Applied intelligence
Phoenix AI
Phoenix is the intelligence of Snowdaddy Solutions. One part of it answers people today: Dove, the assistant inside the 365Entertainment creator studio, running on an open-weight model on a GPU server we operate. The rest of Phoenix is written, and some of it is built and tested. Every piece on this page carries its status.
The roster
Fourteen birds, one intelligence
Phoenix was written as named specialists around one judge: each bird a role, each role a model. Their definitions exist; one bird serves people today, one is built and tested, and the rest wait as written. The roles below are quoted from each definition as written.
- Dove in productionThe assistant of the 365Entertainment creator studio: it suggests replies, and the creator reads, edits and sends them.
- HAWK built and testedA second opinion and search index for Snowdaddy Security's camera events, giving way to Dove on the same GPU.
- Orchestrator writtenSwarm judge, task decomposer, dispatch coordinator and result aggregator.
- Synthesis writtenDecision quality monitor, architectural reviewer, swarm improvement engine.
- Eagle writtenBuild coordinator, code generation, backend and API design, interface work.
- Swan writtenChain-of-thought reasoning, strategic failure analysis, architectural logic.
- Falcon writtenServer infrastructure, deployment planning, security hardening, service management.
- Owl writtenCanadian law: personal-information law, business law, regulatory compliance.
- Heron writtenService health monitoring, resource tracking, queue management.
- Raven writtenAudit trail custody, log analysis, evidence preservation, chain of custody.
- Sparrow writtenFinancial ledger work, accounting verification, revenue reconciliation.
- Kestrel writtenSupport ticket triage, request routing, first responses, initial checks.
- Cormorant writtenDocumentation reconciliation, research synthesis, dependency analysis.
- Pigeon writtenMessage cleanup, multi-bird summaries, the handoff to the operator.
Dove in production
An assistant that asks before every step
Dove lives inside the 365Entertainment creator studio. It answers from an open-weight language model on a GPU server we operate; no creator's words are sent to an outside AI service.
Off until started
Dove is off until the creator starts it, and starting it lets nothing be sent in the creator's name.
One consent per step
Every step beyond suggesting is its own consent, recorded against the exact words shown. Stopping Dove withdraws every step at once.
Nothing sends on its own
Dove suggests; the creator reads, edits and sends. The steps that would let it answer alone stay locked on every platform until each platform's real pages have been tested.
Always left to the creator
A sign that a fan may be under 18, distress, requests to move off the platform, threats, payment outside the platform, or a fan asking if they are talking to a bot.
No transcripts
For each fan Dove keeps a coded reference, counts and single keywords, for that creator only; they are removed after twelve months without contact.
Honest when down
If the model cannot be reached, Dove says it is offline. Nothing is written in its place and there is no fallback service.
HAWK built and tested
A worker built to give way
HAWK is a second worker for Dove's GPU, built for Snowdaddy Security: a second opinion on each camera event and a search index of its pictures. It is built and tested; it is not deployed.
Dove first
A small watcher checks every second whether Dove is busy, freezes HAWK within about a second when it is, and lets it continue after Dove has been idle for a minute.
Work on a lease
HAWK takes its work from a queue on the site's machine as time-limited leases. If it is frozen or stopped the lease runs out and the work is handed out again; each result is stored once.
No people
HAWK never receives a picture of a person, a face or a plate: only small cut-outs of vehicles, animals and objects, kept in memory.
A pinned model
Its model file is pinned by checksum; HAWK refuses to start if the file does not match.
The rules written
How Phoenix was written to work
The orchestration was written for servers that are retired, so none of it runs today. Its rules are the ones every new piece is held to.
- Rounds
- One model reads a request, splits it into tasks, runs the birds in up to four rounds and merges what they return.
- Authority
- A bird has no shell and no network authority of its own. Read-only checks run directly; anything that writes, deploys or restarts waits until the operator approves it.
- Strikes
- Three strikes stop every automated approval until the operator clears them by hand.
- Learning
- Nothing is indexed before review. A model or prompt is promoted only with a repeatable evaluation, a measured gain, no safety regression and the operator's approval. No model has been fine-tuned.
- Knowledge
- Each bird knows only what its role allows; retrieval never carries restricted data into another lane.
Phoenix AI
Intelligence that says what it is
Dove works inside 365Entertainment today; the Phoenix site holds the rest of the story.