We Audited Our Own Social Posts. Every One Failed.

Last week we pointed Digivate's analyst agent at our own published social posts. The verdict: every post recommended for removal.

Here's the actual breakdown.

The setup

The 23-agent pipeline includes an analyst agent (V2, sonnet-class model) that scores published posts against three marketing playbooks: direct response, social content, and persuasion + positioning. Each post gets a composite score (0-100) plus per-playbook scores plus a flag for which of four convergent moves landed: direct-response CTA, social-proof specificity, persuasion lever, positioning clarity.

The audit cron runs nightly at 4am UTC. We pointed it at the last 30 days for the first time on 2026-04-28.

What we audited

Five posts. Two on Twitter, two on Facebook, one Instagram carousel. Combined live engagement: 20 impressions, 0 likes, 0 retweets, 0 clicks across all five posts.

The scores

| Platform | Score | Recommendation |
|---|---|---|
| Twitter post 1 | 38/100 | remove |
| Twitter post 2 | 47/100 | remove |
| Facebook post 1 | 51/100 | remove |
| Facebook post 2 | 57/100 | remove |
| Instagram carousel | 44/100 | remove |

Average: 47.4/100. The auto-publish gate is 75. None of these would have shipped if the gate had been wired up before we started publishing.

Why every post failed

The violations were depressingly consistent.

Banlist hooks fired in 2 of 5 posts. The phrase 'Most SMBs' was on our explicit do-not-use list, written into our brand position memory. The writer agent ignored the banlist anyway because the rule was sitting in passive context instead of being mechanically enforced.

Zero direct-response CTAs. Every post ended on a soft prompt — 'thoughts?', 'swipe through', 'see more'. None named a specific action with a payoff.

Zero proof anchors. None of the five posts referenced the 23-agent pipeline, the quality gate, the audit funnel, the blog, or the Recraft + Supabase image stack. Every post made claims that any competitor could copy-paste verbatim.

Zero convergent moves landed. The four moves we measure — direct-response CTA, social-proof specificity, persuasion lever, positioning clarity — were absent across all five posts.

The fix

We didn't rewrite the posts. We rewrote the producer.

Three changes shipped to the pipeline:

• The banlist moved from passive brand-position context into a hard filter ( enforced both at writer prompt level and at editor scoring level — instant -15 if a banlisted phrase appears).

• The required convergent moves and proof anchors moved from 'consider these' guidance into 'each variation MUST visibly apply at least one' hard-constraint blocks in the writer prompt.

• The editor agent now sees the last 21 days of captions and deducts engagement points if a candidate caption uses the same hook structure as 2+ recent posts. Monoculture detection at the producer side, not just at the analyst side.

The result

We regenerated 10 posts (5 X, 3 FB, 2 IG carousels) through the rewired pipeline.

• Average composite score: 76/100 (up from 47).
• Banlist hook violations: 0/10 (was 2/5).
• Proof anchors present: 10/10 (was 0/5).
• Posts caught by the new monoculture detector and rejected: 5/14 (the system catching itself).

After a revision pass on the rejected ones, 9 of 10 posts cleared the 75-point auto-publish gate.

Why this matters

A pipeline that can't audit itself is theater. We've spent four months claiming to ship 'AI-powered marketing' while producing content that scored 47 against the same playbooks any honest marketer measures with. The audit told us the truth before any client could.

The pipeline is now the product. Not the agents, not the model — the audit + scoring + revision loop that catches the system before it ships.

If you want to see your own content scored against the same playbooks, the audit endpoint is at digivate.org/audit. It uses the exact analyst agent you just read about.