outcomes-ledger

Every list a playbook hands over gets acted on outside GooseWorks: the human emails the prospects, messages the creators. This atom is how what happened comes back, so the next run never repeats a person who said no and gets better from what worked. It never contacts anyone and never requires outcomes: no outcomes, carry on.

Parameters (the loading playbook sets them)

	entity: person	entity: creator
Loaded by	research-audience (writes rows when a list is shown), write-outreach (updates statuses from replies and results)	work-with-creators
Ledger file	research/audience/ledger.csv	creators/roster.csv
Lessons document (file → brand document)	research/audience/lessons/<YYYY-MM-DD>.md → Outreach lessons — <Brand> — <YYYY-MM-DD>	creators/lessons/<YYYY-MM-DD>.md → Creator lessons — <Brand> — <YYYY-MM-DD>
Entity columns (after the required ones)	segment, tier, value_tier, signal_type, email_status, region_flag, email, email_candidate, profile_url, root_domain, route, sequence_id, opener_id, first_touch_date, last_touch, reply_category, recontact_after	role, tier, deal_structure, vetting_depth, wave, wave_criterion, wave_result, quoted_rate, then the roster's own columns
Status values	listed · contacted · replied · meeting · not-a-fit · bounced · unsubscribed · do-not-contact	shortlisted · contacted · replied · quoted · declined · signed · posted · never-again
Suppressing statuses (permanent)	bounced · unsubscribed · do-not-contact	never-again
Timed exclusions (not suppression)	not-a-fit — the account is skipped for 6 months, then eligible again; contacted with no reply after the sequence ended — the person is skipped for 12 weeks on every channel (cool-down)	declined — skipped until the human re-opens it
Match keys, in order	email (or email_candidate) → profile URL → name + root domain; account rows: root domain + route	handle + platform → profile URL
Grouped by (for lessons)	segment, tier, signal type, email status, region flag, sequence, opener and touch, reply category	role, tier, deal structure, vetting depth, wave
Re-check	recheck_by = date listed + 30 days	revet_by = vetted on + 30 days

Files and scope

• Brand coworker workspace (file_* scope agent = the brand's coworker_agent_id from the brand_read list). No coworker → the default workspace with every path prefixed brands/<brand-slug>/, and tell the human teammates won't see it.
• Read before write: file_write replaces the whole file. Read the ledger, change rows, write it back complete.
• Every row a playbook shows the human is ledgered at the moment it is shown — sample rows included — not only at handover. Rows from a sample whose definition the human then rejected get status listed-sample-rejected and are not excluded later (the person was never handed over for outreach).
• Required columns, first and in this order (both entities): row_type, key, key_hash, name_or_handle, account_or_platform, date_listed, list_id, status, status_date, suppressed, suppressed_reason, suppressed_date, recheck_by|revet_by, outcome, notes. key is the first available match key (person: email, else profile URL, else name + root domain; creator: platform:handle). Then the entity columns above, then the loading playbook's own columns. Never drop a column on write-back; unknown columns are kept as they are.
• A new entity (another playbook wanting a ledger) needs its row in the tables above, added by this atom's owner — ask before loading it.

On every run (before any new sourcing)

1. Read the ledger. If it does not exist, create it with the header; carry on.
2. Flag stale rows for free: rows past recheck_by / revet_by that are not yet contacted. Tell the human in one line ("12 people from the 2026-09-10 list haven't been contacted and are past their re-check date"). Re-checking them is paid → the loading playbook's spend step (the spend-gate skill, ../spend-gate/SKILL.md).
3. Ask for outcomes once, never require them — in the run's closing decisions, never before the first result: "Anything to report since <last list date> — replies, meetings, bounces, unsubscribes, declines, quotes? Paste, upload, or tell me in a sentence." Creators past a wave's read window (in creators/<date>/wave-1.md): also ask for each creator's views, orders on their code or link, and whether they posted. No answer → skip to step 6.
4. Match each reported outcome to a row by the match keys in order. Unmatched lines are shown back, never guessed onto a row. A stop, unsubscribe, complaint or bounce from an address not in the ledger gets a suppression-only row: the key_hash, the reason and the date — never the address (§ Suppression without personal data). Keeping the raw address needs the human's yes, asked once in the closing decisions.
5. Update status, status_date, outcome — and for people, first_touch_date when the first send is reported (never overwritten afterwards; it anchors rates and the cool-down), last_touch, reply_category, and recontact_after for a "not now, try in <month>" reply (and for creators quoted_rate when reported). Creators with wave results: score each against its wave_criterion → wave_result = win · partial · miss · no-data. A suppressing status sets suppressed with reason and date, and reduces the row to its suppression form (below); a timed exclusion sets outcome and the date it lapses. Suppression is permanent: a suppressed row is excluded from every later run, of every list, and is never un-suppressed except by the human's own words.
6. Exclude: hand the loading playbook the full exclusion set — every suppressed key_hash (the playbook hashes each candidate's keys the same way and drops any match), every live timed exclusion (including rows whose recontact_after date hasn't come), plus every key already listed and not yet resolved (so the next batch never repeats a person) — but not listed-sample-rejected rows.

Lessons (only when outcomes came back)

1. Compare by the entity's groupings (person: read replies and fixes with write-outreach's references/reading-results.md — many "not relevant" → targeting, many "wrong person" → titles, bounces → list hygiene): reply and meeting rate (person) or reply, quote-vs-range, post rate and result vs the wave criteria (creator).
2. Sample size (one rule everywhere): person — under 50 contacted in a group → counts and what to watch only, no rates judged; 50+ but under ~10 positive replies → "directional only"; creator — under ~5 in a group → "directional only". Said in the lesson.
3. Propose changes to the loading playbook's own rules (signal weights, segments, negative ICP; roles, tiers, price bands, vetting weights) and get the human's yes before any saved definition or plan changes.
4. Save the lessons to the entity's lessons file, then as a dated brand document through the save-to-brain skill (../save-to-brain/SKILL.md) (versioned: keep-history, tags [<playbook>, lessons]). It opens with "Do this first, because…" and the one change the numbers support most. Numbers and groups only — no names, emails, handles of private people or contact data. Named-creator operational notes ("missed the posting window twice") stay in the roster file's notes, never in the brain.
5. A lesson becomes a brand learning only through the save-learning skill (../save-learning/SKILL.md): the human's own words, one idea each, retiring its own superseded ones, text starting [outreach YYYY-MM-DD] (research-audience), [outreach-copy YYYY-MM-DD] (write-outreach) or [creators YYYY-MM-DD] (work-with-creators).

With run-memory (the run-memory skill)

The ledger and the run records hold different things and must agree:

• The ledger is the per-row truth — every person's or creator's status, dates and outcome. Nothing else stores rows.
• A run record holds only the summary — counts and rates by the groupings above, for a stated window — plus the ledger's path. It never holds a row, a key, a name, an email, a handle or a profile URL.
• Recount, don't carry. Each run's "since last time" recomputes its numbers from the ledger (statuses and dates as they stand today), never copies them from the last record. When a recount differs from what the last record said (the marketer reported outcomes late, a row was corrected), the ledger wins and the new record says "restated: <what changed>".
• Lessons and records: a lessons file is the analysis of outcomes (dated, keep-history); the run record links it and lists its proposed changes as recommendations with ids, so the next run can check whether they were adopted and what they did.
• Statuses a run record cites are only the ones the marketer reported into the ledger — never inferred.

Suppression without personal data

A suppression must last forever, but the person's details must not. So:

• key_hash = SHA-256 (hex, prefixed sha256:) of the normalised key: email lower-cased and trimmed; profile URL lower-cased with no protocol, www., query or trailing slash; name + root domain lower-cased, single-spaced, joined by @; creators platform:handle lower-cased. The same rule hashes every new candidate, so a suppressed person is blocked from every later list without the ledger keeping who they are.
• When a row is suppressed (bounced, unsubscribed, do-not-contact, never-again): keep key_hash, suppressed, suppressed_reason, suppressed_date, list_id; blank every other personal field (key, name, email, email_candidate, profile URL, route, notes). Counts by segment stay countable from list_id and the reason.
• Compute the hash with the host's code tool (shell or code execution). A host that can't compute one: don't store the address at all — tell the human to add it to their own sending tool's suppression list, and ask once whether to keep the raw address here so it stays blocked; store it only on their yes, labelled raw, kept on your yes (<date>).
• Rows the human pastes or uploads for scoring follow the same rule when they become suppressed.

Retention

Rows never contacted 12 months after date_listed are deleted from the ledger. Suppression rows are kept forever, in their hashed form only. Say how many were removed.

Never

• Never contact anyone, never send, never mark a row contacted on the agent's own inference.
• Never require outcomes to continue a run.
• Never drop a suppression to reach a list size.
• Never move personal contact data into the brain.