Paid Pilot SLO And Non-SLA Posture

Date: 2026-05-19 Status: protected-preview pilot policy Linear: TIN-1550

This page defines the reliability posture for the first TubeBrain paid design-partner pilots. It is intentionally narrower than a public SaaS SLA.

The protected hosted preview is an operator-assisted pilot surface for research packet delivery. It is not a production dependency, public subscription API, or credit-backed service. The local MCP binary remains the default fallback path.

Customer Posture

Use this wording in quotes and closeout notes:

TubeBrain's hosted preview is provided for a scheduled paid pilot proof, not as
a public SaaS SLA. We will make commercially reasonable efforts to keep the
protected preview healthy during scheduled pilot windows. If the hosted preview
is unavailable, the pilot falls back to the local MCP workflow, is rescheduled,
or is scoped to the evidence that can be produced safely.

Do not promise:

  • uptime credits, service credits, penalties, or contractual availability
  • public API availability outside scheduled pilot windows
  • multi-region or multi-replica stream durability
  • universal source availability
  • raw-audio archival
  • public managed PoToken minting
  • self-serve account, billing, deletion, or export workflows

Internal Pilot Objectives

These are operator objectives, not customer remedies.

Objective Target Evidence
Pre-call readiness All runbook gates green or exceptioned before the scheduled call docs/paid-pilot-runbook.md evidence directory
Scheduled hosted preview health /v1/health returns ok during the pilot window hosted-health.json
Hosted section proof Demo section request completes within the smoke timeout hosted-preview-smoke output or explicit exception
Stream proof Stream start, poll, list, and stop succeed for the selected radio/live source hosted-preview-smoke output or explicit source/STT exception
Admin usage proof /v1/admin/usage returns account/key counters for the pilot key admin-usage.json or explicit rollout/backend blocker
Redaction Evidence contains no raw tokens, cookies, PoTokens, signed media URLs, raw audio, or raw source secrets redaction scan note
Support response Acknowledge pilot-blocking issues by the next business day, or during the scheduled working session when live operator notes

If hosted preview health is down for more than 10 continuous minutes during a scheduled working session, stop treating the hosted path as the proof source for that session. Either fall back to local MCP evidence, reschedule the hosted proof, or document the source/STT/deployment exception in the closeout packet.

Degradation Paths

Use the lowest-risk path that preserves the buyer outcome:

Failure First response Completion metric
Hosted API unhealthy Use local MCP packet and reschedule hosted proof Local packet exists and hosted exception is recorded
Admin usage unavailable Capture rollout/backend blocker and keep usage proof open TIN-1559 or successor has owner, evidence, and next action
STT unavailable Keep deterministic YouTube section proof; mark stream proof out of scope or degraded Stream exception cites source/STT posture
Managed STT unavailable Keep self-hosted Whisper first; do not silently send buyer audio to another provider Managed fallback remains disabled or buyer-approved
Source blocked or unstable Use an approved alternate source or document source exception No raw source secrets appear in evidence

SaaS SLA Prerequisites

Do not offer a public SaaS SLA until the platform has:

  • database-backed account, key, usage, invoice, deletion, and export records
  • cross-replica quota enforcement and stream-session state
  • Redis or equivalent stream cursors/buffers for restart-tolerant sessions
  • alerting, dashboards, error-budget policy, and named on-call coverage
  • backup/restore tests for customer account and billing data
  • documented support tiers, incident response windows, and customer remedies
  • repeatable COGS attribution across infra, STT, CI, support, and operator time

The current SLO posture is enough for paid discovery and GTM learning. It is not enough for broad self-serve SaaS.