Handoff: Meta ad fatigue watcher, merged build (20 Oct 2026)

Status: built, backtested, published. Cron and email still OFF. Nothing in any ad account was changed.

Live example (Pott): https://pott-fatigue-watch.pages.dev/ (19 Oct run) and https://pott-fatigue-watch.pages.dev/2026-08-31.html (the Ad 707 case). De-indexed but public.

What this is

One daily watcher, scripts/meta-fatigue-watch/, that judges every ad still taking real money and puts it on one of five stages, each tied to one action:

Too new and too small are labelled separately.

What changed today

Three fatigue systems existed. The June weekly scoring (raw week on week, age as a trigger), a September daily agent on an unmerged branch that was never recorded anywhere, and the October watcher. Today the best of September was folded into October:

• 28-day baseline instead of 14. Under 21 days of baseline an ad is low confidence and stops at Close, unless it is a 10%+ spender below breakeven.
• Cost per sale before ROAS. ROAS moves with basket size, which is not the creative's doing. A cost-per-sale rise with ROAS up is logged as bigger baskets and not counted. ROAS keeps one job: the breakeven line.
• A chance test on small numbers. At the baseline rate this week's spend should have bought N sales. If seeing this few is more than 20% likely by chance, the fall is not counted (10% for strong).
• Hold rate (ThruPlays over 3-second views) joins click rate and hook rate in the Attention family.
• Click rate net of the account, never worse than the ad's own move.
• Retargeting ad sets get frequency ceilings of 5 and 8 instead of 2.5 and 3. Spotted by ad set name, never by the word "warm" (Pott prospecting is called Cold & Warm), and never where the name excludes customers.
• Check-site guardrail. Purchases collapse while click rate holds means stock, page or pixel, not creative. Routed to Monitor with a site-check action.
• Best-week misfire fixed. Best week on its own, with this week's ROAS and purchases not falling, cannot make the return family "hard". Pairs 446 no longer flags.
• "Cut today" needs a real breakeven. With no breakevenRoas in a client config the bar is only a campaign average, so Fatigued is capped at Close with a note. Pott 2.46x from the margin sheet; H&H 4.0x, Pairs 2.0x, SHF 3.0x from Casey (20 Oct).
• Refresh feed. Every run writes state/refresh-feed.json with each flagged ad's copy, headline, video id and preview link, so the replacement brief starts from the live creative.
• Report. KPIs on top, flagged-ad cards, click any ad for its last 21 days (eight small charts, 7-day rolling line, ad vs account, breakeven line), every-ad grid, then the nine-gate decision tree, the eight checks with trip wires and why, and the rule rationale.
• Backtest harness backtest.cjs: judges past dates and checks what every judged ad did over the following 14 days, per stage.

Kept from October over September: breakeven line, material-only and spend-held gates, over-scaled and pulled-back as separate outcomes, own-best-week slow decay, signal families, five action-defined stages, account-wide guard, persistence against the ledger a week earlier (never yesterday's overlapping run).

Verification

The four hand-picked cases all land where they should:

Pott live run for 13 to 19 Oct: nothing above Monitor. Ad 940 is pulled back (Meta halved its spend), 11 Christmas ads become judgeable from about 21 Oct.

Next-14-day backtest. 113 judged ads across 30 Jul, 13 Aug, 27 Aug, 10 Sep and 24 Sep for Pott, H&H, Pairs and SHF. For each stage, the median of each ad's ROAS over the following 14 days as a share of its own baseline:

Read: Fatigued was right when it fired (Pott 707 on 10 Sep) and None ads kept earning. Highly likely ads kept earning too, which is what an early warning should do. Close is mixed and the sample is tiny. One thing worth watching: Monitor ads whose only signal was an unconfirmed best week went on to earn 0.76x of baseline (17 calls), so that signal has some predictive power even when the week looks fine. Only Pott has a breakeven set, so only Pott could reach Fatigued in this test. Full rows in scripts/meta-fatigue-watch/backtests/.

Decisions needed from Casey

1. Breakeven ROAS for H&H, Pairs and SHF Set 20 Oct from Casey: H&H 4.0x (the floor of "4 or 5 minimum"; 5x is the target), Pairs 2.0x (new customer), SHF 3.0x (blended). These are targets, not margin-sheet breakevens; the KC dashboard's plan engine holds a computed break-even ROAS per client but its keys live only in Railway, so it was not cross-checked. With the real bars: H&H 260 on 14 Oct stays at Close (5.15x is above 4.0x), SHF on 19 Oct reads as an account-wide return dip (6 of 6 ads), Pairs on 19 Oct has Ad 465 at Close and 264 at Highly likely.
2. Cron and email. Crontab line is in the header of run-daily.sh; Casey adds it himself. Keep per-client email false, the combined morning email reads state/last-result.json.
3. Publishing. The daily run writes the report locally only. Say if the Pott page should republish each day.

Files

• scripts/meta-fatigue-watch/watch.cjs judging, shared.json thresholds, report.cjs the page, backtest.cjs harness, run-daily.sh cron wrapper.
• clients/<slug>/meta-fatigue-watch/config.json per client; state/ ledger, last result, refresh feed; reports/<date>.html.
• Flags: --dry-run, --as-of=YYYY-MM-DD, --no-report, --json=<path>, --force-email.
• The September agent, for reference: branch worktree-ad-fatigue-agent, dir .claude/worktrees/ad-fatigue-agent/ad-fatigue/.

Later

Point the weekly optimizer's step 5 at this watcher's output instead of keeping a second scoring system. Pott's Christmas ads are the first live read of the merged rules.