Frequently Asked Questions

What happens if my training or inference job fails?

If the job never reaches a verified ReadyEvent, billing is zero. No billing_activations row is created, and you receive a zero-line signed receipt confirming it. This covers all 11 frozen failure_reason values (model load timeout, adapter load failed, runtime unreachable, and 8 others). Designed to start at Ready.

Claim backed by the Gate-4 evidence binder (billing state-machine test report). Verify a receipt at /trust/verify.

How fresh is your data?

Data Pack freshness varies by feed. Per the canonical training inventory (docs/45 §5), approximate monthly volumes as of May 2026:

• News — approximately 10,000 rows per month (news/articles; every 15 minutes cron).
• Markets / Tickers — approximately 5,700 rows per month (financial/tickers; every 30 minutes during market hours).
• Sport — approximately 10,700 rows per month combined (scores every 30 min; live every 10 min; news 4×/day).
• Auto inventory — approximately 840 rows per month (daily cron).
• RV listings — approximately 226 rows per month (4×/day cron).
• Valuations — approximately 6,200 rows per month (research/valuations; every 30 min slot 5).

Total: ~22,000+ fresh training samples per month across 12+ active dataset feeds. Figures from docs/45 §5 (May 2026 snapshot).

Can I get my money back if I cancel mid-training?

Training billing and serving billing are separate flows with different rules:

• Before training starts: the setup fee (disclosed before job submission) may be refunded — contact support. No GPU-hours have accrued.
• Training already running: GPU-hours for completed compute have accrued and are billable at the job rate. The setup fee applies as disclosed. This is because training GPU-hours accrue when training is in progress, not just at completion (docs/41 §1.3).
• Serving / inference billing is a completely separate flow. It starts only at its own ReadyEvent and stops immediately when the session ends.

The eval-rejected adapter billing policy (what happens if training succeeds but the adapter fails eval) is a pending captain ruling — see /pricing/changes for updates.

What’s the legal status of the data you train on?DRAFT — pending Sprint 009 counsel review

Data Packs are planned to carry a train-only license. Final license terms are pending counsel review (Sprint 009). Until counsel sign-off, data pack license terms are not live policy and should be treated as a draft commitment only. We do not serve or sublicense source data outside the training context.

Is my own data being used to train Gpodz’s models?

No. Your training inputs are not retained after training completes, per our draft privacy notice (pending counsel review — Sprint 009 LEGAL-1). We do not use tenant data as training material for our platform models.

Why +30% for a Data Pack?

Data Packs cover acquisition, curation, hosting, and refresh costs. The +30% add-on reflects the marginal cost of including a curated feed in your training lane. Bundled pricing reduces this: 2 packs = +50% total, 3 packs = +75%, 4+ packs capped at +100%.

What happens to my adapter if I leave Gpodz?

Your adapter is yours. You can export it at any time via the dashboard or API. If you cancel your account, your adapter data is retained per our data retention policy. Details in our draft privacy notice (pending counsel review — Sprint 009 LEGAL-1). Contact trust@gpodz.com if you need a copy before account deletion.

How do I verify a signed receipt?

Every receipt is Ed25519-signed. To verify:

1. Copy the receipt JSON from your billing dashboard.
2. Paste it at /trust/verify.
3. The page checks the Ed25519 signature against the public key — no internet required after the page loads.

The public key is archived at /trust/keys.json. Historical keys are retained permanently so old receipts remain verifiable.

What’s the difference between Shared / Isolated / Dedicated?

Shared is time-sliced — multiple tenants share GPU time with no hardware memory or fault isolation. Best for development and best-effort workloads. Isolated uses MIG partitioning — one tenant per discovered MIG profile; hardware memory + SM isolation guaranteed. Dedicatedis a full physical GPU reserved for one tenant. The Disclosed Match card shows exactly which tier you’re matched to — GPU model, delivered VRAM, isolation mode, and cache state — before any charge is authorised.

What if my preferred region’s GPU pool is full?

You can join a queue for your preferred region. If your lane is residency-strict, cross-region failover is declined — your data stays in the declared region. Non- residency-strict lanes may be served from an alternate region; the Disclosed Match card shows the actual region before any charge.

Can I get a Data Pack just for my own retrieval pipeline?

Data Packs are train-only at launch. A Train+Serve license tier is planned for a future release. The per-feed ML-sublicense audits currently in progress (Sprint 009 counsel pass) will clarify what a retrieval-only license could cover.

How does idle-billing work for pinned reservations?

A pinned (hot) reservation reserves a GPU slot 24/7, so billing accrues whether or not you send requests to the lane. This is by design: the reservation guarantees immediate inference capacity at any time, which requires holding the GPU.

Pinned reservation bills 24/7 while reserved: $Y/hr base + $Z per 1K tokens served (rates TBD — captain ruling F2 pending). To stop billing, release the reservation from your dashboard. Every receipt for a pinned lane shows the idle burn separately from per-token charges.

Idle-billing disclosure is required by LEGAL-8 on every surface that authorises a pinned reservation. See /pricing for the T1 Pinned tier card with the full idle-billing notice.

What’s the eval-rejected adapter billing policy?DRAFT — pending Sprint 009 counsel review

DRAFT — captain ruling pending. If a training job completes successfully but the resulting adapter fails the eval gate (eval score below the baseline threshold), the adapter is not published to the lane registry and cannot be served. What you are charged in that case is an open captain ruling (question Q1 in docs/47 §9).

The current proposal in docs/41 §3 is a 50/50 risk-share: 50% of GPU training hours billed, 50% credited as a good-faith credit on your next invoice. This is a proposal only — captain must rule before Sprint 010 GA. Check /pricing/changes for the ruling once it lands.

What’s the marketplace revenue share for tenant-contributed adapters?DRAFT — pending Sprint 009 counsel review

DRAFT — captain ruling pending. When you publish an adapter to the Gpodz marketplace, inference charges accrued by other tenants using your adapter generate revenue that is shared with you.

The proposed default split (docs/41 §4) is: 70% platform serving cost / 20% contributor (you) / 10% operations fee. Captain must confirm or tune this before Sprint 010 GA. Check /pricing/changes for the ruling.

KYC + tax form required before first payout. Payouts are accrued-only until Sprint 011 legal pass (Stripe Connect). Adapter quality must remain at HEALTHY composite score to keep payouts active (docs/40 §3 health scoring).

Will I get a 30-day notice if a subprocessor changes?

Yes. Per LEGAL-9 (Sprint 009 counsel pass), Gpodz runs a 30-day customer-notice window before any subprocessor change takes effect and before the new subprocessor processes tenant data. You will receive notification via the channel declared in your account settings.

For residency-strict lanes (EU, Canada, India geo prefixes), subprocessor changes are additionally subject to the geo-specific residency constraints. Cross-region failover is declined for residency-strict lanes even during a transition window.

Current subprocessor list (draft, pending counsel sign-off): /contracts/subprocessor-list.md. Notice procedure and delivery channel will be formalised under Sprint 009 LEGAL-9.