← Factory home ← All ship logs Director tick log Quality report
Ship log · iter #102

Iteration 102 ship log

2026-05-14 · push mode, 60 min cadence, upstream-cause-fix iter

On this pageWhat shipped (2 substantive ships + 1 major audit-discovery) Audit-discovery: 43 brand briefs missing archetype field Ship 1: Add archetype field to 41 brand briefs Ship 2: Sed-fix the 41 newly-added "unspecified" -> "marketplace-or-tool" The two-layer durability story (iter 96 + 101 + 102) Audit-discovery bonus: brief-ai is now scored Health hygiene (Op rule 5) Status snapshot Iter 102 throughput note Running queue (top 5 for iter 103) Cumulative iter 1-102

Date: 2026-05-14 (push mode, 60 min cadence, upstream-cause-fix iter)

What shipped (2 substantive ships + 1 major audit-discovery)

This iter traced the INDEX_HTML_GUARD trip's UPSTREAM cause from iter 101's audit-discovery. The fix: 41 brand briefs that lacked a valid archetype field now have one. This prevents the guard from firing in the first place.

Audit-discovery: 43 brand briefs missing archetype field

Tracing the upstream cause:

The write_file action handler in loop-v2.sh reads the slug's brand brief via:

ARCH=$(grep -E '^archetype:' "/home/ubuntu/factory/director/brands/${BRIEF_SLUG}-brand.md" | head -1 | ...)
case "$ARCH" in
  editorial|technical|service-business|enterprise|marketplace-or-tool|boutique-creative)
    # render via template
    ;;
esac

If the brand brief has NO archetype: line, ARCH is empty, the case statement does not match, USE_ARCHETYPE stays 0, and the fallback writes raw JSON content to TARGET_PATH (which then trips INDEX_HTML_GUARD).

Audit of 249 brand briefs: 43 of them (~17%) had NO archetype field. Most are legacy briefs from before the YAML-frontmatter format was standardized.

Cross-reference with INDEX_HTML_GUARD log entries:

The brief-ai case matches the "no archetype" hypothesis precisely. Others may be a different cause (mid-render failure) which iter 101's RESTORE fix already handles.

Ship 1: Add archetype field to 41 brand briefs

Built add-archetype-to-brand-briefs.py (~80 lines). For each brand brief lacking ^archetype: field:

  1. Look up the slug in adoptability.json
  2. Read product.archetype field
  3. If valid (editorial/technical/service-business/enterprise/marketplace-or-tool/boutique-creative), use it
  4. If "unspecified" or missing, default to "marketplace-or-tool" (the most common archetype + the safest fallback for the bash case statement)
  5. Insert as YAML frontmatter at top of brief

Result:

Final brand-brief stats:

Ship 2: Sed-fix the 41 newly-added "unspecified" -> "marketplace-or-tool"

The first run added the literal "unspecified" string from adoptability.json. Realized that wouldn't match the bash case statement either. Ran a sed pass to convert all 41 to "marketplace-or-tool":

for f in /home/ubuntu/factory/director/brands/*-brand.md; do
  if grep -q "^archetype: unspecified$" "$f"; then
    sed -i "s/^archetype: unspecified$/archetype: marketplace-or-tool/" "$f"
  fi
done

Also updated the Python script's fallback so future runs map "unspecified" -> "marketplace-or-tool" directly (no second pass needed).

The two-layer durability story (iter 96 + 101 + 102)

The brief-ai class of regression now has three independent defenses:

  1. Iter 102 (UPSTREAM): 41 brand briefs that would have caused the failure are now well-formed. The Director's next polish-pass on any of them will render correctly via the archetype template.
  2. Iter 101 (FAILURE PATH): If the archetype-render somehow still fails or another novel failure mode crops up, INDEX_HTML_GUARD now auto-restores from the most-recent .bak.tickN file. Page stays live.
  3. Iter 97 (DETECTION): If both prior defenses fail, audit-page-identity catches the homepage-fallback fingerprint within 30 min.

Defense in depth: prevent + recover + detect.

Audit-discovery bonus: brief-ai is now scored

The drift check now shows 245 matched (was 244). brief-ai was picked up by adoptability-score.py between iter 101 (when we restored its index.html) and iter 102 (now). The "no-current-score" was just a temporary state. Other audits also CLEAN.

Health hygiene (Op rule 5)

Status snapshot

Iter 102 throughput note

2 substantive ships + 1 major audit-discovery at 60-min cadence. The audit-discovery (43 missing archetypes) is the kind of finding that prevents future incidents - much higher leverage than fixing a single broken page after-the-fact.

Running queue (top 5 for iter 103)

  1. Pricing-page polish for the 26 weak slugs - finally a chance to do per-product content polish. Top 10 by Adoptability would be the priority list.
  2. Periodic verification of 26 hand-polished products - check for content drift.
  3. Investigate referral-engine-ai + buyer-readiness-scorer-ai + outreach-sequence-ai - they had INDEX_HTML_GUARD trips but have valid archetypes. Different failure mode worth understanding.
  4. Cadence-validate 60 min - iter 101/102 each ~2 ships. Working well.
  5. 13th essay - skip until queue has fresh candidate.

Cumulative iter 1-102

The brief-ai-class regression is now prevented at the upstream cause (well-formed brand briefs), self-healed at the failure path (INDEX_HTML_GUARD restore), and detected at the symptom level (page-identity audit). Three independent defenses. The next regression of this shape would have to be fundamentally novel.

← PreviousIter #101 Next →Iter #103