disclose

You estimate how much Claude contributed to one deliverable, show the evidence, and hand the number to the user to confirm or revise.

Live context

echo "author: $(git config user.name 2>/dev/null || echo '(unset — ask)')"
root="$(git rev-parse --show-toplevel 2>/dev/null)"
if [ -n "$root" ] && git check-ignore -q "$root/.ai/disclosures.jsonl"; then
  echo "ledger: enabled ($root/.ai/disclosures.jsonl)"
else
  echo "ledger: disabled (not a git repo, or .ai/ not gitignored)"
fi

Anti-Rationalization (hard gate)

Do not emit a final disclosure line until the user has replied with a rating. Show the estimate, ask, stop. One question, one turn.

Headless / unattended runs (no user to answer): log the estimate with user_soi: null, user_reply: null, and no statement; emit no disclosure line. /handoff surfaces it as unrated later.

Scale (0–10)

Anchors from MediaWise's AI and Your Audience: Disclosing AI (Poynter), as relayed by the user — these four are the source's:

Intermediate guidance — interpolation, not from the source; label it so if asked:

• 2–4: human originates the ideas and writes the substance; AI assists narrowly (brainstorming, lit search, line edits, code snippets the human integrated).
• 6–9: AI produces most of the substance (drafting, analysis, structure); the human directs, reviews, and edits. Less human intervention → higher number. 9 = a small amount of human intervention (a few edits or corrections). 10 = no intervention after the prompt — reading or glancing at it without changing anything still counts as 10.

Estimate

Identify the artifact (the argument, or the most recent shareable thing in this session). Then assess from transcript evidence only, per role:

No numeric weights; weight roles by what the deliverable is: for prose, ideation and drafting carry it; for code or analysis, method design and implementation; for research products, synthesis — unverified counts for more than verified. Review pulls SOI lower than producing the same work would: the deliverable stays the user's, so weigh how much the critique changed it, not how long the critique was.

Blind spot: you see only this session, so prior work the user brought in makes you overstate Claude's share.

Ask

Present, then stop:

Artifact: <!-- path or short label -->
Claude's role: <!-- roles, e.g. "brainstorming and line editing" -->
Evidence:
- <!-- one bullet per role that mattered, 2–4 total -->
Estimated SOI: <!-- n --> (<!-- one-line reason -->)
Human intervention: <!-- what the user changed after the prompt, or "none seen" -->
Blind spot: session-only view; prior or later work not counted.

Confirm, or give your rating (0–10) and any role corrections.

Disclosure line

After the user replies, write the line with their rating and roles:

Content disclosure: <!-- Author --> prepared <!-- "a draft of" | "this" --> <!-- artifact noun -->
with <!-- roles --> assistance from Claude (spectrum of involvement: <!-- n -->).

Role phrases for the roles slot (join with "and"; list only roles that mattered): brainstorming (ideation) · drafting · analysis and coding · editing · review · research and synthesis. Human intervention is never a listed role.

Adjust the verb to stay truthful at the ends of the scale:

• 0: no line.
• 6–9: "Claude drafted this <noun> at the direction of <Author>, who reviewed and edited it (spectrum of involvement: <n>)."
• 10: "This <noun> was generated by Claude from a prompt by <Author> without further human editing (spectrum of involvement: 10)."

Return the line in the reply, and when the artifact is a file you can edit, offer in one line to insert it — default to the end, but ask where. Insert only on a yes.

Drift check: before printing, confirm the line keeps the template's shape and matches the user's number, roles, author and artifact. If they dictated wording, use it, but flag substantive drift (missing SOI, a role added or dropped, wording that implies more or less involvement than the number).

Ledger

Unless --no-ledger or the context block says disabled, append one JSON line to <repo>/.ai/disclosures.jsonl (mkdir -p the .ai/ dir first; never create or edit .gitignore):

{"date": "YYYY-MM-DD", "artifact": "<path or short label>", "roles": ["..."], "evidence": ["<one short clause per role that mattered>"], "estimated_soi": 4, "user_soi": 3, "user_reply": "<verbatim>", "statement": "<line>"}

evidence is the estimate's supporting bullets, compressed to a clause each. user_reply is the user's rating message quoted verbatim — it makes a skipped gate checkable against the transcript. user_soi is null if the user declined to rate; then write no statement either. Best-effort: a failed write is mentioned, never blocks. Keep both SOIs; the pair is the calibration record.

Verification

• Estimate shown with per-role evidence and the blind-spot line, then stopped
• The user replied with a rating (or explicitly declined) before any disclosure line
• Disclosure line uses the user's number and role corrections, not the estimate
• Printed line follows the template shape, and any substantive drift from the user's reply was flagged
• Insertion offered when the artifact is an editable file
• Ledger entry carries evidence and the user's verbatim reply, or the ledger was disabled and said so