◆ FORGE · FAQ
Why did my run's episodes all reject?
Why did my run's episodes all reject?
Because the gates are real. Every render is scored by the
action-consistency verifier (ratio, corr, mae against calibrated
thresholds) plus task-completion and identity-drift checks, and
candidates that miss any of them reject. Historical yield runs roughly
30–75%, so a 2-candidate run can honestly score 0/2, run 6 or more.
If you disagree with a specific verdict, overrule it: the episode is
kept, and your verdict is recorded on the campaign (who, what, when). The
certificate carries the machine’s verification readings. See
concepts for what each gate reads.
Why can't I review this episode yet?
Why can't I review this episode yet?
Reviews return
409 until the episode is graded: a machine accept
or reject. A verdict parked on a still-pending render would silently
flip to “kept” once something lands under that name, so ungraded episodes
can’t hold one. Wait for verification to finish, then confirm or overrule.What actually costs money?
What actually costs money?
Delivery. 1 credit per kept episode entering the export, each episode
billed once, ever on its campaign. Uploads, health checks, sampling,
runs, renders (including playground splices), and reviews are free.
Details in billing.
Can I re-deliver? Will it charge me again?
Can I re-deliver? Will it charge me again?
Yes, and no. The paid set persists on the campaign, so a re-delivery
bills only episodes kept since the last one. Re-exporting,
re-confirming, or toggling verdicts never double-bills; an unchanged kept
set re-delivers for free.
Why doesn't my other account see my campaigns?
Why doesn't my other account see my campaigns?
Campaigns are per-user. Every read is scoped to the bearer token’s
account, and a foreign campaign id answers
404: the API doesn’t
confirm that other people’s resources exist. There’s no sharing or org
layer yet; if your team needs one, talk to us.Where do described changes actually run today?
Where do described changes actually run today?
Honestly: it depends on the deployment. A described change always becomes
a real batch bound to your episode (
wb/<campaign>/<episode>). Where
the live render engine is connected, items come back as clips with
verification verdicts. Where it isn’t, the item says
“recorded (live engine pending”) and runs when the engine is on.
You’ll never get a fabricated clip standing in for a real one.Which models power which transforms?
Which models power which transforms?
Two families. New trajectories (and both playground verbs) run on the
campaign’s action-conditioned render engine: a fine-tune of the Cosmos
world-model line that obeys spliced or edited action streams; the base
model doesn’t, which is the whole point. Object swap and
environment restyle are trajectory-preserving, so they use general
video-to-video transfer models instead, swap’s backend isn’t connected
to the workbench yet, and restyle is blocked on a fidelity floor. The
Transform step shows the live readiness of each.
How do I export programmatically?
How do I export programmatically?
Deliver first, then download the archive with your bearer token:The archive is rebuilt whenever a newer delivery exists and is served
uncached, so you always get the current export. Calling it before any
delivery returns
404.What exactly is in the archive?
What exactly is in the archive?
One directory per billed episode, LeRobot-friendly:
top.mp4, left.mp4, right.mp4 (the three camera views),
actions_<fps>fps.npy (model fps) and actions_<source_fps>fps.npy
(upsampled to the source rate), and certificate.json: the provenance
record: source head and tail episodes, splice frame/blend/seed, every
verification reading with its thresholds, and a statement that the frames
were generated by the fine-tuned world model. (Your review decisions live
on the campaign record, not inside the archive.)Are there rate limits?
Are there rate limits?
No forge-specific ones yet. The play API’s per-key limiter does not apply
here; what you’ll meet in practice are Cloudflare’s platform-level
protections and the pipeline’s own pacing (renders run sequentially per
job). If you build something that hammers the API, expect limits to
arrive, with headers, like the play side’s.
Where is my data stored?
Where is my data stored?
On the forge service’s own storage: uploaded datasets, campaign state,
render artifacts, and exports live on the server’s persistent volume,
scoped to your account. Nothing is published anywhere by default, the optional Hugging Face publish step only targets a repo you name, and
only in namespaces enabled on the server. Deleting-on-request and
retention controls are talk-to-us today: contact us.
How do I get more credits?
How do I get more credits?
The wallet chip (◍) in the forge topbar sells one-time top-up packs
starting at single credits, pay for exactly the episodes you keep. The
wallet is shared with the rest of Alakazam. Running a pilot or pricing a
larger augmentation campaign? Talk to us.
Where do I see my balance, history, and usage?
Where do I see my balance, history, and usage?
Your account page,
forge.alakazam.gg/?view=account, or the avatar menu on
any forge surface. Balance with one-click top-up, the full credit ledger
(every grant and spend, newest first), daily generation charts, and each
campaign’s measured renders and cost. Programmatic:
GET /api/billing/credits, GET /api/billing/credits/history,
GET /api/workbench/account/usage.
