distill — independently scanned and version-tracked by SaferSkills.
SaferSkills independently audited distill (Agent Skill) and scored it 100/100 (green). The audit ran 55 deterministic rules across Security, Supply Chain, Maintenance, Transparency, and Community; it found 0 high-severity and 0 lower-severity findings. The full rule-by-rule trace and per-finding evidence are below. Free, methodology-open.
Findings & checks · 0 flagged
Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
scripts/preflight.sh passes (jq, bats, bash required).scripts/parse-ratio.sh and scripts/validate-essence.sh are executable.| Flag | Required | Description | |
|---|---|---|---|
| `--mode reduce\ | construct` | Yes (when called directly) | reduce = essence reduction, construct = structural expansion |
--ratio 0.05–0.5 | No (used in reduce mode) | Natural-language values accepted — parsed by parse-ratio.sh | |
--audience {target} | No | If omitted, the adapter handles it via HITL | |
--weak-decomposition | No | Allow Weak D/E/V/R decomposition | |
--input {path} | No | File path instead of stdin |
Persist source text to .galmuri/tmp/source-{slug}.txt and reset the retry counter (clears leftovers from prior sessions). If --mode is specified, skip Step 2 and jump to Step 3.
mkdir -p .galmuri/tmp
SLUG=$(echo "$INPUT" | head -c 40 | tr -cs '[:alnum:]' '-' | tr '[:upper:]' '[:lower:]')
cat > ".galmuri/tmp/source-${SLUG}.txt"
rm -f ".galmuri/tmp/retry-count.${SLUG}"Skipped when --mode is explicit. Otherwise: count tokens, suggest a mode (< 80 → construct, ≥ 80 → reduce), confirm via HITL — "Proceed with reduce or construct? e.g. reduce (summarize), construct (structural expansion)."
TOKEN_JSON=$(bash scripts/count-tokens.sh ".galmuri/tmp/source-${SLUG}.txt")
TOKEN_COUNT=$(printf '%s' "$TOKEN_JSON" | python3 -c "import sys,json; print(json.load(sys.stdin)['tokens'])")
[ "$TOKEN_COUNT" -lt 80 ] \
&& echo "[distill] short input (${TOKEN_COUNT} tokens) → construct mode recommended" \
|| echo "[distill] long input (${TOKEN_COUNT} tokens) → reduce mode recommended"Apply the three method references in order:
| # | Method | Reference | Output |
|---|---|---|---|
| 1 | Essence reduction | references/prompt.md | each claim → one-line subject-verb essence |
| 2 | First-principles decomposition | references/decomposition.md | 4 D/E/V/R questions (Weak mode under --weak-decomposition) |
| 3 | Socratic validation | references/socratic_probe.md | 3-axis probe → failed units → dropped[] |
Retry contract (deterministic): (A) produce candidate essence_units → (B) schema-check via bash scripts/validate-essence.sh → (C) in reduce mode, re-apply if |actual_ratio − target_ratio| > 0.05 (max 2 retries; past that, fall through to Step 5 HITL).
COUNT_FILE=".galmuri/tmp/retry-count.${SLUG}"
CURRENT=$(cat "$COUNT_FILE" 2>/dev/null || echo 0)
echo $((CURRENT + 1)) > "$COUNT_FILE"Final adjudication after the Step 3 LLM loop. If |drift| > 0.05 persists: re-enter Step 3 when retry-count < 2, else fall through to Step 5 HITL.
bash scripts/validate-essence.sh < output.json
bash scripts/count-tokens.sh output.jsonvalidate-essence.sh pass is mandatory → emit EngineOutput JSON. Fail → exit 1 + HITL ([a]ccept / [r]e-target / [c]ancel). Either branch cleans up the retry counter before exit.
bash scripts/validate-essence.sh < final-output.json || {
echo "[distill] validation failed" # HITL: [a]ccept / [r]e-target / [c]ancel
exit 1
}
cat final-output.json
rm -f ".galmuri/tmp/retry-count.${SLUG}"EngineOutput JSON (matches essence-schema.json):
{
"units": [...],
"mode": "reduce|construct",
"ratio": 0.2,
"dropped": [...],
"source_ref": ".galmuri/tmp/source-{slug}.txt"
}No persistence here — the adapter receives EngineOutput and decides whether to persist.
~30 seconds. Free. No account. Every finding cites a rule and a line of evidence.