire on provider releases

    Three silent behaviour changes detected in band 1 this quarter that no vendor changelog mentioned. This is what gives Argus teeth, and it is cheap: the suite already exists for internal evals.

    MLTB42d left
    Where it came from
  • Question3 asking5d
    Which framework do the insurers actually audit against?

    Third asker, third vertical, in the same fortnight. Nothing in the corpus.

    SLTBAH54d left

    Folded, not merged. Each keeps its own asker, so taking this tells all of them.

    Where it came from
  • Correction3 asking1d
    Tell p-hana — Health insurer claims-assistant proposal, architecture section

    Used it as proposal, and it is still in front of somebody.

    SLOV16d left

    Folded, not merged. Each keeps its own asker, so taking this tells all of them.

    Where it came from
  • Signal3d
    Real-world evidence and outcomes analytics

    The nearest adjacency rather than the most exciting one — this is recognisably the firm's existing competence in a vertical it does not serve, where the pharma R&D cluster above is neither. Sponsored by Sylvia against TH-1, with a decision point in Q1 FY28.

    SL42d left
    Where it came from
  • Question5d
    Has anyone tried Gemini for claims triage?

    Logged as an intake event. No run on this model against this task; the nearest evidence is a different model on a different corpus and saying so is more useful than guessing.

    SL54d left
    Where it came from

health

5 · 17d · 8%
  • Signal5 asking3d
    AU data-residency inference

    Dropped by Amber from a banking RFP and by Aadhi 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 — the pattern with the shortest distance between a client question and a defensible answer.

    AHAV42d left

    Folded, not merged. Each keeps its own asker, so taking this tells all of them.

    Where it came from
  • Question2 asking5d
    Will agents clean the warehouse or just make more mess in it?

    Same gap. Two verticals, two vocabularies, no join.

    DBSL50d left

    Folded, not merged. Each keeps its own asker, so taking this tells all of them.

    Where it came from
  • Signal3d
    Deployment-time weight updates

    Orphan density: nineteen signals in the quarterly survey fitting no existing field and developing internal coherence around post-deployment adaptation. Cost of being wrong is the highest in the pool — it invalidates the assurance position, not just a recommendation — and that axis is deliberately kept apart from impact.

    ML42d left
    Where it came from
  • Signal3d
    Design-of-experiment agents in wet-lab pipelines

    Strongest capability signal outside the served verticals, and it has no sponsor — which means it belongs to nobody and dies quietly the moment the room gets busy. The standing caution applies with force: the trend being real is necessary and nowhere near sufficient, and the entry cost here is credibility the firm does not hold.

    DD42d left
    Where it came from
  • Signal3d
    Real-world evidence and outcomes analytics

    The nearest adjacency rather than the most exciting one — this is recognisably the firm's existing competence in a vertical it does not serve, where the pharma R&D cluster above is neither. Sponsored by Sylvia against TH-1, with a decision point in Q1 FY28.

    SL42d left
    Where it came from

energy

4 · 12d · 5%
  • Signal11 asking3d
    Legacy comprehension agents

    The largest pipeline value attached to any ask in the record, and the capability side is real — long-context retrieval over whole repositories moved twice this quarter. It ranks fourth in its own lane because the join is fuzzy and uncommitted, which caps convergence at the weaker side. That is the ranking behaving correctly on bad data, not a mis-weighting.

    AHMA42d left

    Folded, not merged. Each keeps its own asker, so taking this tells all of them.

    Where it came from
  • Signal3d
    Agents against formal specifications

    Split signal: the verification literature and the agent literature have started citing each other after two years of not doing so. Cross-community citation is the earliest structural tell the survey has, and it precedes the naming event rather than following it.

    Found by a detector42d left
    Where it came from
  • Signal3d
    Learned world models for operational planning

    Cross-pollination candidate rather than a field on its own: enabling join to network operations, where the constraint has always been simulator fidelity rather than the optimiser. Serves no stated objective, which is optionality or drift and the plan cannot yet tell which.

    DY42d left
    Where it came from
  • Signal3d
    Distributed flexibility forecasting

    Regulatory signal ahead of technical signal — three market-body consultations inside a quarter, which is the pattern the detectors weight highest in verticals with a regulator. No sponsor, no engagement history, no way for Engel to see it at all.

    Found by a detector42d left
    Where it came from

defence

1 · 3d · 1%
  • Signal3d
    Agents against formal specifications

    Split signal: the verification literature and the agent literature have started citing each other after two years of not doing so. Cross-community citation is the earliest structural tell the survey has, and it precedes the naming event rather than following it.

    Found by a detector42d left
    Where it came from

pharma

1 · 3d · 1%
  • Signal3d
    Design-of-experiment agents in wet-lab pipelines

    Strongest capability signal outside the served verticals, and it has no sponsor — which means it belongs to nobody and dies quietly the moment the room gets busy. The standing caution applies with force: the trend being real is necessary and nowhere near sufficient, and the entry cost here is credibility the firm does not hold.

    DD42d left
    Where it came from