Product · Ask
One question, opened all the way down
Ask takes an industry question and returns an answer in which every sentence that asserts a fact points at a passage in a primary source. This page shows one such answer, the work that produced it, and the claim the validator refused to let through.
Modes
Three ways to ask, and what each one costs.
The mode is chosen before the question runs and is recorded with the answer. Quick and Deliberate differ in how long synthesis deliberates; the citation check is the same in both.
Quick
mode=speculative
Retrieval-first synthesis. Same citation validation, less deliberation.
- Wall time
- ~30 s
- Per query
- ~$0.01
Deliberate
the sample belowmode=deliberate
The full specialist pipeline — deeper retrieval, longer synthesis.
- Wall time
- 1–2 min
- Per query
- ~$0.10
Conversational
POST /v1/tutor · follow-up turns
A session you can keep asking into. Every turn is cited the same way.
- Wall time
- follow-ups
- Per query
- ~$0.01
published wall-time expectation and per-query price for each mode · not a measurement of any run
The work, made visible
What a reader watches while the question runs.
Ask streams its progress. The coordinator reports 13 named stages over the response as NDJSON frames, and each of them advances exactly one of the five rows below — so a row only ever moves because a frame said it did. Nothing is estimated forward.
01
Look for an existing answer
idempotency check against answers already produced
Wire stages that advance this row
idempotency_lookupTypical
0.4 s
02
Retrieve passages
embedding, vector search and rerank over the live corpus
Wire stages that advance this row
retrievalTypical
2.4 s
03
Synthesis
the specialist models write the cited answer
Wire stages that advance this row
specialist_llmspecialists_pass1head_composespecialists_retryTypical
60–110 s
04
Validate citations
verbatim substring check, claim by claim
Wire stages that advance this row
index_chunksvalidatevalidate_pass1validate_pass2Typical
10–20 s
05
Save the answer
persisted as a brief with its provenance manifest
Wire stages that advance this row
persistpayload_loadcitation_snapshotTypical
0.6 s
five reader rows and the 13 Ask wire stages that drive them · durations are measured medians of past deliberate runs, not this page's run
Before the first frame, the clock is the only fact
Until a frame arrives, the only real number on screen is the elapsed wall clock. The stepper is driven by the typical timings above and is labelled as such, and it is held short of the last row: a progress display may not show the answer saved before the answer has arrived.
typical timing — this run reported no stage frames
Cancel stops the waiting, not the work
Cancelling aborts the request to the API, and the abort really does reach the upstream fetch. It does not stop the pipeline: the run finishes and the answer is persisted. The work is kept, so asking the same question again returns it rather than running it a second time — the first stage below is that check.
idempotency_lookup · hit → the stored answer is returned
The answer
One answered question, with its claims.
Everything from here to the end of the withheld claim is an illustrative sample: it was written to show the format and was never produced by the engine. It carries no date, because a date would be a claim that something happened.
not engine output · not a delivery · not dated
No answer is published to the public sample endpoint. Every ad-hoc answer in the showcase workspace is agent traffic rather than content chosen for publication, and publishing whichever one is newest is the rule we removed. So the prose below is hand-authored to show the shape — the numbered claims, the sources rail, the withheld block — and the only counted numbers on this page are the validator's, at the foot.
Question
How do PJM capacity prices affect Virginia data-center buildouts?
answer · 6 citations · 6 verified · 1 withheld
PJM's Base Residual Auction sets the price Virginia load-serving entities pay for generation availability. The 2025/26 auction cleared at $269.92/MW-day RTO-wide, roughly nine times the prior year 1; the 2026/27 auction cleared at the $329.17 collar agreed in the FERC settlement 23.
Dominion Energy Virginia is the largest load zone affected. Dominion reported approximately 40 GW of contracted data-center capacity in its pipeline at year-end 2024 4, and its 2024 Integrated Resource Plan filed with the Virginia SCC projects zone load roughly doubling by 2039 5. Capacity costs are passed through to customers via rate adjustment clauses 6.
Challenge (private review aid)Add to caseExport · DOCX PPTX MD
sample answer in the signed-in register · illustrative
Sources cited
- 1PJM Interconnection2025/2026 Base Residual Auction Report2 passages
- 2PJM Interconnection2026/2027 Base Residual Auction Report1 passage
- 3FERCOrder accepting settlement, Docket EL25-461 passage
- 4Dominion EnergyForm 10-K, fiscal 20243 passages
- 5Virginia SCCDominion 2024 Integrated Resource Plan2 passages
- 6Virginia SCCRate adjustment clause filings1 passage
6 sources · 10 passages matched
sources rail as a delivered answer renders it · illustrative
Provenance
One claim, followed to the passage behind it.
Clicking a numbered claim in a delivered answer opens this chain. It is the whole product in one figure: five rungs from the sentence you read to the exact substring of the document it was matched against, each rung independently addressable.
Citation 4 · provenance chain
claim → source → document → chunk → verbatim
- 01claim
≈40 GW contracted data-center capacity, YE2024
the sentence in the answer that asserts a fact
- 02source
Dominion Energy, Inc.
the organisation that published it
- 03document
Form 10-K, fiscal 2024 · SEC EDGAR
the filing it was published in, with its identifier
- 04chunk
chunk · page · item
the locator — which chunk, which page, which item — so the passage can be found again
- 05verbatim
not shown — in a delivered answer this rung holds the exact substring of the source document that the claim was matched against
in a delivered answer this rung holds the exact substring of the chunk the claim was matched against; here there is no document behind it to quote
provenance chain for one claim · the verbatim rung is empty because this sample has no source document · illustrative
The two lower rungs carry no text here on purpose. A verbatim span is a quotation from a real filing, and writing one for a sample would put an invented quotation under a named 10-K on the page whose entire subject is that we do not do that. In a delivered answer both rungs are filled from the retrieved chunk, and the match is an exact substring — paraphrase does not pass. Citation discipline sets out the check.
What was withheld
The claim that did not make it into the answer.
A withheld claim is not a defect in the answer. It is the evidence that the check is load-bearing — a validator that never withholds anything is not being used.
Effect of 2026/27 prices on 2027 construction starts
No verbatim match found in corpus after two validation passes. The model drafted this claim; the corpus did not support it as written. It was removed from the answer rather than softened, hedged or rewritten until it passed, and the rest of the answer stands unchanged.
the sample answer's withheld claim and its two failed passes · illustrative
Counted, not illustrated
How often the validator actually withholds.
The sample above shows one withheld claim because the format has one. These are the real counts, read from the production database when this page was rendered. They are the only numbers on this page that describe something that happened.
1,733
passed validation
every claim matched a verbatim passage
queried 2026-10-19 06:03 UTC
597
withheld
no verbatim match after two passes
queried 2026-10-19 06:03 UTC
25.6%
withhold rate
of 2,330 through the validator
queried 2026-10-19 06:03 UTC
briefs through the validator, showcase workspace, all time · counted 2026-10-19 06:03 UTC
A brief counts once, in the state it finished in. We publish the withhold rate because a reader cannot judge the briefs that passed without knowing how many did not. Transparency carries the same counts beside the corpus and citation figures.
Ask is metered per query at the prices above. A question that returns nothing the corpus supports returns nothing, and says so.
Start free How a claim gets made