> Source: https://vitna.ai/docs/cli/aperture/private
> Raw Markdown: https://vitna.ai/docs/cli/aperture/private.md
> Corpus index: https://vitna.ai/llms.txt

# Your own infrastructure

Registering something this organisation runs that no provider publishes a price
for, and entering what it costs.

A company-hosted deployment, an internal fine tune, a regional vendor under
contract, an owned cluster. Until these commands existed, a workload could only
name infrastructure Vitna could observe, which is a narrower thing than the
infrastructure an organisation can actually use.

## Register one

```bash
npm run aperture:private:add -- --id acme/vllm-prod-a --name "vLLM, Llama 3.3 70B, 4xH100" --class model
```

| Flag | Meaning |
|---|---|
| `--id` | `org/slug`, the same shape a workload id has. The org half is what scopes it |
| `--name` | What a person reading a brief will see |
| `--class` | The market it competes in, from the ones the record already names |
| `--note` | What it actually is, in your own words. Nothing derives anything from it |

## Enter what it costs

```bash
npm run aperture:private:price -- --sku acme/vllm-prod-a --dimension gpu_hour_on_demand --amount 2.49 --from 2026-08-01
```

| Flag | Meaning |
|---|---|
| `--sku` | Which of yours |
| `--dimension` | What the rate is per. It has to be one the market quotes, or no unit profile could spend it |
| `--amount` | USD per one of that dimension's units |
| `--from` | The date the rate became true. A date in the future is refused |
| `--note` | Where the figure came from in your own world: an invoice, a contract, a capacity model |

## It is not a sku on the tape, and it never becomes one

This is the part worth being exact about, because the word "sku" invites the
opposite assumption.

A private sku lives in `aperture`, scoped to one org. The boards, the spreads,
the [reference rates](../../concepts/reference-rate.md) and the JSON Feed all
query `witness.sku`, and it is not in that table, so it cannot reach any of them
by construction rather than by a filter somebody has to remember.

Nothing about it is published. Vitna makes no claim about it to anybody.

## The price is entered, and every surface that reads one says so

There is no fetch behind an entered rate, no archived document, no sha256 and no
[attestation](../../concepts/attestation.md). What the record can say is who
entered it and when.

So a decision priced partly from one:

- names the `aperture.private_price` rows in `inputs.enteredPriceIds`, its own
  list, separate from the `witness.price_period` ids
- prints them under their own heading in the brief
- carries a sentence saying a reader cannot recompute them from anything they
  can fetch, which a witnessed price they can

A figure a reader can recompute from a fetchable document and a figure somebody
typed are not the same claim, and a brief that put them in one list would be
saying the second is the first.

## A rate is not revisable

`aperture.private_price` is append-only like every other record table. A rate
that turns out to be wrong is corrected by entering the better one from the date
it became true, and the old row keeps saying what a decision taken last quarter
was taken over.

The engine reads the newest rate effective on or before the day it runs. A rate
dated ahead of today is refused rather than entered and silently ignored,
because a person who entered one would otherwise believe the next run reads it.

## What a private sku cannot have

**A watch on the market.** Four of the six [trigger](triggers.md) kinds fire on
a `witness.market_event`: a price move, a delisting, a deprecation, a launch.
Nothing observes a private deployment, so no event will ever name one, and a
watch armed over one would sit in the register looking like a watch and be
skipped silently on every collector run. Arming one is refused.

The two that stay are the two that watch the organisation's own record rather
than the market: `evidence_added` and `date`. Those are exactly the watches a
private deployment can have.
