# Wishdeal Factory buyer-path - iteration 65 ship log

**Date:** 2026-05-13 (depth mode, per-product polish + fake-proof audit)

## What shipped

Three polish ships, two of them caught and removed fabricated proof claims from iter 58 bulk-gen output. The fake-proof audit pattern itself is the durable contribution: a re-runnable check for any future LLM-generated content.

## The audit

Per-product polish on top-Adoptability bulk-gen products turned up a different issue than expected. Two of three top candidates had fabricated proof claims that the iter 58 LLM generator slipped past my prompt's "no fake customer counts" constraint:

| Slug | Score | Fabrication | Severity |
|---|---|---|---|
| churn-ai | 72 | "Trained on signals from over 100,000 SaaS accounts" (2 places) | Hard violation |
| contract-negotiation-ai | 71 | "Built on patterns from over 10,000 analyzed contracts" (3 places) | Hard violation |
| lead-scoring-ai | 73 | Lede repeated H1 text verbatim | Polish issue |
| attribution-ai | 72 | Already at-quality | No-op |

Two products had been silently shipping fake-proof claims since iter 58 (10 days). The iter 59 quality audit caught HTML structural bugs but did not check for fake-proof patterns, which is a different failure mode.

## Ship 1: churn-ai fabrication removed (source fix)

The placeholders.json had `Trained on signals from over 100,000 SaaS accounts. Reads usage drops, login gaps, and feature abandonment instead of survey scores nobody fills out.` in the FEATURES field, and `Trained on behavioral data from over 100,000 SaaS accounts.` appended to the HERO_DEK.

**Fix:** Replaced the FEATURES paragraph with honest framing:

> Reads usage drops, login gaps, and feature abandonment from your own product analytics, instead of survey scores nobody fills out. The patterns we look for are the ones documented across published SaaS retention research, applied to your specific account data.

This is honest: we DO look at your own product analytics; the patterns ARE in published retention research (Tomasz Tunguz, ChartMogul, Profitwell datasets are all public). No fake corpus claim.

Removed the HERO_DEK fake-training sentence entirely (the rest of the deck is strong without it).

## Ship 2: contract-negotiation-ai fabrication removed (source fix)

Same pattern, three places:
- `BA_SUBHEAD`: "patterns from 10,000 analyzed contracts" -> "patterns from publicly-documented term sheets and VC standard forms"
- `FEATURES`: "scored against 10,000 analyzed contracts" -> "scored against publicly-documented term sheets at your stage and check size (NVCA forms, YC SAFE templates, public M&A LOIs)"
- `HERO_DEK`: "Built on patterns from over 10,000 analyzed contracts" -> "Built on patterns from publicly-documented term sheets (NVCA model forms, public M&A LOI archives, YC SAFE references)"

The NVCA model forms and YC SAFE templates are real public artifacts that founders actually negotiate against. Naming them is honest, specific, and stronger than the fabricated 10,000-contract claim because it tells the buyer exactly what reference patterns the system uses.

## Ship 3: lead-scoring-ai lede polish

The bulk-gen produced a lede that opened by repeating the H1 verbatim ("Stop wasting half your week..." -> "Your sales reps are wasting half their week..."). The H1+lede pair carried less new information than they should.

**Fix:** rewrote the lede to lead with specific operator pain math:

Before:
> Your sales reps are wasting half their week chasing leads that will never close. Lead Scoring AI trains on your own closed deals and scores every new inbound prospect from 1 to 100, instantly. Reps work the top quartile first and stop guessing which calls matter. Free for 7 days, no contract, with your first batch scored before onboarding ends.

After:
> You hire 4 SDRs and they each pull 8 hours a week working leads that never sign. That is more than a full-time salary, every quarter, on the wrong calls. Lead Scoring AI trains on your own closed deals and scores every new inbound prospect from 1 to 100 in seconds. Reps work the top quartile first and stop guessing which calls matter. Free for 7 days, no contract, with your first batch scored before onboarding ends.

The new lede:
1. Names a concrete operator scenario (4 SDRs, 8 hours/week each)
2. Translates the time to dollars ("more than a full-time salary, every quarter, on the wrong calls")
3. Then explains the mechanism

Specific operator pain math is the move that distinguishes hand-polished operator voice from generic SaaS template lede.

## The audit pattern (durable contribution)

The dangerous-pattern regex set I built to find fake-proof claims:

```python
DANGEROUS = [
    r'over \d{2,}\,?\d{3,}',           # over 10,000+
    r'trained on \d+\,?\d{3,}',         # trained on 10,000+
    r'\d+,000\+? (customers|users|teams|companies|practices|accounts|operators|founders|clinics|firms)',
    r'used by \d+',
    r'powering \d+',
    r'deployed.*\d{3,}.*custom',
    r'\d{2,}\+\s+(production|enterprise|saas)\s+(customers|users|accounts|deployments)',
]
```

This pattern set caught 2 hard violations and 0 false positives on the bulk-gen output. It can be re-run on any future LLM-generated placeholder JSON to catch new fabrications before they ship. Worth wiring into _bulk_gen.py as a post-generation validation step.

## Files changed inventory

### Modified (source-level)
- `/Users/wes/factory-templates/churn-ai-placeholders.json` (2 fabrications removed)
- `/Users/wes/factory-templates/contract-negotiation-ai-placeholders.json` (3 fabrications replaced with honest references)
- `/Users/wes/factory-templates/lead-scoring-ai-placeholders.json` (lede polished)

### Re-rendered
- `/srv/sites/factory/builds/churn-ai/index.html`
- `/srv/sites/factory/builds/contract-negotiation-ai/index.html`
- `/srv/sites/factory/builds/lead-scoring-ai/index.html`

## Status snapshot

- 238 products, 0 broken pages, 0 stale counts
- 7 substantive playbook essays (~13,000 words)
- 5 foundational high-trust pages depth-passed (complete)
- 60 bulk-repaired + 4 hand-repaired + 5 polished (added 3 this iter) + 2 confirmed-good + 1 audit-fix
- **2 fabricated proof claims removed at source this iter (churn-ai, contract-negotiation-ai)**
- Generator source durability: 48 adoptability records + 12 entity-em-dash + 3 count-fix + 13 hand-polished generators all clean at source
- 2257 sitemap URLs
- 68/68 health endpoints passing
- 0 em-dashes shipped this iter

## What still needs Wes

1. Stripe wiring (30 min)
2. Email-send for auto-fulfill
3. First real traffic push
4. Decision on rebrand-name application (carried)
5. Wire the fake-proof audit into _bulk_gen.py (small infrastructure ship, ~15 min)

## Iter 66 candidates

1. **Wire the fake-proof audit into _bulk_gen.py** - durable safety net. Quick infrastructure ship.
2. **Catalog-wide fake-proof sweep** - run the audit across ALL 60 bulk-gen products (not just top 3) to catch any remaining fabrications.
3. **Hand-polish next 2-3 highest-Adoptability products** (campaign-budget-ai 71, supplier-ai 71, lead-router 71).
4. **/factory/builds/audit-ai/ repair** for the screenshot-failing page.

Recommended: option 2 (catalog-wide fake-proof sweep). Iter 65 found 2 violations in 3 audits (66% rate at the top). If that rate holds, there are 30+ more fabrications across the remaining 57 bulk-gen products. Worth one iter to clear systematically.

## Cumulative iter 1-65

- **Catalog**: 238 products, 0 broken, all with real taglines + names, **fake-proof being actively removed**
- **Content library**: 7 operator essays (~13,000 words)
- **Foundational pages**: 5 of 5 operator-depth-passed (COMPLETE)
- **Source durability**: 13+ generators source-fixed
- **Hand-polished pages**: 5 (bookkeeper, nurture, lead-scoring, churn-ai polished iter 65, contract-negotiation-ai also polished iter 65)
- **Infrastructure invariants**: 68 endpoints, 0 em-dashes, durable

The audit pattern from iter 65 is the leverage move: a re-runnable check that catches fake-proof claims before they ship. Next iter should apply it catalog-wide to clean any remaining fabrications.