Registered Work · project · experimental
SISO Compute Pool · Free Qwen Worker Tier
A controller that serves Qwen3.8-27B on free Kaggle TPU sessions behind a private Bifrost gateway, so agents gain a zero-cost worker tier. Transport is a Cloudflare Worker relay the notebook dials out to, because the account holds no zone. The delivery path is verified end-to-end against a mock origin; no TPU allocation, model throughput or output quality has been measured, and continuous service is not established.
For agents
SISO · Open authorized private source entry ↗ · Machine-readable dossier
No reviewed run or test commands are recorded for this Work. Read its owning instructions before execution. The selected release and source availability are shown below; a source link alone is not installation or access permission.
Source & upstream links
Relationships
SISO Harness Lab · Agent EffectivenessShares the agent-effectiveness programme's evidence discipline; this project supplies a candidate zero-cost worker tier to it.
Evidence & receipts
From the OVH VPS against a mock origin: probe returned ready true, register returned registered true with gateway_verified true, and one chat completion traversed Bifrost and the Cloudflare relay in 373 ms with a Cf-Ray header. Relay auth rejects an absent key with 401 and an absent worker with 503; streaming relays frame by frame and reconnects after client loss. This is transport evidence only.siso-compute-pool:docs/PROGRESS-20260910.md
Kaggle permits one concurrent TPU session per account, refused at push time rather than queued, with a roughly nine hour session cap and a twenty hour weekly quota. Observed cold start is 25 to 50 minutes and one probe queued 44 minutes. Queued kernels consume no quota, so a successor can be pre-queued, but overlap requires a second account. Unattended use is therefore a bounded batch, not continuous service.siso-compute-pool:docs/FINDINGS.md
Qwen has never served a token under this controller. TPU allocation, model startup, sustained throughput and output quality remain unmeasured. Readiness gate one exists because Kaggle has previously returned CPU silently in place of a requested accelerator.siso-compute-pool:docs/PROGRESS-20260910.md
Provenance
- Work ID
- gls:work:766cc7a3-79ed-4bfe-9602-8cea1f6f7f51
- Registry source
- registry/works/work-siso-compute-pool.json
- Origin
- siso
- License / redistribution
- NONE (not_applicable)