Skip to main content

​
Scoring-only dress rehearsal

DCinside sends real candidates from the full admin extraction queue to Surf and receives model scores. DCinside records what it would have published and compares those choices with operators’ actual picks. No posts are published by the model during this rehearsal; operators continue as usual. The rehearsal checks the integration. Model quality needs later evaluation with real placements.
These pages incorporate DCinside’s September 30 timing and recovery requirements. They are not a declaration that the deployed service meets the SLA. Verify the active image, full-queue latency and readiness checks before the three-day run. See Rehearsal and acceptance.

​
Data flow

  1. DCinside applies its 24-hour age filter and sends each new candidate’s text and images once.
  2. Every cycle, DCinside sends engagement updates for all pending candidates, including captured between-cycle entrants with source="unknown" when needed. Saved/hidden labels stay client-side and do not trigger removals under the operator-independent policy. Posts remain pending after first scoring until DCinside sends a deleted/aged_out removal; the client owns the 24-hour age filter. Original clocks are preserved.
  3. Surf returns waiting before queue entry plus 15 minutes, then a final score, post identity, model version and scoring time. Eligible pending candidates are refreshed each cycle.
  4. DCinside logs would-be picks. Late or stale responses are skipped for that cycle; recovered responses remain useful for reconciliation.
All connections are initiated by DCinside over HTTPS, with inline cycle responses and a retrieval endpoint for recovery. The rehearsal uses fixed five-minute cycles and allows 60 seconds from sending, including at most one retry. About 2,200 candidates/day is a volume estimate, not a quota.

​
Rehearsal endpoints

​
Reference endpoints — Not in scope for trial

​
Model and timing

The intended model is the delivered Model 0924 with down-vote weight 0. A transport-test adapter is not a substitute for that model. Per-part predictions are optional. DCinside requires at least 99% of candidates to receive a first score within two cycles after the 15-minute queue-entry mark. Before that mark, return waiting. This target is unchanged and does not end retention. Saved/hidden decisions remain in the denominator; immediate deleted/aged_out removals before the mark retain the early-removal exclusion. The proposed 2026-10-15 00:00 KST switch is unconfirmed, and the earlier first-score-or-25-minute hold was never activated. Record the actual switch and keep historical cohorts distinct. The response deadline is 60 seconds from sending, including at most one retry. A result recovered later is reconciliation data and cannot be used for a decision. Hidden alone does not veto a model pick. The client skips posts currently on the real board; API last-20 tier windows do not cover every board post. Post-decision scores use cycle-time evidence and do not reconstruct a score before the operator decision. Previously removed posts cannot be revived by resending; any retrospective backfill remains a separate, undecided migration question. See the worker handoff and cycle contract. Read the conventions and rehearsal acceptance criteria before integrating. How Scoring Works is marked Not in scope for trial. The machine-readable specification is OpenAPI JSON.