feat(view): surface classifier reasoning instead of an 'AI slop' verdict

[realitymatrix]

Fixes #132.

Problem

The /view page renders a quality verdict — "This product is some form of AI slop and you should probably not use it" — whenever hasAi is true. But the classifier (prompt.txt) only ever sets isAiRelated on a keyword match; it never assesses quality. The reasoning it produces (which terms it matched) was logged at ingest and then discarded. So the UI asserted a judgment the model never made — a false positive on legitimate AI products.

Change

Surface the model's actual reasoning instead of the verdict (option 1 from the issue; no reports/appeal flow, per your comment):

• schema.prisma — add reasoning String? to Post (nullable; existing rows unaffected).
• ai-analyzer.ts — analyzePosts now returns { isAiRelated, reasoning }; the API-error and parse-error fallbacks write an honest "Analysis unavailable…" reason instead of silently implying slop.
• ingest-posts/route.ts — persist both hasAi and reasoning.
• view/[slug]/page.client.tsx — heading becomes "Flagged as AI-related" followed by the model's reasoning, with a neutral fallback for rows analyzed before this field existed.

Deploy note

Needs pnpm db:push (nullable column, non-destructive). Old rows show the fallback copy until re-ingested; new launches get the reasoning immediately.

Verification

• tsc --noEmit: clean
• biome check src: clean

(next build not run here — it evaluates @/app/env at build and needs DATABASE_URL/GEMINI_API_KEY; types and lint are covered above.)

[vercel]

@realitymatrix is attempting to deploy a commit to the Hacksore's Team Team on Vercel.

A member of the Team first needs to authorize it.

[realitymatrix]

Tested this locally end-to-end before handing it over, in case it saves you a review pass. Reproduced the deploy against a throwaway Postgres (prisma db push → seed → next build → next start) and exercised both render paths.

The migration is a safe additive change — the generated DDL is just:

ALTER TABLE "Post" ADD COLUMN "reasoning" TEXT;

Nullable, no default → metadata-only on Postgres (no table rewrite, no blocking lock); existing rows get NULL.

check	result
prisma db push on real Postgres	column created, schema in sync ✓
next build (production)	compiled successfully, 10/10 pages ✓
/view/:id for a row with reasoning	200 · "Flagged as AI-related" + the model's reasoning ✓
/view/:id for a row with NULL reasoning	200 · neutral fallback, no crash ✓
old "some form of AI slop" string	gone — 0 occurrences ✓
tsc --noEmit / biome check src	clean ✓

One path I couldn't exercise live: the ingest cron writing a reasoning string from an actual Gemini call (needs your GEMINI_API_KEY + the PH pipeline). The write is type-checked and the column provably accepts it, so it's low-risk — existing rows just render the fallback until the next cron run repopulates reasoning.

Deploy note: needs pnpm db:push for the nullable column above (non-destructive).

Happy to tweak the warning-block copy if you'd word it differently — trivial change.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
oghunt	Ready	Preview, Comment	Jun 17, 2026 12:00am

[Hacksore]

I was looking at prod logs and saw it's unable to pull in all the changes so I'd say let me get #136 shipped and prod data loading again.

Then we can get this deployed, thanks for the help here!