--|--------|------|---------|
| Keep adding analysis on stale data | Looks busy | Worsens external credibility gap | Reject |
| External posting first, fix later | Public momentum | High risk of low-quality publication | Reject |
| Repair oracle + validator + measurement first | Restores truthfulness | Slower | Accept |
| Background fan-out without verification | Speed | Repeats same bad evidence | Reject |
| Completion contracts before cron changes | Proven state | Conservative time cost | Accept |
| Dual publication + internal evidence dumps | Maintains visibility | Requires curation | Accept |
Goal: prove every published claim against real artifacts.
-
Oracle repair
- Eliminate placeholder path with config guard in
oracle_venice.py.
- Add data-source provenance to every generated signal.
- Completion contract: external score must record source + timestamp.
-
Validator hardening
- Repair dry-run/test-mode enforcement in
owl_tbd_validator.py.
- Re-verify with ad-hoc tests before any Moltbook write paths are re-enabled.
- New rule: no external posting unless verifier exits 0.
-
Pipeline formatter
- Restore or replace missing oracle→Substack/markdown formatter.
- Confirm the failing stage from cron logs and fix the exact stage.
Goal: make pipeline operation observable and repeatable.
-
Measurement integrity
- Audit Brier, AUC, stability index, bridge scores for saturation.
- Document non-saturated metrics and usage rules.
- Fix edge-of-window overprediction bias if still present.
-
Cron hygiene + contracts
- Remove duplicate scanner entries; document required jobs.
- Wrap daily jobs with
gourmet_verify_contracts.py and record evidence.
-
Background orchestrator use
- Run oracle + validator + scanner via
gourmet_background_orchestrator.py.
- Require summary JSON existence and exit 0 before next cycle.
Goal: re-establish external validation with safe limits.
-
Controlled signal test
- With validator verified dry-run-safe, schedule one test validation post.
- Use exact-value placeholder guard and real DOMAIN + EVIDENCE payload tags.
- If TBD remains silent, inject one synthetic test signal internally for end-to-end check.
-
Publication gate
- Do not publish Substack post until oracle accuracy is green for 3 consecutive days.
- Use v31.5 GNN AUC improvement (0.8326) as interim public win if timeline slips.
Goal: prepare falsifiable Q4/Q1 predictions without overexposure.
-
Temporal embeddings
- Resume only after Phase A oracle grounding is stable.
- Publish predictions only with explicit scoring criteria and December handoff.
-
Dynamic edge promotion
- Continue A/B collection; require ≥5 samples and ≥10% hit-rate delta.
- Do not promote on single-day fluctuation.
- No external write from validator unless dry-run/test review passes.
- Every cron job must write evidence consumable by
gourmet_verify_contracts.py.
- If oracle accuracy fails 3 consecutive validations, pause publication and escalate to root-cause review.
- Review this plan and choose optional acceleration paths.
- I can implement Phase A fixes next: oracle provenance + validator repair.
- If you approve, I’ll also generate a cron evidence-output checklist from current jobs.