14
Blueprint

Taste Evaluator Blueprint

A build-ready blueprint for a generative-to-evaluator pipeline that scales human taste at machine speed — a taste corpus, an on/off-brand filter, the 1,000→100→10→1 funnel, and a human-in-the-loop loop that prevents the Flattening Effect.

ShareXLinkedInFacebookEmail
Curation funnel
Generated
1000
Shortlist
100
Finalists
10
Shipped
1

What it does

When execution commoditizes, judgment becomes the scarce differentiator. Chapter 7 makes the case through Midjourney: competitors have the same diffusion architectures, much of it open source, yet none has reproduced Midjourney's output, because "the moat is taste — and taste is not something any open-source license can distribute." Chapter 9 names the person who holds that moat — the Force Multiplier — and the mechanism that scales her judgment: the evaluator agent. Figure 9.4 is the architecture in one line: a generative swarm produces 1,000 variations, a curated library narrows them, the Force Multiplier's filter keeps the top 100, and the best is deployed — "scaling human judgment at machine speed."

This blueprint is the build spec for that pipeline. It is not a scoring tool; it is an engineering brief for a system that operationalizes Taste as a Moat. The core insight from Chapter 9 is what the pipeline must respect: agents "detect violations of existing aesthetic rules with ruthless precision. Only humans sense when the rules themselves need revision." So the system generates and filters at machine scale, but it keeps a human in the loop at the exact point where the rules might need to change — and it actively defends against the Flattening Effect, the documented drift where a Force Multiplier's rejection rate falls from 50% in week one to under 10% by month six, not because she got better but because her taste converged with the agents'.

The pipeline has four stages — taste corpus, on/off-brand filter, the 1,000→100→10→1 funnel, and the human-in-the-loop calibration loop — and Sofia Marchetti's "staleness detector" from Chapter 9 is built in: a penalty for recommendations too similar to recent output, measured across silhouette, layering, palette, and accessory dimensions over rolling windows.

Who it's for: Force Multipliers and the Architects who build alongside them, in any AI-Born firm where output is commoditized and brand-distinctive taste is the durable advantage — fashion, design, content, product.

Figure: The four-stage evaluator architecture this blueprint specifies — generative swarm to curated library to Force Multiplier filter to deployed best.

Where output is commoditized and brand-distinctive taste is the durable advantage.
Stage 1 — Taste corpus

The Force Multiplier’s accumulated judgment, versioned. Borderline stays a distinct label.

Corpus total: 240.
The aesthetic axes Sofia’s staleness detector measures across.
Stage 3 — Funnel ratios
Default 1,000 → 100 → 10 → 1 (Figure 9.4). The human narrowing cannot be skipped.
Exploration randomness30
Sofia's lever — temperature 0–1. Raise it to force the swarm out of a local maximum.
Anti-flattening guardrails
Default 2-week window; 12% floor (Chapter 9). Below the floor → recalibration gate.
Funnel simulation
Recent-output similarity (rolling window)35
Drives the staleness penalty. Push it high to watch the funnel converge on a local maximum.
Same seed + corpus version reproduces identical evaluator scores — audits replay exactly.
100010010 → 1 · funnel run · corpus ve18a4f
Deploy held

Rejection rate below the flattening floor — recalibration required before the next deploy.

Drift dashboard
Rejection rate (floor 12%)0%
Mean batch staleness29
Shortlist escalated to human (0/10)0%
Flattening alert · deploy blocked

Rejection rate has fallen below 12%. This is not skill improvement — it’s taste convergence with the agents. Rotate through non-AI environments and recalibrate before the distinctiveness erodes invisibly. The next deploy is gated until recalibration.

Evaluator scorecard · shortlisted 10 (top 10)
#On-brandStalenessNetRoute
961853154auto
106823052auto
610823052auto
580813051auto
119813051auto
534803051auto
23803051auto
995803051auto
886803051auto
35792950auto

The human-in-the-loop narrowing at 10010→1 is mandatory — the evaluator detects rule violations; only the human senses when the rules themselves must change. Identical seed and corpus version reproduce this table exactly.

Operationalizes the Taste as a Moat framework.
Further reading
From the books
  • Book 1, Chapter 9 — "The Force Multiplier: Taste-Transmitter and Player-Coach" (the Taste-as-Filter evaluator architecture, Sofia Marchetti's staleness detector, and the Flattening Effect mitigations).
  • Book 1, Chapter 7 — "The Invisible Moat" (Taste as a Moat; the Midjourney example).
The Dispatch — N°01

Essays from
the lineage break.

New essays, framework studies, excerpts and pre‑order news. Sent rarely. Never noise.