TFM · Review hub · 2026-08-26

Advantage+ Exposure Audit

Read-only — no ad-account changes were madeStatus refresh 15:30 ET — queue items C, E, G, I, J closed; H Tier 1 done 44 ad accounts in account_mapping Meta Graph API v21.0 · creative nodes queried directly for degrees_of_freedom_spec

accounts with usable data came back clean — every one has at least one auto-enhancement opted in

Tier-1 accounts to review first (2 are finpubs)

accounts the API can't see — SHARE / dynamic-template creatives, needs an Ads Manager check

Sora 2 dependencies — Sep 24 shutdown needs no remediation

What this is. Meta's Advantage+ creative enhancements (standard_enhancements, image touch-ups, GenAI recomposition, text optimization, auto-crop, carousel-to-video) silently alter the rendered ad. "High-risk" below counts only the features that change text, images, or product depiction — the class behind the Business Insider complaints. Moderate features (translation, CTA tweaks, inline comments) are counted in the total but not the ranking.

Ranked exposure

Bar = high-risk OPT_IN instances across sampled active creatives, relative to the worst account. Up to 25 creatives sampled per account.

#	Account		High-risk opt-ins	Total	Coverage	Top features
1	street-sheet	FINPUB TIER 1	157	290	25/25	standard_enhancements, brightness/contrast, cv_transformation, image_animation, video_auto_crop
2	shriver-media	TIER 1	98	193	14/14	image_enhancement, pac_relaxation (GenAI), standard_enhancements, auto-crop/uncrop — contradicts the registry's "all-ANALYZE by design"
3	written-word	TIER 1	92	193	25/25	standard_enhancements, brightness/contrast, pac_relaxation, image_templates
4	the-intelligence	TIER 1	84	162	16/23	standard_enhancements, brightness/contrast, text_optimizations, cv_transformation
5	student-loan-planner	FINPUB TIER 1	72	97	22/25	text_optimizations + cv_transformation on loan-terms content — compliance-relevant
6	smbdealhunter	TIER 2	62	143	25/25	standard_enhancements, image_touchups, carousel_to_video
7	alphasignal	TIER 2	54	74	9/15	every high-risk feature on all 9 sampled
8	arnolds-pump-club	TIER 2	51	122	25/25	standard_enhancements, brightness/contrast
9	the-flyover	TIER 2	47	90	20/25	standard_enhancements
10	stocks-and-income	FINPUB TIER 2	25	70	15/22	standard_enhancements, pac_relaxation
11	stocks-news	FINPUB TIER 2	21	59	25/25	standard_enhancements, pac_relaxation
12	big-desk-energy	TIER 2	20	43	8/9	standard_enhancements, image_uncrop
13	insight-links	TIER 2	16	36	5/25	low coverage — most creatives untyped
14	daily-drop	TIER 2	15	63	25/25	brightness/contrast
15	houck	TIER 2	12	48	6/6	video_auto_crop, video_uncrop
15	creator-spotlight	TIER 2	12	39	9/15	brightness/contrast, standard_enhancements
17	unicorner	TIER 2	11	45	6/6	standard_enhancements
17	marketbeat	FINPUB TIER 2	11	11	11/25	standard_enhancements only — but on 11/11 sampled
19	bytebytego	TIER 2	4	15	5/25	low coverage
19	stockearnings	FINPUB TIER 2	2	2	6/6	standard_enhancements
—	contrarian-thinking	NO HIGH-RISK	0	21	25/25	only video_filtering
—	a-media-operator, workweek-tmm/-ftt/-gtm/-hospitalogy	NO HIGH-RISK	0	0–3	low	only enhance_cta / inline_comment, or too few typed creatives to conclude

Offboarded accounts still carrying live ads in the DB (just-womens-sports 5, status-news 2, rnt-fitness 0 high-risk) are omitted from the ranking — nothing to act on.

What the API could not see

Needs an Ads Manager check — 6 accounts

Active ads returned, but degrees_of_freedom_spec is absent on every sampled creative. These use SHARE / dynamic-template creative objects where the field isn't populated. The only verification path is the UI: Ads Manager → open the ad → "Advantage+ creative" panel.

capital-news FINPUB
high-yield-harry FINPUB
secretcfo FINPUB
experiential-hospitality
workweek-fttb
workweek-ihih (partial — 3/17 typed; looked clean but inconclusive)

The five 403s — explained, not a token problem

Resolved same day from your roster ground truth. None of these need a BM access fix.

how-to-ai — client removed TFM's Meta access; toggles are theirs to manage now
raging-bull — Meta media runs on a Whop-owned account (CLH_0152 - Whop Ads); any RB audit goes through Whop-side access
the-points-guy — churned; the October return is off
vendry — churned (offboarded 2026-04-03)
quartz — paused, no work in months

Decisions that are yours

Shriver Media discrepancy. The registry says its account is "all-ANALYZE by design"; the API shows the second-heaviest Advantage+ opt-in in the portfolio, including GenAI recomposition. Either the note is stale or the toggle drifted after it was set. Worth a direct look before the Tier-1 pass — it's also one of two clients carrying an undocumented compliance question from the config sweep.
Turn-off order. Recommendation is Tier 1 in rank order — street-sheet, shriver-media, written-word, the-intelligence, student-loan-planner — then Tier 2 opportunistically. Two of the five are finpubs, where text_optimizations on loan/market copy is a compliance exposure, not just a brand one. Nothing has been toggled.
Finpub list is a hypothesis. No vertical field exists in the registry or DB; the 13-account finpub set was inferred from newsletter content. Confirm before it drives anything.

Also delivered today

All in the vault; paths are relative to the-feed-media/. Local commits only — nothing pushed to origin.

Advantage+ / Sora audit — full report + raw JSON
reports/sora-advantage-audit-2026-08-26.md · reports/sora-advantage-audit-2026-08-26-advplus-raw.json
SOURCE
Client-config sweep — 21 → 39 clients eligible for self-serve creative, 0 blocked
Harvest-first, no invented compliance language. 10 TODO-GM fields across 7 clients (alts-co, filterbuy ×2, half-baked, ims ×2, interactive-offers ×2, secretcfo, shriver-media) + 2 standalone compliance flags (secretcfo, shriver-media). Independently re-verified.
reports/client-config-sweep-2026-08-26.md
VERIFIED
Vision tagger bake-off, n=30 — gemma4:26b confirmed as default
Gemma wins background (83% vs 63%), key visual (87% vs 77%), layout; ~24% faster. Real finding: both models miss ui_chrome on iPhone/iMessage mockups (gemma 40%, qwen 0%) — fix is a folder-path rule, not prompting. Note: gemma was already the shipped default, so this was retroactive verification.
research/ml/5090-vision-bakeoff-n30-2026-08-26.md
COMMITTED
Roster ground truth recorded everywhere
TPG → churned; how-to-ai access revoked; quartz dormant; RB on Whop; Vendry reconfirmed. Registry rows + counts + notes, TPG frontmatter, system/CLAUDE.md, audit addendum, memory.
system/client-registry.md · clients/the-points-guy/client-config.md · system/CLAUDE.md
DONE
Two roster bugs fixed on the way
The creative gate treated only "offboarded" as terminal, so TPG and FinMC were offerable for new ads; and every roster count used exact-match on status, silently dropping any annotated cell like churned (…) or active (Test Phase). Both now prefix-match. l2_fill suite 20/20.
system/lib/l2_fill/core.py · system/lib/vault_io.py
TESTED
Per-GM ask drafts for the 10 TODO-GM fields
Ready-to-paste Slack messages: Brady (alts-co, interactive-offers), Aubree (secretcfo + shriver-media compliance reviews). filterbuy, half-baked, ims have no assigned GM in the registry — their 5 asks are routed to you in a separate section (Rabii can likely answer half-baked). Nothing sent.
reports/todo-gm-asks-2026-08-26.md
DRAFTED
Unslop visual analysis + first per-client profile (marketbeat)
Headline: 81% of generated "ads" were strategy decks about an ad, not the ad itself — the wrong artifact shape, directly relevant to the Higgsfield pipeline. MarketBeat advisory profile built; 5 items dropped for conflicting with approved_language. No skill wiring (gate-controlled).
research/tooling/unslop-visual-2026-08.md · clients/marketbeat/unslop-profile.md
COMMITTED
Autoresearch loop closed — 9 wins adjudicated, nothing shipped
Repro passed on 5 (all SHELVED/DEFER — arch swaps for marginal gains or no production path), 3 STALE, 1 already shipped. Two incidents caught and fixed: best.json clobbered by --dry-run (restored byte-identical) and conflict markers baked into 5 results.tsv by a merge collision (repaired, 0 remain).
tmp/ship-and-evals.md · tech-debt entries + HANDOFF.md
VERIFIED
Gauntlet installs (separate session, logged)
fatigue-scan v2, tech-debt v2, vault-integrity v4, weekly-enrichment v2 installed live. Rejected on gate fails: portfolio-pulse v2, scan-triage v2, creative-qa v2, deliverable-qa v2. creative-qa reverted to its last gate-passing state. Watchdog sealed run still in flight.
~/.claude/state/shumer-gauntlet/log.html
INSTALLED

Next to tackle

A
Verify Shriver, then switch off Advantage+ on the five Tier-1 accounts
Still open — your call. street-sheet, shriver-media, written-word, the-intelligence, student-loan-planner. Re-run the audit after to confirm OPT_OUT.
B
UI check on the six SHARE-creative accounts
Still open — needs Ads Manager access. capital-news, high-yield-harry, secretcfo first (finpubs).
C
✅ TODO-GM asks drafted — your move: forward them
Slack drafts ready in reports/todo-gm-asks-2026-08-26.md. Brady ×3, Aubree ×2, and 5 routed to you (filterbuy / half-baked / ims have no assigned GM).
D
Gauntlet: 4 installed, 4 rejected, watchdog in flight
Installed: fatigue-scan v2, tech-debt v2, vault-integrity v4, weekly-enrichment v2. Rejected: portfolio-pulse, scan-triage, creative-qa, deliverable-qa v2s. Watchdog sealed run finishing.
E
✅ Step 3 closed — creative-qa reverted to gate-passing state
Both Layer-4 wirings failed their gates (60.71 and 63.14 vs 64.14 bar). comparative_bar.py stays built and directly runnable; docs parked in reports/.
F
iMessage ui_chrome folder rule in the tagger
Still open — small. */imessage-ad/* → force background_type: ui_chrome in vlm_tagger.py.
G
✅ Vault pushed — level with origin
Merged + pushed 15:14 ET; later commits ride the sync daemon.
H
Git slim: Tier 1 done (gc, 3.5G → 2.6G). Tier 2 armed, awaiting "force-push go"
Tier 2 rewrites history (−24% remote packs) and breaks every teammate clone — needs a Slack heads-up + re-clones, so it fires only on your explicit word.
I
✅ Unslop follow-ups done
Visual analysis run (81% deck-not-ad finding) + marketbeat pilot profile committed. Next optional step: wire profiles as advisory into concept generation.
J
✅ ship-and-evals resumed and finished
9 autoresearch wins adjudicated, nothing shipped (all shelved/defer/stale/already-live), 2 data incidents repaired.