Reference Controls Enricher

Reference Controls Enricher

What this skill does

For a given framework YAML, walk every assessable requirement and propose 1–5 reference control URNs from the central doc-pol library (the CISO Assistant Key Reference Controls). Outputs:

• a reviewable xlsx per framework
• a YAML patch applied in place (adds URNs to each requirement's reference_controls field, adds doc-pol to dependencies)
• a coverage report

Claude is the reasoner — no embedders, no local models. The pipeline is pure stdlib + pyyaml (+ optional openpyxl for xlsx).

When NOT to use this skill:

• For frameworks with >500 assessable items: the conversation context fills up. Chunk the run across sessions or adapt tools/mapping_builder/map_v2.py for offline use.
• For frameworks that already reference doc-pol:* URNs on most requirements: run coverage_report.py first; if coverage is already good, don't rewrite.

How to use the helpers

Stdlib + pyyaml scripts under .claude/skills/reference-controls-enricher/scripts/.

SKILL=.claude/skills/reference-controls-enricher/scripts
MB=.claude/skills/mapping-builder/scripts  # reused for framework parsing

# Parse a framework YAML (same parser as mapping-builder)
.venv/bin/python $MB/parse_framework.py backend/library/libraries/<framework>.yaml > /tmp/fw.json

# Parse the key-controls library
.venv/bin/python $SKILL/parse_controls.py backend/library/libraries/key-reference-controls.yaml > /tmp/controls.json

# Write a reviewable xlsx from the running verdicts
.venv/bin/python $SKILL/write_review.py /tmp/fw.json /tmp/controls.json /tmp/verdicts.jsonl /tmp/<name>_review.xlsx

# Apply URN assignments to the framework YAML (adds doc-pol dependency if needed)
.venv/bin/python $SKILL/apply_enrichment.py backend/library/libraries/<framework>.yaml /tmp/verdicts.jsonl --bump

# Audit coverage + low-confidence verdicts + zero-URN requirements
.venv/bin/python $SKILL/coverage_report.py /tmp/fw.json /tmp/verdicts.jsonl

Verdicts file format

One JSON object per line. target_urns is the list of doc-pol URNs to link. Empty list → requirement is consciously left unlinked.

{"source_urn": "urn:intuitem:risk:req_node:<fw>:<ref>",
 "target_urns": ["urn:intuitem:risk:function:doc-pol:pol.access",
                 "urn:intuitem:risk:function:doc-pol:tech.mfa"],
 "confidence": 8,
 "rationale": "Access control obligation with explicit MFA — maps to access policy + MFA capability."}

Persist /tmp/<framework>_verdicts.jsonl and append after each section so progress is recoverable if the conversation gets interrupted.

Always read source_urn from the parsed framework JSON's items[].urn — do NOT construct it from ref_id. Ref_ids may use mixed case (e.g. 2.A.1) but the URN suffix is always lowercase (e.g. 2.a.1). Constructing URNs by string concatenation is a known bug trap.

Parsed framework shape (important)

parse_framework.py emits a flat items list, each item carrying section_urn, section_ref_id, section_name — items are not nested inside sections[].items. Iterate parsed["items"], not parsed["sections"][i]["items"]. This differs from common heuristic expectations and has bitten at least one implementation mid-session.

Workflow

Step 1 — confirm the request

Confirm with the user:

• Target framework YAML path (the one being enriched)
• Whether the framework already has a doc-pol dependency — if yes, existing reference_controls on requirements must be preserved (this skill appends, never replaces)
• Whether they want the skill to apply in place or emit a review-only pass

Grep the framework for urn:intuitem:risk:function:doc-pol: to check prior linkage. Report coverage before touching anything.

Step 2 — parse both sides

.venv/bin/python .claude/skills/mapping-builder/scripts/parse_framework.py \
  backend/library/libraries/<framework>.yaml > /tmp/fw.json

.venv/bin/python .claude/skills/reference-controls-enricher/scripts/parse_controls.py \
  backend/library/libraries/key-reference-controls.yaml > /tmp/controls.json

Read both JSONs. Note n_assessable for the framework and the doc-pol prefix breakdown (doc / pol / proc / tech / phys / train).

Size gate: if n_assessable > 500, stop and recommend a chunked or offline approach.

Step 3 — build a section → prefix-family affinity map (DO NOT SKIP)

Read the sections array from the parsed framework (typically 5–30 entries). For each section, reason about which doc-pol prefixes are most likely relevant. doc-pol prefixes group by control-type:

For a governance section, expect mostly pol.* + doc.*. For a monitoring section, pol.monitor + proc.logging + tech.siem + train.logging. Etc.

Print the affinity table to the user; they sanity-check before Step 4. This narrows the 170+ controls per section to ~10–25 candidates.

Step 4 — per-section deep dive

For each section, pull:

1. The section's assessable items (read their full_sentence).
2. Candidate doc-pol controls filtered by the section's prefix families — read name + description + annotation for each.

For each requirement, decide which controls apply (1–5 is typical; keep it tight). Emit a verdict per requirement:

{"source_urn": "...",
 "target_urns": ["...", "..."],
 "confidence": 0-10,
 "rationale": "one-sentence why"}

Confidence scale (calibrated from real runs):

• 10: obvious / near-identical semantic match (e.g. "implement the PSSI" → pol.main). Use sparingly — reserved for the ~15% of cases with zero interpretation.
• 8–9: clean semantic match — this is the norm and should be where most verdicts land.
• 7: plausible, some interpretation or scope trimming required.
• 6: mitigation / exception patterns — recognise the pattern "when X is not possible, entity implements risk-reduction measures". These legitimately map to proc.exception_mgmt + the parent topic at confidence 6. Don't chase these to higher confidence — they're correctly low.
• 0–5: weak — rare; only include if the requirement is thin and this is the best available.

Single-URN verdicts are fine when many requirements are sub-clauses of the same parent topic (e.g. ten pol.access sub-clauses in an IAM objective). Don't force artificial expansion to meet an imagined 2-URN minimum.

Four-plus-URN verdicts must be justified explicitly — e.g. a requirement listing "at minimum: crypto policy, access policy, audit policy, account policy" genuinely deserves 4 URNs. Default 4+ linkages are usually over-linking; 2–3 is the sweet spot.

Prefer strong high-level policy URNs (e.g. pol.access) over too-specific technicals when the requirement is broadly worded. Prefer specific tech URNs (e.g. tech.mfa) when the requirement explicitly mentions a capability. Don't over-link — 3 crisp URNs beat 8 soft ones.

Affinity is a starting heuristic, not a fence — if a requirement obviously needs a control from a different prefix family, pull it.

Append verdicts to /tmp/<name>_verdicts.jsonl section-by-section so progress is recoverable.

Step 5 — borderline review with the user

.venv/bin/python .claude/skills/reference-controls-enricher/scripts/coverage_report.py \
  /tmp/fw.json /tmp/<name>_verdicts.jsonl --threshold 5

Surface to the user:

• Verdicts with confidence < 5 — borderline calls to confirm or drop
• Requirements with zero target URNs — either a true gap (add to doc-pol backlog) or a miss (reconsider)
• Requirements with >5 URNs — likely over-linking

Apply the user's edits to the running jsonl before moving to Step 6.

Slug usage as a sanity signal. Check the most-cited doc-pol slugs — they should match the framework's character. Example: for a French NIS2-style framework (RECYF), pol.access, pol.network, proc.pam dominating is expected. For a privacy framework, expect pol.privacy, proc.dpia, doc.ropa at the top. If the top slugs don't match the framework's theme, something's off (too shallow, or affinity was wrong).

Also check the unused slugs: they should fall in out-of-scope domains relative to the framework. A security framework using 0 of the privacy/AI/OT slugs is normal. If a supposedly broad framework leaves 70%+ of slugs unused, reconsider affinity.

Step 6 — write review xlsx

.venv/bin/python .claude/skills/reference-controls-enricher/scripts/write_review.py \
  /tmp/fw.json /tmp/controls.json /tmp/<name>_verdicts.jsonl \
  /tmp/<name>_review.xlsx

Open the xlsx with the user and walk through flagged rows (LOW / ZERO). Edit the jsonl if anything changes; re-run write_review.py to refresh.

Step 7 — apply in place

Once the user is comfortable:

.venv/bin/python .claude/skills/reference-controls-enricher/scripts/apply_enrichment.py \
  backend/library/libraries/<framework>.yaml \
  /tmp/<name>_verdicts.jsonl \
  --bump

This:

• Appends target_urns to each matching requirement's reference_controls (dedup against existing)
• Adds urn:intuitem:risk:library:doc-pol to the framework's dependencies (if new URNs were added)
• Bumps the framework's version (if new URNs were added)

Run the standard post-check:

.venv/bin/python backend/manage.py storelibraries \
  --path backend/library/libraries/<framework>.yaml

Expect "Successfully stored library". If the library fails to load, inspect the stderr and revert via git checkout -- backend/library/libraries/<framework>.yaml.

Step 8 — final summary, no auto-PR

Report:

• Enrichment coverage % (from coverage_report.py)
• Number of URNs added, number of requirements touched
• Path to the review xlsx
• Path to the verdicts jsonl (for replay / audit)
• Suggested commit message

Do NOT auto-commit or auto-push. The user confirms.

Common mistakes to avoid

• Over-linking. 3 tight URNs beat 8 loose ones. The suggestions UI exposes all linked controls to the auditor — noise hurts.
• Constructing source URNs from ref_id. Case mismatches (2.A.1 vs 2.a.1) break resolution silently. Always use the urn field from the parsed framework JSON.
• Iterating sections[].items. parse_framework.py emits a flat items list. Iterate parsed["items"] and group by section_ref_id if you need section structure.
• Missing the affinity step. Without pre-filtering by prefix family, you'll waste context reading irrelevant controls for each requirement.
• Replacing existing reference_controls. apply_enrichment.py only appends, but if a future script or hand-edit clobbers the list, audit evidence gets orphaned.
• Forgetting the dependency line. apply_enrichment.py adds it automatically when any URN is added, but if you hand-edit the verdicts file, the library will still fail to load on import with URN-resolution errors.
• Ignoring zero-URN requirements. A truly uncovered requirement tells you either: (1) the framework is asking for something doc-pol doesn't have (flag for doc-pol extension), (2) the requirement is too abstract to link (leave empty by design), or (3) you missed a match (revisit).
• Re-applying after editing verdicts. apply_enrichment.py only appends and dedups — it cannot remove URNs from a prior apply. To change verdicts after an apply, git checkout -- <framework.yaml> first, then re-run with the updated jsonl.
• Chasing confidence 10 for mitigation cases. Requirements phrased "when X is not possible, implement mitigating measures" correctly map to proc.exception_mgmt + the parent topic at confidence 6. Don't force them higher — the lower score accurately signals the exception nature.

Quick reference