Content as Black Box — one message in, published media out

[xg-gh-25]

---

一句话到可发布内容，不需要人修间距、对齐、字体。

Two Black Boxes

Coding as Black Box: one requirement → verified code delivery.
Content as Black Box: one message → published media package.

Same pattern. Same philosophy. Different domain.

How Content as Black Box works

INPUT:  "没有记忆就没有理解"
OUTPUT: Poster (1080x1080) + narrative text + platform variants
MIDDLE: generate → verify gates → fix → re-verify → publish

The pipeline (called Pollinate):

1. EVALUATE — Is this worth producing? (ROI scoring: knowledge differentiation × audience match × timeliness)
2. THINK — What format serves this message best? (poster? video? narrative? all three?)
3. BUILD — Generate the content with design system constraints
4. VERIFY — Check against quality gates:
   • Typography: hierarchy, spacing, readability scores
   • Color: contrast ratios, palette coherence
   • Layout: visual balance, safe zones respected
   • Content: GEO signals (citations, structure, authority markers)
5. FIX — If any gate fails, regenerate with specific corrections
6. CONVERGE — Only publish when ALL gates pass simultaneously

Why "simultaneously" matters

If you fix typography and break color, then fix color and break layout — you're playing whack-a-mole. The convergence gate requires ALL dimensions to pass in the same render. This eliminates the most common AI content failure: "looks good on one axis, broken on another."

What makes content harder than code

Code has a natural "test suite" — it either compiles and passes or doesn't. Content has no equivalent. A poster can be technically correct (right dimensions, readable text, valid colors) and still be bad — wrong tone, wrong hierarchy of attention, wrong emotional register.

Pollinate solves this by decomposing taste into measurable proxies:

Human judgment	Machine proxy	Gate
"Looks clean"	Whitespace ratio ≥ 30%, max 3 font sizes	Typography
"Colors work"	WCAG AA contrast, ≤ 4 palette colors, no adjacent warm-warm	Color
"Feels balanced"	Visual weight centroid within 15% of geometric center	Layout
"Sounds authoritative"	Citation density, structured claims, no hedging language	GEO

None of these individually capture "good content." But ALL passing simultaneously is a strong proxy — stronger than any single human reviewer's gut check, because humans satisfice on 1-2 dimensions and miss the rest.

This is the unique insight: you can't test content the way you test code, but you can decompose quality into orthogonal dimensions and gate on convergence.

The human role shifts

With both black boxes running:

• Human provides: judgment (what to build, what to say, what matters)
• AI handles: execution (how to build, how to format, how to verify)
• Neither touches: the other's domain

This maps to Karpathy's insight: "You can outsource thinking, not understanding." The black boxes outsource thinking. Understanding — knowing WHAT is worth building and WHO should hear WHAT message — remains irreducibly human.

Every poster in this gallery was produced this way

Zero manual design. Zero Photoshop. Zero template tweaking.

One prompt → Pollinate pipeline → quality convergence → publish.

The prompt for this series was literally: "6 thesis-driven posts about SwarmAI philosophy."

Questions

• Where does "content as black box" break? (Humor? Cultural nuance? Brand voice?)
• Is verification harder for content than code? (Code has tests; what's the "test suite" for a poster?)
• Should the human even SEE intermediate outputs, or does that introduce bias toward "good enough"?

---

Gallery: xg-gh-25.github.io/swarm-content — every piece here was produced by Pollinate.