# Outreach: draft, lint, approve, send The outreach module never sends or fetches a message. It records a local lifecycle: draft, lint, approve, sent, reply, close or drop. Sending belongs to a separate tool and requires the operator's authorization for the actual message. Set `RH_CONTACTS` to a private JSON file containing an array of objects with a nonempty `name` string. Names must be unique ignoring case. An optional `language` field holds the recipient's language as a code such as `pt`; unset or `en` means English and changes nothing. Extra fields may hold addresses for your separate sending tool; the harness uses names and that one field. Nothing in this repo contains contacts. Do not commit that file. Ledger metadata rejects email-address-shaped strings; draft bodies remain private files. Set `RH_VOICE_COMMAND` to a shell-quoted executable and arguments. It receives the draft on stdin; lines starting `FAIL:` or a nonzero exit refuse lint. It is executed as an argument vector without a shell. For a social reply, the harness appends `--reply`. A missing checker refuses. No personal voice policy is provided. The tiny checker under tests is a synthetic fixture, not a writing standard. `OUTREACH.toml` sets question limits and follow-up timing. Optional `science_checks = true` invokes `RH_SCIENCE_CHECK_COMMAND`, which receives JSON with `text`, `language` and `attachments` on stdin and uses the same failure convention. That replaces an installation-specific Python import. A checker that ignores `language` is unaffected by its arrival. **Two-language email.** When a recipient's `language` is not English, that language is the default for the whole email, and the email opens and closes in it. English is required only where the content is technical: numbers, variables, equations, methods, measurements. Attachments are always English. The rule holds even for a recipient who writes their own science in English. Lint fails a technical paragraph in the recipient's language, an opening or closing that is not in it, and an attachment that is not English; ordinary prose in English anywhere is a warning, because it belongs in their language too. A paragraph counts as technical when it carries a maths symbol, a variable-like token, a statistical acronym, a measured number (a decimal, or a number with an abbreviated unit such as `7 s`), a run of three or more bare numbers, or a dense run of distinct technical terms. Counted numbers are not measurements: a year ("I left in 2009"), a duration in words ("four years ago", "thanks for the 20 minutes"), a numbered reference ("Remark 4.2", "Table 3"), a link and the European Portuguese intensifier "100%" are all ignored. Neither is the vocabulary alone enough ("a separate model re-checked the result" is prose), and a term repeated, as in "data science and data engineering", counts once. **Independent re-check and pod-touched evidence.** Two gates, from the friends-lane postmortem (Ricardo, 11-05). A draft names the claim directories it reports with `--claim DIR`, and lint fails when a claim has no recorded independent re-check (`RECHECK_MISSING`), when the re-run did not reproduce it (`RECHECK_FAILED`), when the re-check dropped it (`RECHECK_DROPPED`), or when evidence produced on a pod was last re-checked before the configured host incident (`POD_EVIDENCE_STALE`); `CLAIMS_REQUIRED`, `CLAIM_MISSING` and `RECHECK_LEDGER_MISSING` cover a missing claim, directory or ledger, and `POD_EVIDENCE_UNDATED` warns when pod evidence has no incident to be dated against. Both read one artifact the lane already writes: `RECHECK.csv` in a re-check root, one row per claim, with the claim's directory name in `id`, the independent re-run's outcome in `rerun` and the re-check's own verdict in `verdict`; an optional `utc`, `date` or `checked_utc` column dates a row, and the ledger's mtime dates it otherwise. A claim counts as pod-touched when its directory holds entries named `pod_*` or `pod-*`, which is deliberately generous: a leftover pod script is answered by a re-check after the incident. Configure in the lane's `OUTREACH.toml`: `recheck_roots` (a list of directories or globs), optional `pod_incident_utc` (without it, pod-touched evidence only warns, since an installation with no incident has nothing to re-run after) and optional `require_claims = true` (a lane whose emails always report results). A misconfigured value is refused rather than quietly disabling the gate, and `claims` also reaches `RH_SCIENCE_CHECK_COMMAND` in its JSON payload. [claim_evidence.py](../harness/claim_evidence.py). **No infrastructure or operations in a message to a person (Ricardo, 11-06).** A message to a friend carries only what the friend needs; our provider, our pods, a security incident, a key rotation, the harness and its watchdogs, and why something was re-run are none of them. The rule came from one sentence in an email of 11-05 — "The first two ran on a RunPod server that has since had a security incident" — which the recipient wrote back to ask about: it cost a round trip, moved the subject from the result to our plumbing, and reported a host compromise he had no way to act on. Lint fails any non-public message (email, `x_dm`, `linkedin`) carrying one of those terms, as `OPS_DETAIL: `; a public post is not covered, because naming one's own tools and hosts in public is a different decision. The vocabulary is deliberately narrow, because these are working scientists and several ops words are ordinary words in their fields: bare `incident` is an incident wave, `lane` is a traffic lane, `compromised` is a compromised cell line, `modal` is a modal split, and `key` is a key result. So it matches provider names, unambiguous ops nouns and multi-word phrases (`security incident`, `key rotation`, `server crashed`) rather than their component words, and two ambiguous words are checked with their sentence for company: `pod` fails beside rented-machine or incident wording and not in "each seed pod holds four to six ovules", and an agent's name (`codex`, `claude code`) fails only beside incident wording, since the tool that did the work is part of the story. When the recipient raised the subject themselves, `rh outreach draft --asked-about TERM` drops that one term to a warning so a reply can answer them; it takes a term from the check's own list, so a typo cannot quietly turn the check off and the ledger records our word for the subject rather than the recipient's wording. Only the matched term is ever reported, never the surrounding text. `ops_mentions.terms(text)` returns those term names as a set, for a caller that wants to decide for itself — running it over an incoming message and passing what it finds as `asked_about` is how a lane with the recipient's own text in hand can exempt what they raised, without parsing a failure message. [ops_mentions.py](../harness/ops_mentions.py). **Paper-cut drafts.** With `RH_VOICE_BANKS` set (a JSON file naming the author's private message banks and their languages), lint requires every draft to be cut from the author's own messages: `DRAFT.cuts` beside the draft lists the bank lines, one `BANK:LINE` each, and lint fails as `VOICE_CUTS_MISSING`, `VOICE_CUTS_BAD`, `VOICE_CUTS_LANGUAGE` (an English bank cut into a Portuguese draft), `VOICE_CUTS_WORDS` (under `voice_cuts_min_words`, 0.70, of the words from the cited lines) or `VOICE_CUTS_PAIRS` (under `voice_cuts_min_pairs`, 0.40, of the word pairs). Words of the message given with `--incoming` count as subject swaps; greeting and sign-off lines are left out. A lane sets `voice_cuts = "off"`, or `"on"` to refuse when no banks are configured (`VOICE_BANKS_MISSING`); the default `"auto"` follows the banks. `rh voice-cuts DRAFT [--incoming FILE]` runs the same check alone and records a pass by the hash of the normalized text in `RH_VOICE_CUTS_STATE`, which the email hook reads for a send that has no outreach row. [voice_cuts.py](../harness/voice_cuts.py). Language identification is a small offline heuristic over function words and letters English does not use, currently covering Portuguese against English ([mixed_language.py](../harness/mixed_language.py)). It never touches the network and never loads a model. When a passage is too short or too mixed to call, it reports nothing rather than blocking a send; short openings and closings are decided by their greeting words instead. It reports paragraph positions, never the text, so the lane ledger holds no part of a private draft. Lint checks question count, links, recent contact collisions and similarity across drafts. It normalizes names, links and titles; first/last sentence word-set Jaccard thresholds are 0.6 and body 5-gram Jaccard is 0.35. Similarity is reported as a **warning** and never fails a draft: two emails sharing a conventional greeting or sign-off is normal, and the two-language rule above makes a shared sign-off likelier still. The voice check is what guards against a templated message. These heuristics can miss templated spam or flag legitimate messages. A documented collision override is possible, and `sent --utc` records a message that went out before the ledger caught up. **A message that already went out.** `sent --utc` records a send the ledger caught up with late, but it still needs an approved draft, and approval still needs a lint PASS — so a message that already left and whose text fails today's lint could not be recorded at all. On 11-05 that cost a thread its record: the row had to be dropped, and the friend's reply had nothing to attach to. `record-sent ID --message-id M --via api --utc WHEN --reason WHY` records it anyway. Lint still runs and every finding is kept on the row under `recorded_after_the_fact` (`lint_ok`, `findings`, the operator's `reason`), as warnings rather than a refusal; the follow-up falls due from the real send time, and `reply` can attach to it. It is bookkeeping, never authorization: the row gets no `approval`, and the send hook matches submitted bytes against an approved entry only, so a recorded row can never authorize a send. It refuses a future `--utc`, a row already recorded as sent, and an approved draft (which belongs in `sent`), and `approve` still refuses the row afterwards. Approval requires a current passing lint hash and the operator's approval quote. Any byte change invalidates approval. `sent` requires matching approved bytes and records a message identifier; follow-ups become due after the configured interval. The optional email hook checks exact submitted bytes against the latest approved ledger entry. Approval enforcement defaults to warning; set `RH_OUTREACH_GATE=enforce` in the sending environment. Voice failures deny in both modes. The hook cannot authenticate the person behind an approval quote and matches message bodies, not recipient authorization. ## Reciprocate A reply mirrors the message it answers ([reciprocate.py](../harness/reciprocate.py), [test_reciprocate.py](../harness/tests/test_reciprocate.py)). Draft it with their latest message, `rh outreach draft ... --incoming FILE`, and lint compares the two, with quoted text removed from both: - `RECIPROCATE_LENGTH`: our body is within `reciprocate_min_ratio` to `reciprocate_max_ratio` (0.7 to 1.6) times their word count, up to `reciprocate_max_ratio_answering` (2.0) when they asked something; a message under `reciprocate_short_words` (40) may get that many words back. The floor never exceeds `reciprocate_floor_cap` (120), a common cap on a reply to a friend, so a long message cannot demand more than that; a shorter reply takes `--short-ok` in the owner's words. Life questions and first-person answers are matched in English and Portuguese ("where are you", "estás em", "eu vim", "agora sou"). - `RECIPROCATE_GREETING` and `RECIPROCATE_SIGNOFF`: they greet, we greet, in the same register (hi or ola against dear or caro), by name when they used ours; they sign off, we sign off. - `RECIPROCATE_WARMTH`: "good to hear from you" in, something as warm out. - `RECIPROCATE_QUESTION`: every question gets an answer; asked how we are, we say what we're doing and ask back. - `RECIPROCATE_FAMILY`: they mention family, ours goes in; they don't, ours stays out. - `RECIPROCATE_LANGUAGE` and `RECIPROCATE_EMOTICON`: the language they wrote in, and no more emoticons than they used. When it is our move (the contact's ball is with us) or the draft has `--requested`, a draft without `--incoming` fails as `RECIPROCATE_INCOMING_MISSING`. The matching is lexical, English and Portuguese, so each finding is easy to check: it quotes their line and ours. It does not judge tone; the person approving still reads both. ## Science emails From one lane's postmortem of its emails to scientists (11-08): a plain question first got 3 replies of 3, a paper result first 3 of 10, a bug report first 2 of 8, and every "no" was "this code isn't used" or "the number isn't the interesting part". With `science_email_rules = true` in OUTREACH.toml, an email that reports a result, a bound, a bug, a PR or a paper (by its words, its `--kind`, or a non-image attachment) fails as `SCIENCE_OPENER` when its first paragraph asks nothing (unless their message asked us something), `LIFE_LINE_MISSING` with no line about the sender's own life or work (`life_words` adds the lane's own words, such as places, kept in the private lane config), `USERS_MISSING` for a bug with no evidence of use (the contact's `users` field, a `--target` whose users check passed, or `--users-override`), `MEANING_MISSING` for a tighter number without what it means, and `CALL_MISSING` when the contact has replied before and no call is offered (`call_link`). `--opener-ok`, `--life-ok`, `--meaning-ok` and `--nocall-ok` take the owner's words. The same flag adds `AWAITING_REPLY`: no new science email to someone whose last one (follow-ups included) has no reply yet, unless they asked for it (`--requested`) or their contact row says `cadence: fast`. [science_email.py](../harness/science_email.py). ## Email to a friend A friend, on a call (11-09), said emails cut line by line from the sender's own messages read as AI: the sender writes them one sentence, and these were well-formed paragraphs. To a contact whose `group` is in `friend_groups` (default `["friend"]`), an email fails as `FRIEND_LENGTH` over `friend_max_words` (60, greeting, sign-off and quoted history left out), `FRIEND_PARAGRAPHS` over `friend_max_paragraphs` (2), and `FRIEND_BULLETS` with any bulleted or numbered list. If the contact's `engagement` field says they noticed AI or want one sentence, any draft fails as `AGENT_NOTICED` unless the owner wrote it: `--owner-wrote` in the owner's words, which also takes the text out of the friend and paper-cut checks. For friends the science-email life line, meaning and call offer are optional and the reciprocate length floor is off; the question first stays. [friend_email.py](../harness/friend_email.py). ## Bug-hunt lane A lane that finds bugs in other people's research code and writes to them about it. Its rules are checks that refuse, each with a code and a test ([bughunt.py](../harness/bughunt.py), [contacts_ledger.py](../harness/contacts_ledger.py), [test_bughunt.py](../harness/tests/test_bughunt.py)); none is prose an agent may skip. An example lane config is in [examples/friends-lane](../examples/friends-lane/OUTREACH.toml). 1. **The target is recent and active.** `rh bughunt target-check LANE OWNER/REPO` reads the GitHub API: archived or not, last commit on the default branch, last release or tag, commits in the last 365 days, last issue or pull-request activity, and an "unmaintained" notice in the README. It passes when the repository is not archived, carries no such notice, and has a commit or release within `target_max_age_days` (default 365). A paper alone never passes, and a data deposit (Zenodo, Mendeley Data, figshare, OSF, Dryad) is not maintained code. The verdict and its numbers go to the lane's `TARGETS.jsonl`. An email names its code with `--target OWNER/REPO`, and lint fails it as `TARGET_UNCHECKED`, `TARGET_FAILED`, or `TARGET_STALE` (last pass older than `target_check_valid_days`, default 7). With `target_required_kinds = ["code"]`, a code email without a target fails as `TARGET_MISSING`. `--target-override REASON` (the recipient asked about older code) turns these into recorded warnings. Recent is not enough: someone outside the project must use it. The same check counts outside accounts (authors of issues and pull requests and owners of forks from the last `usage_window_days`, 1095, minus contributors, the owner and bots; an old fork is not someone using it now), stars, and downloads for each package named with `--package INDEX:NAME` (`pypi:`, `cran:`, `bioconda:`, `conda-forge:`), and passes usage on any one of: `usage_min_outside_accounts` (3), `usage_min_stars` (25), `usage_min_downloads_month` (100, PyPI and CRAN) or `usage_min_conda_downloads` (1000, all time). A code email to an unused target fails as `TARGET_UNUSED`; `--users-override "their words"` records the author's own statement that it is in their current work and turns it into a warning. The printed headline and the exit code are PASS only when both recency and users pass. 2. **One email per person per day, batched.** With `emails_per_contact_per_day = 1`, lint and `sent` refuse a second email to the same contact within 24 hours (`DAILY_LIMIT`, across lanes), and lint warns when other unsent drafts to that person exist. `rh outreach batch LANE CONTACT --file OUT` merges them into one new draft and drops the originals; the merged draft carries a marker line and fails lint (`MERGE_UNEDITED`) until it has been edited into one email. A send the recipient explicitly asked for is a reply, not outreach: draft it with `--requested "their words"`, and neither the daily limit nor the recent-contact collision holds it (both are recorded as warnings). 3. **Contacts kept current.** `sent`, `reply` and `bounce` update the contact: last out, last in, whose move it is, and dead addresses. The contacts file (`RH_CONTACTS`) stays the one record and the harness never writes it: updates go to a private ledger (`RH_CONTACT_UPDATES`, default `contact-updates.jsonl` in the harness root, mode 0600) that is read over the file, and `rh outreach contact pending` prints what the file does not have yet, for its owner to land. Addresses are read from a row's `email` field and from free text in `channels`; an address followed by `BOUNCED` is dead. A draft to a contact whose every address bounced fails as `ADDRESS_DEAD`. A contact's `engagement` field says, from their own words, how they prefer to be engaged ("treat people like they want to be treated"); lint prints it on every draft to them, and when it says to just send, a permission or format question ("do you want me to...", "PR or patch?") fails as `ENGAGEMENT`. Set it in the contacts file or with `rh outreach contact engagement NAME TEXT --source S`. `rh outreach contact find NAME [--url PAGE]` reads public sources (OpenAlex for recent works and the ORCID iD, ORCID's public record, and any lab or university page given) and lists candidate addresses with the URL each came from; `--record ADDRESS` keeps one, and refuses an address that was not found in a source read. 4. **More checking for some readers.** A contact with `scrutiny: extra` (set in the contacts file, or with `rh outreach contact scrutiny NAME extra`) needs `extra_rechecks` (default 2) independent re-check directories named with `--recheck DIR`, each holding a verdict file (`VERDICT.md`, `VERDICTS.md`, `RECHECK.csv` or `verdict.json`) that records a confirmed result (`SECOND_RECHECK_MISSING`), and a passing review of the exact final text, `rh outreach review LANE ID --by WHO --verdict pass` (`TEXT_REVIEW_MISSING`). Any edit after the review invalidates it. 5. **The question we want answered.** `required_questions` lists keyword sets, one per question, with `required_question_names` and `required_questions_kinds`; an email of those kinds fails as `QUESTION_MISSING` unless a sentence ending in "?" carries a keyword from each set. Keywords may cover more than one language. `rh outreach contact answered NAME QUESTION --source S` records that a contact already answered, which turns the finding into a warning. 6. **Counts.** Every email draft says what it reports, `--kind paper|code-bug|code-speedup|other` (`KIND_MISSING`; repeat `--kind` for an email that carries more than one, such as refereed bugs and a verified speedup, and the rules of each kind apply; rows stored as `code` before the split count as `code-bug`, and a config list naming `code` covers both code kinds); an email that went out before that rule gets its kind with `rh outreach set-kind LANE ID KIND --reason R`, bookkeeping that is refused for an unsent draft. `rh outreach stats [LANE] [--contact NAME] [--json]` prints emails sent, replies, reply rate, bounces and the split by kind (an email counts once in the totals and once under each kind it carries), per contact and in total, with last out, last in and whose move it is. `--aggregate` prints totals only, with no names, for publishing. 7. **Speedups, verified.** A `code-speedup` email names its speedup with `--speedup DIR`, a directory holding `SPEEDUP.json` (the target, a local clone, their commit, the patch, the command that writes the outputs, the outputs, an optional numerical tolerance, the number of runs, the profile and the hot spot with its numbers). `rh bughunt speedup-verify LANE DIR` checks out their commit twice, applies the patch to one copy, runs both versions alternately, compares every output (byte-identical, or within the stated tolerance), and records median wall time, peak memory and the speedup in `VERIFY.json`. It refuses on a controller (`RH_NO_LOCAL_COMPUTE=1`), because it runs other people's code at size. It fails as `SPEEDUP_PROFILE_MISSING`, `SPEEDUP_PATCH_FAILS`, `SPEEDUP_RUN_FAILED`, `SPEEDUP_OUTPUT_MISMATCH`, or `SPEEDUP_TOO_SMALL` (under `speedup_min_ratio`, default 1.5x, or under `speedup_min_seconds` saved, default 1.0). Lint then fails the email as `SPEEDUP_UNVERIFIED` (none, or a failed verification), `SPEEDUP_STALE` (the patch or manifest changed after verification), `SPEEDUP_STALE_BASE` (their latest commit, from the last target check, is not the one verified), or `SPEEDUP_OVERLAP`: the target check lists their own open pull requests and branches about speed, and a lab already working on it is asked first, recorded with `--overlap-checked "how"`. For an extra-scrutiny contact, the second independent check is a second verification of the same patch, run separately. Two bookkeeping rules keep the ledger whole without authorizing anything. `rh outreach adopt LANE ID --kind K --reason R` turns a send logged by hand (a row with `sent_at` and no status) into an ordinary sent row, so counts, the daily limit and follow-ups see it. And `draft` refuses a file in a temporary directory: a restart empties it, and the ledger loses the text that later drafts are compared against. Every draft is also kept by its sha256 when it is drafted and approved (`RH_DRAFT_ARCHIVE`, default `~/.local/state/research-harness/drafts`, mode 0600), and a row whose file is gone is read from there; a comparison draft that is gone everywhere is a warning, not a failure on every later message. The X posting gate (`xr` and the browser hook) refuses an unapproved post unless `RH_X_GATE=warn` is set explicitly; until 11-09 an unset value meant warn. ## Bug-hunt referees From three published skill sets compared with our referee step (11-08; Cloudflare's security-audit-skill most of all), as code ([bug_verdicts.py](../harness/bug_verdicts.py), [bug_coverage.py](../harness/bug_coverage.py), [test_bug_verdicts.py](../harness/tests/test_bug_verdicts.py)): - **Refuted is not blocked.** NOT REPRODUCED mixed "the claim is wrong" with "I could not run it". A referee's `verdict.json` (schema `bug-verdicts/1`, strict: unknown fields, a missing reason or a NOT REPRODUCED fail) says REFUTED with the reason, or BLOCKED with the one fact that would settle it (`needs`) and how to check it (`check`). Refuted claims go to the lane's `REFUTED.jsonl` under a fingerprint of repo, file and function; blocked ones to `BLOCKED.jsonl`, as questions. - **No re-claiming.** `rh bughunt claims-check LANE claims.json` refuses a claim at a refuted place (`REFUTED_BEFORE`) unless the file changed (its `file_sha256` differs) or the claim carries `differs_from` with that fingerprint and why. `referee-prompt` runs the same check first. - **One referee prompt, tested on false claims.** `rh bughunt referee-prompt LANE claims.json WORKDIR` writes the standard prompt ([templates/bug_referee.md](../harness/templates/bug_referee.md)) and the claims, renumbered and shuffled with at least one planted false claim: a real claim with observed and expected swapped, or one claim's root cause placed at another's function, or a list given with `--plants`. The key stays in the lane's `PLANTS.jsonl`. `rh bughunt verdicts LANE verdict.json` validates the file (`VERDICT_MALFORMED`: run the referee again, never repair it), refuses one with no planted batch (`VERDICT_UNPLANTED`) or a missing claim (`VERDICT_INCOMPLETE`), and rejects the whole file when a planted claim was confirmed (`REFEREE_FOOLED`). The prompt needs the hunter's family (`"hunter": {"model", "family"}` in the claims file, or `--hunter-family`; `HUNTER_UNDECLARED` otherwise), kept in `PLANTS.jsonl` and never shown to the referee, and a verdict from the same maker's family (GPT-5 and GPT-6 Codex count as one, as do Opus and Sonnet) is rejected whole (`REFEREE_SAME_FAMILY`; 11-09, a GPT-6 Codex hunter was refereed by GPT-6). Batches written before that record pass with `REFEREE_FAMILY_UNCHECKED`. The referee's family comes from the launcher's selection receipt (`--receipt JOB.log.selection.json`, or found by the verdict folder's name under `RH_LAUNCH_RECEIPTS`), not from the referee's own `referee` field: on 11-09 Gemini 4 Argon run through Codex reported itself as GPT. A verdict with no receipt is checked on its self-report and flagged `REFEREE_FAMILY_SELF_REPORTED`; one whose self-report disagrees with the receipt is flagged `REFEREE_SELF_REPORT_WRONG`. `rh bughunt referees LANE` prints each referee's confirm rate and plants caught, and warns at a plant missed or over 90% confirmed across 20 claims. - **Lint.** A code-bug email whose `--recheck DIR` holds a `verdict.json` fails as `VERDICT_UNRECORDED` until that file was recorded and accepted, and as `BUG_REFUTED_BEFORE` when a confirmed claim sits at a place another referee refuted (a third referee decides). Only accepted structured verdicts count toward `extra_rechecks`. - **Coverage.** Each hunt round records its units (module times bug class, covered, candidate or deferred) with `rh bughunt coverage LANE record FILE`; a round that covers nothing new at that commit is refused (`COVERAGE_REPEAT`). A target is done after two rounds, the last with no new candidate and nothing deferred (`coverage LANE show OWNER/REPO`). A code-bug email warns `COVERAGE_UNRECORDED` or `COVERAGE_ONE_PASS` until then. The planted claims are made from the real ones, so a careful referee can sometimes see which of two contradicting claims is false; they catch a referee that confirms without running anything, which is the failure we had. ## Growing a small account on X A research account starts with no followers, and its first results reach nobody if it only posts originals. The advice below is Nikita Bier's, from his posts while he ran product at X (July 2025 to October 2026), which is the closest public source to the ranking team's own view. What the ledger can check is enforced; the rest is guidance, and this page says which is which. 1. **Reply first, originals second.** Bier's ladder: 0 to 500 followers, be a reply guy; 501 to 2K, niche bangers [1]. Below about 500 followers, effort goes into replies under larger accounts in the lane's field, where there is a real thought or question. Above it, one strong original per real result, with its figure. A ranking change he announced in July 2026 had replies up 3.15% and small-account reach up 1.19% [2]. *Guidance, not checked: the ledger does not know the follower count.* 2. **Space posts, never a burst.** The For You feed loads posts in batches and pushes extra posts from one author to later batches, so posts sent together lose reach (reported secondhand [7]). Originals are spaced by `x_original_gap_minutes` (default 120, or the older `RH_X_MIN_GAP_MINUTES`); lint, approve and the posting gate refuse one that falls inside the gap of another original from any lane. A round's replies are approved together but leave one at a time: the posting gate refuses a reply within `x_reply_gap_minutes` (default 3) of another reply's send or posting attempt, and its refusal names the time the next one may go. Approval is not counted for replies, only the moment one goes out. 3. **One line to a light post.** "Never respond to playful humor with an essay" [3]. A joke or a one-liner. *Guidance, not checked: lint does not see the post being answered.* 4. **No filler.** A greeting with nothing in it ("gm", "good morning", "happy Friday") fails lint as an original. A greeting with something real after it is not matched. 5. **Original work, not reposted reach.** Bier said X would penalise big accounts reposting others' videos hours later, and reward original creators [4]. Short talking videos about one result are his top lever for small accounts [5]; they need the person on camera, so they are proposed to them, never generated. 6. **A person approves replies.** In July 2026 X removed 42,000 accounts that automated replies with chatbots: "using AI to programmatically engage with users without a human in the loop runs counter to our mission" [6]. With `x_reply_human_approval = true` (the default), `approve` refuses an `x_reply` approved with the standing-approval quote, so each round's replies need a person's approval in their own words. A lane whose owner has decided otherwise sets it to `false` in its `OUTREACH.toml`, and that is the owner's decision to record, not the agent's. A post deleted from X after it went out (a wrong video, say; X has no edit) is recorded with `rh outreach retract LANE ID --reason TEXT [--replaced-by URL]`. It refuses while X's public oEmbed still returns the post, and when X cannot be asked, takes `--deleted-confirmed` with how the deletion was checked. A retracted post, like a dropped one, holds no slot in the gap, the daily limit, the 48 h topic count or reply spacing, so its replacement can go out at once; the row keeps its `sent` record. ```toml # OUTREACH.toml (all optional; these are the defaults) x_original_gap_minutes = 120 x_reply_gap_minutes = 3 x_reply_human_approval = true ``` References: 1. Nikita Bier, "Twitter Growth Strategy", 2022-03-03 (he reposted it in September 2025 as "Still true today"). https://x.com/nikitabier/status/1499452301185929216 2. Nikita Bier, ranking change results, 2026-07-14. https://x.com/nikitabier/status/2076917480518558035 3. Nikita Bier, reply on comms advice, 2026-02-05. https://x.com/nikitabier/status/2019346324173537337 4. Nikita Bier, on reposted videos and original creators, 2026-05-25. https://x.com/nikitabier/status/2058810803625881883 5. Nikita Bier, on talking videos, 2026-04-12. https://x.com/nikitabier/status/2043404323485589763 6. Nikita Bier, 42,000 chatbot-reply accounts removed, 2026-07-24. https://x.com/nikitabier/status/2080747924380856519 7. X trending summary on burst posting, where he called it "not ideal" (secondhand; the post itself not found). https://x.com/i/trending/2060629176668094923 Code: [outreach.py](../harness/outreach.py), [email-voice-gate.py](../tools/email-voice-gate.py). Tests: [test_outreach.py](../harness/tests/test_outreach.py), [test_outreach_gate.py](../harness/tests/test_outreach_gate.py).