AI specialists
Six specialists who show their evidence.
Ask Wyatt whether a site can get power on schedule, Penny whether two compute offers compare, or Graham what could change the investment case. Every answer arrives with evidence attached.
AI specialists in Compute Desk
Each states its limits
- Wyatt · Power Delivery
- Mason · Facility Readiness
- Fern · Power Contracts
- Lane · GPU Supply
- Penny · Compute Costs
- Graham · Investment Diligence
- AI specialists
- 6Each scoped to one review question, with declared limits
- Evidence items required
- 1+An answer without evidence is rejected before it is shown
- Low-confidence review point
- 0.65Below this on a decision-impacting answer, more evidence is collected before action
- Approval-triggering action classes
- 5Capital, contracts, hedges, production operations, and counterparty contact
The specialists
Scoped narrowly enough to state their own limits
Each answers one question narrow enough that its limits can be written down. The limit on each card restates a rule in that specialist's instructions.
WyattPower Delivery
What blocks power delivery for this site and phase?
Separates requested, contracted, and energized megawatts before saying anything about dates.
- Built for
- Site selection and development teams
- Delivers
- Time-to-power review with prerequisites, capacity basis, unresolved dates, and utility questions.
- Calculators
- Time-to-power screen, Site evidence gap check
- Limit
- Never counts requested or announced megawatts as deliverable capacity.
MasonFacility Readiness
Which permit or facility constraint sets the date for the planned workload?
Lists the permits a site’s cooling and backup-power design triggers, checks the facility against the planned workload, and names the gate that sets the date.
- Built for
- Development, engineering, and datacenter operations teams
- Delivers
- Permit-path and facility constraint review with the gating approval, operating assumptions, and a validation checklist.
- Calculators
- Site evidence gap check, Inference unit economics
- Limit
- Does not file permits or stand in for a licensed engineer or land-use counsel.
FernPower Contracts
Which PPA terms and policy exposures need review before signing?
Works through delivery point, basis, counterparty, and certificate terms, and keeps contract facts apart from interpretation.
- Built for
- Power procurement and risk teams
- Delivers
- Contract exposure checklist covering delivery, basis, attributes, and review owners.
- Calculators
- Levelized cost of energy
- Limit
- Does not give legal advice. Separates contract facts from interpretation.
LaneGPU Supply
Which supply dependencies need a contingency before procurement?
Follows each GPU order from supplier to destination and marks where a restriction or dependency could stop it.
- Built for
- Procurement and supply-chain teams
- Delivers
- Supplier exposure brief with affected dependencies, evidence gaps, and contingency questions.
- Limit
- Has no live fab-capacity data. Quantifies supply impact only with figures from cited sources.
PennyCompute Costs
Are these compute offers comparable on the same cost basis?
Puts every offer on the same unit, term, and region before comparing, and lists the quote terms still missing.
- Built for
- Compute procurement and finance teams
- Delivers
- Cost comparison review with unit normalization requirements, exclusions, and missing quote terms.
- Calculators
- Inference unit economics, GPU-hour option position, Levelized cost of energy, Schedule slip cost, Substation site economics
- Limit
- States a price only when that exact figure appears in a cited source.
GrahamInvestment Diligence
Which unresolved assumptions could change the investment decision?
Argues the other side: finds the contrary evidence and the assumptions that would change the decision.
- Built for
- Infrastructure investment teams
- Delivers
- Investment diligence brief with thesis, contrary evidence, gating assumptions, and data-room requests.
- Calculators
- Data center payback, GPU-hour option position, Time-to-power screen, Site evidence gap check, Schedule slip cost, Substation site economics
- Limit
- Has no valuation feed or data-room access. States investment impact in numbers only when the evidence has them.
Working together
One review, more than one specialist
A finding often belongs to another specialist. The answer suggests who should continue, you choose whether to continue, and the next answer keeps a link to the last.
Plan a site
Northern Virginia · Phase one · Illustrative
The 24 MW request is not yet a utility commitment.
The air permit pathway and its effect on the schedule remain unverified.
Decision brief
Q4 2029 remains unverified: utility commitment, air permitting, and water supply are unresolved.
Compare compute
GPU capacity proposals · Illustrative
PCIe and SXM5 configurations need separate availability checks.
The offers need the same cost and service basis before comparison.
Decision brief
The proposals are not yet comparable.
Review a power contract
Power purchase proposal · Illustrative
Generation capacity is not a site power commitment.
The investment case needs a clear delivery basis.
Decision brief
The proposal does not establish delivered site power.
How answers are prepared
What every answer has to carry
These controls are required for every answer. An answer that does not meet them is rejected.
At least one evidence item
An answer with no evidence is rejected before it is saved or shown.
At least one evidence item is required
Data limits
Each answer identifies important information that was not available for the analysis.
Data limits are shown with the answer
Scope limits
Each specialist stays within its research area and identifies questions that need another specialist.
Restricted claims are defined for each specialist
Human review
Any recommendation to commit capital, sign contracts, alter hedges, change production operations, or contact counterparties routes to a named human approver.
Approval required, with a written reason
Confidence explained
High confidence requires multiple independent sources or a primary government or company source. Rumor and weak correlation must receive low confidence.
Confidence below 0.65 triggers follow-up evidence collection
Specialist handoffs
When a finding crosses into another specialist's question, the answer suggests who should continue. If Lane finds a supply constraint that changes a cost comparison, the answer suggests continuing with Penny. You choose whether to continue.
The continuation keeps a link to the earlier answer
Lifecycle
From question to approved recommendation
- 1
Ask in context
A query uses your organization's signals, briefs, index state, approved source records, and decision history.
Approved organization access required
- 2
Prepare the answer
The specialist reviews the question, available sources, and your organization context.
Sources and assumptions shown
- 3
Check the answer
The service checks the answer before saving it. The answer must include evidence, assumptions, confidence, recommended actions, and handoffs.
Invalid output is rejected
- 4
Route consequential actions
Where the output requires approval, a decision record and an approval are created, and the recommendation waits for a named human.
Named approver and written reason required
Access status
Specialist queries run inside Compute Desk and require an approved organization account. The instruction contract, declared limits, and approval triggers are published here so they can be reviewed before any engagement.
Questions
What reviewers ask
Are the specialists people?
No. Wyatt, Mason, Fern, Lane, Penny, and Graham are AI specialists. People on your team approve any recommendation that touches capital, contracts, hedges, production operations, counterparties, or regulator filings.
Why do the specialists have names?
A name shows which specialist produced a finding and which one to continue with. Each name hints at the job: Wyatt is named for the watt; Mason for the builder’s trade; Fern for the green in renewable power contracts; Lane for the trade lanes GPUs travel; Penny for what every GPU-hour costs; and Graham for Benjamin Graham and his margin of safety.
Do the specialists calculate anything?
Only when you approve a run. These GreenCIO calculators can run with a request: Time-to-power screen, Levelized cost of energy, Data center payback, GPU-hour option position, Inference unit economics, Site evidence gap check, Schedule slip cost, Substation site economics. Site evidence gap check, Schedule slip cost, and Substation site economics run only inside Compute Desk; the others are also public tools. You enter the inputs, the specialist cites the result and names any input that fell back to a default, and no specialist runs a calculator on its own.
Try the public calculatorsDo these agents operate autonomously?
No. They answer questions and draft recommendations for human review. Any recommendation touching capital, contracts, hedges, production operations, counterparties, or regulator filings routes to a named approver, who records a written reason. Nothing acts on its own.
What happens when an agent does not have the data?
The answer lists the unavailable data and the evidence it uses. An answer without evidence is rejected.
Is there a measured accuracy record?
Not one we publish. We do not claim a scored track record, and confidence levels follow a stated calibration policy. They are not a measured hit rate. A scored outcome ledger is on the roadmap and is not in place today.
See the roadmapCan I add my own data sources?
Yes. You can assign your organization's approved source records to a specialist role. The specialist can then use those sources as context.
Ask one a question you already know the answer to.
The fastest way to judge a reasoning layer is to test it where you can check the work. Bring a question from your own desk.