decorum-voice-judge

decorum-keeper.py measures the DETERMINISTIC rungs of professionalization — spelling (vendored typo map), bio/positioning staleness (git authored age), and narrative accuracy (profile claims vs the VVLTVS contribution mix). It cannot judge whether the public prose reads professionally: tone, brand-voice, "does this sound like a founder or a first draft." This skill is that missing rung.

The keeper detects when a prose surface's bytes change and no current voice-judgment covers them, and queues a review (a polish/voice-judge finding, and — when the fleet effector is armed — a Haiku-tiered task pointing here). This skill reads the changed surface, scores it, and appends a verdict pinned to the exact bytes it judged (content_sha256) to the register, which clears the finding. A later edit changes the sha and re-queues review — so judge against the CURRENT bytes only.

Tier: Haiku-first (per the model-tiering policy). Voice scoring is a bounded, cheap judgment; do not escalate unless the prose is genuinely ambiguous.

Steps

1. Find what's queued. Run python3 scripts/decorum-keeper.py --sweep --json and read egg_face_findings[] for entries with source: "voice-judge" — each names a changed prose surface (a repo-relative path like README.md or docs/positioning/<x>.md). Or take the path from the task context that invoked you.

2. Read the CURRENT bytes and compute the sha. Read the file at that path, then python3 -c "import hashlib,sys;print(hashlib.sha256(open(sys.argv[1],'rb').read()).hexdigest())" <path>. This exact hex is what pins your verdict — record it verbatim.

3. Score the prose 0-5 on four axes (be a demanding reader who represents the target visitor — an investor, a hiring manager, a paying customer):

   • tone — confident without bluster; not hedging, not cringe, not LLM-boilerplate ("delve", "in today's fast-paced world", empty superlatives).
   • brand_voice — sounds like THIS operator's estate (precise, systems-minded, build-in-public), not a generic template; consistent with the other public surfaces.
   • clarity — the value proposition lands in the first sentence; no jargon wall; a stranger understands what this is and why it matters.
   • credibility — claims are backed (links, numbers, artifacts); nothing reads as vaporware or over-claim; no unsupported superlatives.
   • verdict = pass if every axis ≥ 3 and none is a visible embarrassment; else fail.

4. Append the verdict to institutio/observatory/decorum-judgments.yaml under judgments: keyed by the surface path (newest row last). This file PUBLISHES — keep every note public-safe (no PII, no private URLs, no secrets). Row shape:

   judgments:
     README.md:
       - judged_at: "<UTC ISO8601, e.g. 2026-07-23T00:00:00Z>"
         content_sha256: "<the hex from step 2>"
         model: "claude-haiku-4-5"          # or your handle if a human judged
         verdict: pass                        # pass | fail
         scores: {tone: 4, brand_voice: 4, clarity: 5, credibility: 4}
         defects: ["opening line buries the value prop under a stack list"]
         suggested_fix: "lead with the one-sentence outcome; move the stack list below the fold"
   

5. Confirm it cleared. Re-run python3 scripts/decorum-keeper.py --sweep — the voice-judge finding for that surface should be gone (a pass row whose content_sha256 matches the current bytes clears it). If it persists, your row's sha does not match the current file — recompute and fix the row.

Laws

• Pin to bytes, not to the path. A verdict without the exact content_sha256 is worthless — it cannot be invalidated on the next edit. Always record the hash of the bytes you actually read.
• Public-safe. The register publishes; write no PII, no private URL, no secret in any note.
• Judge current only. If the file changed after you computed the sha, recompute — never record a verdict against bytes you didn't read.
• A fail is a gift. A specific suggested_fix is worth more than the score; it is what the next fix task acts on.