Skip to content

Reasoning Pod: Dry-Run Deployment Guide

This guide defines the dry-run flow for a reasoning pod. All steps are local-first and offline-friendly.

Objectives

  • Validate manifests and resource expectations without contacting hosted services.
  • Produce artifacts (MD + JSON) suitable for PR review and promotion gates.

Control Surface (Knobs)

  • Curriculum phases: configs/training/reasoning/curricula/*
  • Trace capture mode: trace_capture.mode ∈ {weights, activations} (see configs/training/reasoning/baseline.yaml)
  • Evaluation presets: configs/evaluation/reasoning/*
  • Deployment preset: configs/deploy/reasoning_pod.yaml

Formalism (signal tracking): let R be reasoning-readiness and A be artifact completeness. We model readiness heuristic as: R = α·E + β·T + γ·D, where E=evaluation pass ratio, T=trace coverage, D=deployment dry-run parity. Choose α,β,γ per your milestone; ensure R ≥ R_min before promotion.

Dry-Run Steps

1) Repo Map (Reasoning) 

codex repo-map --reasoning > docs/status_updates/repo_map_reasoning.txt

2) Status Report (Artifacts) 

python tools/status_report.py \
  --emit-md docs/status_updates/status_report.md \
  --emit-json docs/status_updates/status_report.json

3) Compose Deployment (Dry-Run) 

python tools/selection_report.py --config configs/deploy/reasoning_pod.yaml \
  --dry-run \
  --emit-md docs/status_updates/deploy_dry_run.md \
  --emit-json docs/status_updates/deploy_dry_run.json

4) Link in PR Include the above artifacts in your promotion PR.

Promotion Checklist (excerpt)

  • Status report (MD+JSON) attached.
  • Dry-run deploy artifacts (MD+JSON) attached.
  • Trace capture mode documented (weights or activations).
  • Evaluation preset recorded (e.g., configs/evaluation/reasoning/base.yaml).

Notes

  • This flow intentionally avoids CI and remote deployment to remain offline-first.
  • For actual hosting, adapt these manifests to your environment (k8s, container runtime, etc.), preserving the artifact trail.