earos-remediate

name	earos-remediate
description	Generate a prioritized improvement plan from an EAROS evaluation. Triggers on "how do I fix this", "improve this artifact", "remediation plan", "how to pass EAROS", "fix the assessment", "improvement plan", "what's wrong with my architecture", "how to get a better score", or any request to improve an artifact based on evaluation results.

SKILL: earos-remediate

Generate a prioritized, actionable improvement plan from an EAROS evaluation record.

References

• Patterns and before/after examples: references/remediation-patterns.md
• Output template: references/output-template.md
• Rubric files: discovered at runtime via earos.manifest.yaml
• Evaluation schema: standard/schemas/evaluation.schema.json

---

Workflow

Step 1 — Load the evaluation record

Ask the user to provide the evaluation record (file path or paste content).

• Read the file if a path is given.
• Identify: rubric_id, profile_id, overlay_ids, status, all criterion scores, gate results, dimension scores, and the narrative fields (evidence, rationale, actions).
• If the record is missing required fields, note which ones and proceed with what is available.

Step 2 — Load the rubric files

Read earos.manifest.yaml to locate the rubric(s) referenced by the evaluation record.

For each referenced rubric (core + profile + overlays):

• Load the YAML file.
• Extract for every criterion: id, question, gate, scoring_guide, examples.good, examples.bad, anti_patterns, remediation_hints, decision_tree.

This is the authoritative source for what "good" looks like. Do not rely on training data for rubric content.

Step 3 — Classify and prioritize issues

Read references/remediation-patterns.md before this step.

Triage every criterion that scored below 3 (or triggered a gate failure):

Tier 1 — Blockers (fix these first; evaluation cannot pass until resolved)

• Any criterion with a critical gate that failed (status = Reject)
• Any criterion with a major gate that scored < 2 (caps status at Conditional Pass)

Tier 2 — High Impact (most likely to change the overall status)

• Criteria scored 0 or 1 (Absent / Weak)
• Criteria in high-weight dimensions (weight >= 1.1)
• The criterion closest to tipping a dimension above the 2.0 floor threshold

Tier 3 — Incremental (polish, turn Conditional Pass into Pass)

• Criteria scored 2 (Partial) — especially those where the gap to 3 is narrow
• Dimension average between 2.4 and 3.2 (one bump in a criterion could cross the Pass threshold)

Tier 4 — N/A review (challenge whether N/A was correctly applied)

• Any criterion marked N/A — check the justification is sound

Within each tier, order by: gate severity descending → weight descending → score ascending.

Step 4 — Generate remediation items

For each prioritized criterion, produce a structured remediation entry using the rubric's own content:

Criterion: <id> — <question>
Current score: <score>/4   Gate: <severity or none>
What score 3 looks like: <scoring_guide["3"]>
What score 4 looks like: <scoring_guide["4"]>
Good example: <examples.good[0]>
Anti-patterns to avoid: <anti_patterns>
Decision path: <decision_tree excerpt>
Specific fixes:
  - <remediation_hints item 1>
  - <remediation_hints item 2>
  - [context-specific fix derived from the evaluation's rationale]
Effort estimate: <Low | Medium | High>  (see references/remediation-patterns.md)

Derive effort from the gap and criterion complexity:

• Score 3→4: Low (usually evidence and cross-referencing)
• Score 2→3: Medium (content gaps to fill)
• Score 0–1→3: High (structural rework required)
• Gate failure on critical: High (non-negotiable)

Step 5 — Compute projected impact

After fixing Tier 1 and Tier 2 items, estimate what the new status would be:

• Recalculate dimension averages assuming target scores (conservative: assume score reaches the minimum needed, not the maximum possible).
• Check gate conditions.
• State the projected status: Pass / Conditional Pass / Rework Required.

If the current status is Reject due to a critical gate, state clearly: fixing the gate failure is the prerequisite for any other improvements to matter.

Step 6 — Output the remediation plan

Read references/output-template.md before producing output.

Structure:

1. Summary (current status, projected status after fixes, number of issues by tier)
2. Tier 1 — Blockers (gate failures)
3. Tier 2 — High Impact items
4. Tier 3 — Incremental improvements
5. Tier 4 — N/A review (if any)
6. Quick wins (Tier 3 items that are Low effort — do these first)
7. Next step: offer to run earos-assess on the revised artifact

---

Operating Principles

• Rubric-anchored. Every fix recommendation must cite the rubric's own scoring_guide, examples.good, or remediation_hints. Do not invent criteria or guidance not in the rubric.
• Triage ruthlessly. A plan with 15 equal-priority items is useless. The output must be prioritized so the architect knows exactly where to start.
• Feasibility-aware. Flag if a Tier 1 item requires organizational change (e.g., a governance decision) vs. document change.
• Evidence anchored. Where the evaluation record includes evidence anchors, quote them back to the architect so they know exactly what text triggered the low score.
• Status awareness. If the artifact is already a Conditional Pass and the architect's goal is Pass, focus on Tier 3. If it is Reject, Tier 1 is the only thing that matters.