Skip to content
Arxo Arxo

Policy and CI Gates

Policy and CI Gates

Section titled “Policy and CI Gates”

Use these policies to move from visibility to enforcement for fine-tuning architecture risk.

Visibility-First Policy (Initial Rollout)

Section titled “Visibility-First Policy (Initial Rollout)”
metrics:
  - id: finetuning_architecture


policy:
  invariants:
    - metric: finetuning_architecture.overall_finetuning_health
      op: ">="
      value: 0.65
      severity: warning
      message: "Raise fine-tuning architecture health above minimum baseline"
    - metric: finetuning_architecture.reproducibility_score
      op: ">="
      value: 0.65
      severity: warning
      message: "Improve reproducibility controls"
    - metric: finetuning_architecture.data_integrity_score
      op: ">="
      value: 0.65
      severity: warning
      message: "Improve data integrity and eval maturity"
    - metric: finetuning_architecture.safety_governance_score
      op: ">="
      value: 0.65
      severity: warning
      message: "Improve artifact trust/privacy/governance controls"

Standard Policy (Most Teams)

Section titled “Standard Policy (Most Teams)”
metrics:
  - id: finetuning_architecture
    config:
      profile: "sft"
      require_base_pinning: true
      require_eval_harness: true
      require_full_determinism: false
      require_checkpoint_eval_lineage: true
      require_safe_serialization: true
      privacy_profile: "recordkeeping"


policy:
  invariants:
    - metric: finetuning_architecture.overall_finetuning_health
      op: ">="
      value: 0.70
      severity: warning
      message: "Overall fine-tuning architecture health baseline not met"
    - metric: finetuning_architecture.base_model_versioning_score
      op: ">="
      value: 0.80
      severity: warning
      message: "Base model pinning baseline not met"
    - metric: finetuning_architecture.eval_absence_score
      op: ">="
      value: 0.75
      severity: warning
      message: "Eval readiness baseline not met"
    - metric: finetuning_architecture.checkpoint_eval_lineage_score
      op: ">="
      value: 0.80
      severity: warning
      message: "Checkpoint-eval lineage baseline not met"
    - metric: finetuning_architecture.artifact_trust_surface_score
      op: ">="
      value: 0.85
      severity: warning
      message: "Artifact trust/safe serialization baseline not met"
    - metric: finetuning_architecture.dataset_contamination_score
      op: ">="
      value: 0.75
      severity: warning
      message: "Dataset contamination controls are insufficient"

Strict Policy (Production Fine-Tuning)

Section titled “Strict Policy (Production Fine-Tuning)”
metrics:
  - id: finetuning_architecture
    config:
      profile: "rft"
      require_base_pinning: true
      require_eval_harness: true
      require_full_determinism: true
      require_preference_eval: true
      require_checkpoint_eval_lineage: true
      require_safe_serialization: true
      privacy_profile: "strict"


policy:
  invariants:
    - metric: finetuning_architecture.overall_finetuning_health
      op: ">="
      value: 0.80
      severity: error
      message: "Overall fine-tuning architecture health baseline not met"
    - metric: finetuning_architecture.method_integrity_score
      op: ">="
      value: 0.80
      severity: error
      message: "Post-training method integrity baseline not met"
    - metric: finetuning_architecture.checkpoint_eval_lineage_score
      op: ">="
      value: 0.80
      severity: error
      message: "Checkpoint-eval lineage baseline not met"
    - metric: finetuning_architecture.artifact_trust_surface_score
      op: ">="
      value: 0.85
      severity: error
      message: "Artifact trust/safe serialization baseline not met"
    - metric: finetuning_architecture.privacy_recordkeeping_score
      op: ">="
      value: 0.80
      severity: error
      message: "Privacy and recordkeeping controls are insufficient"
    - metric: finetuning_architecture.dataset_contamination_score
      op: ">="
      value: 0.85
      severity: error
      message: "Dataset contamination controls are insufficient"

Baseline No-Regression Policy

Section titled “Baseline No-Regression Policy”
metrics:
  - id: finetuning_architecture


policy:
  baseline:
    mode: git
    ref: origin/main
  invariants:
    - metric: finetuning_architecture.overall_finetuning_health
      op: ">="
      baseline: true
      severity: error
      message: "Overall fine-tuning architecture health regressed vs baseline"
    - metric: finetuning_architecture.eval_absence_score
      op: ">="
      baseline: true
      severity: error
      message: "Eval readiness regressed vs baseline"
    - metric: finetuning_architecture.checkpoint_eval_lineage_score
      op: ">="
      baseline: true
      severity: error
      message: "Checkpoint-eval lineage regressed vs baseline"

CI Command Examples

Section titled “CI Command Examples”
Terminal window
# Focused fine-tuning architecture run
arxo analyze --path . --metric finetuning_architecture
Terminal window
# AI preset run (includes finetuning_architecture)
arxo analyze --path . --preset ai --config arxo.yml --fail-fast

Rollout Guidance

Section titled “Rollout Guidance”
  1. Start with warning-level thresholds for 1-2 release cycles.
  2. Fix recurring low-score categories in central training modules first.
  3. Promote strict gates to error once score trends stabilize.
  4. Keep baseline no-regression checks enabled to prevent drift.
Section titled “Read Next”