Delete identical comment campaigns across linked channel posts (#81) #37

Merged

backmeupplz merged 3 commits from fix/81-repeat-campaign into main

2026-10-12 17:31:05 +00:00

backmeupplz commented

2026-10-12 17:09:28 +00:00

(Migrated from github.com)

Agentboard #81 (https://agentboard.icefish-betta.ts.net/#81). Implements Nikita's option B ("yeah do B").

What

A bot-side repeat-campaign rule. When the same actor posts an identical comment (NFKC + lowercase + whitespace-collapsed, 40+ chars) under 3+ distinct linked channel posts within 10 minutes, every qualifying copy is deleted, including earlier copies the model kept. Later copies in the same window are deleted as they arrive.

Copies are tracked in a dedicated per-chat, per-actor CampaignLedger (10-minute TTL, 20 copies). It is independent of the classifier's kept-message history, so copies already deleted by the model, the cache or the rule still count towards the posts. A copy edited into an ineligible message is forgotten. A copy whose delete was attempted is never retried within the window. history.clear and the 10-message history cap don't hide copies, and a classifier failure doesn't skip the rule.
A "post" is only a same-chat is_automatic_forward reply whose sender_chat is a channel. Replies to another member's message (for example their own job request), non-linked channel chatter, and messages that are not replies never count and are never deleted by this rule.
Known gap (deliberate, conservative): a copy that replies to any other comment is exempt, because the bot can't tell a member's request from a random comment. A spammer replying to random comments under each post evades this rule; the model still judges each copy.
In scope by decision: one actor pasting any identical 40+ char text under 3 posts within 10 minutes is deleted whatever its content, whether an offer, a job search or a warning (Nikita accepted this for "looking for work"). Several different members echoing the same warning are untouched.
Only identical qualifying copies are deleted. Interleaved messages from the same actor and other actors' histories are left alone.
When the model deletes a copy itself, the rule also deletes the earlier kept copies that its weak links (~0.72 < 0.75) would have missed.
The Laya model, checkpoint, prompt, classifier input/context serialization and the 0.80 gate are all unchanged. No forced score. The classifier input and its kept-message history are untouched; the ledger is separate state.
Content-free repeat_campaign_detected log and rule on spam_deleted. Audit snapshots exist for every deleted copy. Stats source stays jev.
Cache: CACHE_POLICY_REVISION is a digest of bot.ts and friends, so it changes automatically. campaign.ts was added to the digest list.

Why the model can't do this

See the #81 findings: identical prior copies lower Laya's score, because its training history only ever held kept messages (survivorship bias). Retraining is forbidden by decision.

Tests

src/campaign.test.ts: normalization, post detection, ledger threshold/selection, deleted-copy counting, edits, TTL expiry and bound.
src/bot.test.ts: real grammY routing with channel-post reply entities. Covers the incident (model keep 0.79/0.61/0.54 → deletes 1,2,3, and a 4th copy deleted on arrival), the model-delete widening, a model-deleted first copy still counting, classifier failure, 11 messages of chatter between copies, members echoing a warning/report (kept), the same actor mass-posting a warning (deleted), an edited-away copy (forgotten), a failed delete (not retried), two posts / same post / edits / other actor (not triggered, other history kept), requested-offer replies to members, interleaved messages, short chatter, non-linked channel, distinct controls, and audit per copy.
src/campaign.live.test.ts (gated RUN_LIVE_LAYA=1 JEV_URL=…): before/after on the approved ckpt-v2 (laya-jev-ckpt-v2-examples-856e52e3, checkpoint sha d899eecd…) at 0.80, through the real handler. SYNTHETIC reconstruction:
- incident, no bio: model keep 0.7723 / 0.6236 / 0.4989 → after: deletes 1,2,3
- incident, decorative bio: keep 0.4711 / 0.4021 / 0.4456 → after: deletes 1,2,3
- controls never trigger the rule. Pre-existing model verdicts are unchanged: requested offer to member requests 0.892 delete (no bio) / 0.6555 keep (bio); job seeker 0.9263 (no bio) / 0.8203 (bio) delete; wage, warning, report and chatter are kept.

bun run check: 486 pass, 0 fail.

Agentboard #81 (https://agentboard.icefish-betta.ts.net/#81). Implements Nikita's option B ("yeah do B"). ## What A bot-side repeat-campaign rule. When the same actor posts an identical comment (NFKC + lowercase + whitespace-collapsed, 40+ chars) under **3+ distinct linked channel posts** within 10 minutes, every qualifying copy is deleted, including earlier copies the model kept. Later copies in the same window are deleted as they arrive. - Copies are tracked in a dedicated per-chat, per-actor `CampaignLedger` (10-minute TTL, 20 copies). It is independent of the classifier's kept-message history, so copies already deleted by the model, the cache or the rule still count towards the posts. A copy edited into an ineligible message is forgotten. A copy whose delete was attempted is never retried within the window. `history.clear` and the 10-message history cap don't hide copies, and a classifier failure doesn't skip the rule. - A "post" is only a same-chat `is_automatic_forward` reply whose `sender_chat` is a channel. Replies to another member's message (for example their own job request), non-linked channel chatter, and messages that are not replies never count and are never deleted by this rule. - Known gap (deliberate, conservative): a copy that replies to *any* other comment is exempt, because the bot can't tell a member's request from a random comment. A spammer replying to random comments under each post evades this rule; the model still judges each copy. - In scope by decision: one actor pasting any identical 40+ char text under 3 posts within 10 minutes is deleted whatever its content, whether an offer, a job search or a warning (Nikita accepted this for "looking for work"). Several *different* members echoing the same warning are untouched. - Only identical qualifying copies are deleted. Interleaved messages from the same actor and other actors' histories are left alone. - When the model deletes a copy itself, the rule also deletes the earlier kept copies that its weak links (~0.72 < 0.75) would have missed. - The Laya model, checkpoint, prompt, classifier input/context serialization and the 0.80 gate are all unchanged. No forced score. The classifier input and its kept-message history are untouched; the ledger is separate state. - Content-free `repeat_campaign_detected` log and `rule` on `spam_deleted`. Audit snapshots exist for every deleted copy. Stats source stays `jev`. - Cache: `CACHE_POLICY_REVISION` is a digest of `bot.ts` and friends, so it changes automatically. `campaign.ts` was added to the digest list. ## Why the model can't do this See the #81 findings: identical prior copies *lower* Laya's score, because its training history only ever held kept messages (survivorship bias). Retraining is forbidden by decision. ## Tests - `src/campaign.test.ts`: normalization, post detection, ledger threshold/selection, deleted-copy counting, edits, TTL expiry and bound. - `src/bot.test.ts`: real grammY routing with channel-post reply entities. Covers the incident (model keep 0.79/0.61/0.54 → deletes 1,2,3, and a 4th copy deleted on arrival), the model-delete widening, a model-deleted first copy still counting, classifier failure, 11 messages of chatter between copies, members echoing a warning/report (kept), the same actor mass-posting a warning (deleted), an edited-away copy (forgotten), a failed delete (not retried), two posts / same post / edits / other actor (not triggered, other history kept), requested-offer replies to members, interleaved messages, short chatter, non-linked channel, distinct controls, and audit per copy. - `src/campaign.live.test.ts` (gated `RUN_LIVE_LAYA=1 JEV_URL=…`): before/after on the approved ckpt-v2 (`laya-jev-ckpt-v2-examples-856e52e3`, checkpoint sha d899eecd…) at 0.80, through the real handler. SYNTHETIC reconstruction: - incident, no bio: model keep 0.7723 / 0.6236 / 0.4989 → **after: deletes 1,2,3** - incident, decorative bio: keep 0.4711 / 0.4021 / 0.4456 → **after: deletes 1,2,3** - controls never trigger the rule. Pre-existing *model* verdicts are unchanged: requested offer to member requests 0.892 delete (no bio) / 0.6555 keep (bio); job seeker 0.9263 (no bio) / 0.8203 (bio) delete; wage, warning, report and chatter are kept. `bun run check`: 486 pass, 0 fail.