Case Study: Biophilic vs. Brutalist Spaces Poster
This worked example shows how the verified-infographic workflow would have prevented the 10 fact errors found in a real one-shot-generated poster.
The Original One-Shot Poster
A side-by-side infographic titled "THE EVIDENCE: PEOPLE ARE HAPPIER & HEALTHIER IN BIOPHILIC SPACES" was generated from a single text-to-image prompt. It contained the following claims:
Biophilic side: +15% happiness, -23% stress cortisol, +12% cognitive performance, +26% health/recovery, +8–15% productivity.
Brutalist side: -10% happiness, +15% stress, -9% cognition, -20% health, -7–12% productivity.
Cited sources: Browning et al. (2014); Human Spaces / Terrapin (2015–2023); University of Exeter; Singapore NUS; Finnish studies.
Fact-Check Results
80% of numeric claims were unsupported. Specifically:
| # | Claim | Problem |
|---|---|---|
| 1 | -23% cortisol | No primary source; round-trip marketing stat |
| 2 | +12% cognitive | Misattributed — 12% is a greenery dose, not a cognitive uplift |
| 3 | +26% recovery | No study produces this number; Ulrich 1984 showed ~8.5% |
| 4 | -10% happiness (brutalist) | Fabricated mirror-statistic |
| 5 | +15% stress (brutalist) | Fabricated mirror-statistic |
| 6 | -9% cognition (brutalist) | Fabricated mirror-statistic |
| 7 | -20% health (brutalist) | Fabricated mirror-statistic |
| 8 | -7–12% productivity (brutalist) | Fabricated mirror-statistic |
| 9 | "Finnish studies" | Misattributed — research is Japanese |
| 10 | Browning 2014 as empirical source | Category error — it's a meta-synthesis |
Only +15% well-being and +8–15% productivity were supportable.
How the Skill Would Have Prevented Each Error
Phase 1 (Claim Planning) catches symmetry bias
The "versus brutalist" structure triggers the symmetry_warning flag. Claude would tell the user up front: "The brutalist side may end up with almost no verified data. We will not invent mirror-statistics. Are you OK with an asymmetric layout, or would you like to reshape this as a single-subject biophilic-benefits poster?"
This one question alone would have prevented errors #4–#8.
Phase 2 (Source Discovery) catches round-trip marketing stats
For the "-23% cortisol" claim, Claude runs multiple searches and finds:
- Antonelli et al. (2019) meta-analysis: significant cortisol reduction but no clean -23%
- "Human Spaces" marketing pages cite the number with no primary source
- No peer-reviewed paper produces this exact figure for indoor biophilic design
Claim is classified as QUALITATIVE-ONLY — descriptor "Significant" is used instead of a fake percentage. Error #1 prevented.
Same process catches error #2: Claude notices that "12%" in the literature refers to an optimal greenery coverage ratio (Lei et al. 2021), not a cognitive uplift. The actual effect is 14% short-term memory improvement (Yin et al. 2019) — the number gets corrected, with a proper citation.
Phase 3 (Verification) catches the Browning meta-source error
When Claude attempts to verify a specific percentage against the Browning 2014 paper, it finds the paper is a synthesis of 500+ other studies, not an originator of the percentage. The citation is demoted — Browning can appear in a general reading list but cannot be cited as the source of a specific number. Error #10 prevented.
Phase 3 (Verification) catches geographic misattribution
Claude searches for "Finnish biophilic cortisol studies" and finds the cortisol-reduction literature is overwhelmingly Japanese (Shinrin-yoku, Miyazaki, Park, Li). The citation is corrected. Error #9 prevented.
Phase 4 (User Checkpoint) catches anything the automated phases missed
The Verification Report would show the user:
- Biophilic side: 3 VERIFIED / DIRECTIONAL claims, 2 QUALITATIVE
- Brutalist side: 1 VERIFIED claim (sick days), 1 QUALITATIVE (attention restoration), 3 REJECTED
The rejection rate on the brutalist side is 60%, far above the 20% abort threshold. Claude halts and proposes reshaping the poster.
The user has three good options:
- Drop the comparison, make it a single-subject poster.
- Keep the comparison but render asymmetric (3 vs. 2 rows).
- Substitute a different comparison subject that actually has data (e.g., "windowless offices" instead of "brutalist spaces").
All three produce honest posters. None require fabrication.
Phase 8 (Post-Render Audit) catches image-model drift
Even with a locked prompt, image models sometimes render "+15%" as "+51%" or misspell author names. The Phase 8 audit reads the rendered image multimodally and flags any mismatch between pixels and spec. Up to 3 regeneration attempts are allowed; beyond that, the user is alerted.
Final Poster Content (What Gets Published)
Based on the verified evidence, the final poster would contain:
Title: "Biophilic Spaces Improve Well-Being at Work"
Verified claims (left column, 5 rows):
- Well-being: +15% (Human Spaces, 2015)
- Stress reduction: "Significant" (Antonelli et al., 2019 meta-analysis)
- Short-term memory: +14% (Yin et al., 2019)
- Surgical recovery: -8.5% days (Ulrich, 1984)
- Creativity & productivity: +6% to +15% (Human Spaces, 2015; Exeter, 2014)
Conventional-space claims (right column, 2 rows — asymmetric):
- Sick leave: +18% days with limited daylight (UK GBC, 2016)
- Attention restoration: "Reduced" (Kaplan, 1995)
Sidecar source file sources.md contains full citations with URLs
and quoted passages, so any reader can verify every number.
Error Count: Before vs. After
| Approach | Numeric claims | Fabricated / misattributed | Error rate |
|---|---|---|---|
| One-shot text-to-image | 10 | 8 | 80% |
| verified-infographic-generator | 7 | 0 | 0% |
The verified-infographic skill produces a slightly smaller poster (fewer rows, because empty rows are not invented to fill space) but every claim is defensible, every citation is real, and the poster has a reader-facing sidecar so any fact can be checked in seconds.