Price period
One stretch during which a price did not move. The history unit.
Where an observation is a reading, a period is the claim built from readings: this sku cost this much, for this dimension, from here to there.
Fields
From witness.price_period:
| Field | Meaning |
|---|---|
skuId, dimensionId | The cell this period belongs to |
amountUsd | The price during the period |
validFrom | When it started |
validTo | When it ended, null while current |
firstObservationId | The reading that opened it |
lastObservationId | The most recent reading confirming it |
confirmed | Whether a second consistent read has confirmed it |
flap | Whether the value moved and moved back |
validTo IS NULL is the tape
The current price of any cell is its one open period. A partial unique index enforces exactly one open row per (skuId, dimensionId), so the record cannot hold two current prices for one thing. That is a constraint rather than an assumption, because the alternative is a tape that is quietly self- contradictory.
A second constraint requires validTo >= validFrom, so a period cannot run backwards.
Confirmation
confirmed is what separates a reading from a price. A figure seen once is recorded, but a move is published only on the second consistent read. One malformed page, one partial deploy on the provider's side, or one transient parse failure therefore does not enter the log as a price change.
The cost is latency: a real price cut shows up on the second read rather than the first. That is the trade the record makes on purpose, because a log that is occasionally wrong is worth less than one that is slightly late.
flap
Set when a value moved and moved back. Providers do deploy a wrong price and withdraw it, and a record that reported both moves as decisive would be reporting noise as signal.
Reading history
curl "https://vitna.ai/api/v1/skus/gpt-5-pro/history?dimension=input_per_mtok"History exists only in the live record. A checkout serving the committed snapshot answers with the sku, an empty periods array, and a meta.note saying why, rather than inventing a single period out of the current price.
