How quality scoring works

Every generation scored. 100 points, four dimensions.

Archiet runs a 4-dimensional scorer on every generation and refuses to deliver a ZIP that regressed. Marketplace templates need 80+; your own projects use the same scorer so you know what shipped before you open the archive.

Try it — no credit card Sample handoff pack

The four dimensions

Build integrity

Does the code compile and run?

25 points

—Generated files parse correctly in their target language
—No imports reference non-existent modules
—Package manifests (requirements.txt / package.json / composer.json) are valid and installable
—Database migrations form a linear chain; every model field maps to a migration
—No files contain `raise NotImplementedError` as generated output (runtime crash detector)

Security posture

Does the output follow Archiet's hard security rules?

25 points

—No JWTs in localStorage (frontend) or AsyncStorage (mobile) — always httpOnly cookies
—CSRF protection on every state-changing route
—Passwords hashed with bcrypt (work factor 12)
—All routes validate input via schema definitions; no raw SQL for data queries
—No hardcoded secrets in source — all credentials read from environment
—Authentication required on every protected endpoint

Production readiness

Is the app actually deployable to production?

25 points

—Dockerfile + docker-compose.yml ship for the chosen stack
—Alembic migrations applied cleanly from zero state
—Required auth pages present: forgot-password, reset-password, verify-email, settings, onboarding
—Mobile app includes App Store compliance: EAS config, privacy-policy screen, review prompt, OTA update check
—CI pipeline (GitHub Actions) scaffolded with test + lint stages
—README includes run-locally instructions that actually work

PRD fidelity

Does the generated app match the PRD you uploaded?

25 points

—Every extracted functional requirement maps to at least one generated file
—Every extracted user story has a screen that implements its flow
—Every extracted entity exists as a model with the declared fields
—Every extracted state machine is enforced at the DB layer via migration
—Orphan FRs (declared but not implemented) are surfaced explicitly, not hidden

Handoff completeness bonus (+0 to +5)

Separate from the 100-point base score — so existing calibrations stay stable — there is an additional 0–5 bonus for how complete the docs/ tree is. The marketplace and per-user dashboards can surface this independently.

+docs/decisions/ — Architecture Decision Records present (+1)
+docs/traceability.md — story → FR → file chain present (+1)
+docs/diagrams/*.mmd — Mermaid architecture diagrams present (+1)
+docs/cost-estimate.md + security-posture.md + performance-notes.md — all three operational briefs present (+1)
+docs/compliance/ — compliance docs cover every declared flag (+1)

Two gates, one scorer

Delivery gate

Hard block — the ZIP does not ship

—Overall score < 80 (production stacks) or < 30 (preview stacks)
—CRITICAL security issues remain after auto-repair
—Any must-priority FR has zero generated files
—File count < 30 (template rendering collapsed)
—SBT sandbox boot failed (app crashed on startup during synthetic boot test)

Marketplace publish gate

Higher bar for the Archiet marketplace

—Overall score ≥ 80
—No single dimension < 15 (prevents a perfect-in-three, terrible-in-one passing)
—Passes delivery gate
—Non-trivial feature depth (not just CRUD boilerplate)

Why this exists

Generative coding tools have a consistency problem: the model produces something that looks reasonable, you open it, find half the imports broken, give up, delete the ZIP. The quality scorer is Archiet's answer to that failure mode. If a generation would have landed you in that scenario, the delivery gate fails it before you see it.

Over 300 marketplace solutions have calibrated the score today; new template additions slot into that distribution so you know how your output compares.

See your own score

Free trial. No credit card. Generate something real; the score shows up in the codegen output page before you download.

Try it For CTOs

Loading…

The four dimensions

Build integrity

Does the code compile and run?

25 points

—Generated files parse correctly in their target language
—No imports reference non-existent modules
—Package manifests (requirements.txt / package.json / composer.json) are valid and installable
—Database migrations form a linear chain; every model field maps to a migration
—No files contain `raise NotImplementedError` as generated output (runtime crash detector)

Security posture

Does the output follow Archiet's hard security rules?

25 points

—No JWTs in localStorage (frontend) or AsyncStorage (mobile) — always httpOnly cookies
—CSRF protection on every state-changing route
—Passwords hashed with bcrypt (work factor 12)
—All routes validate input via schema definitions; no raw SQL for data queries
—No hardcoded secrets in source — all credentials read from environment
—Authentication required on every protected endpoint

Production readiness

Is the app actually deployable to production?

25 points

—Dockerfile + docker-compose.yml ship for the chosen stack
—Alembic migrations applied cleanly from zero state
—Required auth pages present: forgot-password, reset-password, verify-email, settings, onboarding
—Mobile app includes App Store compliance: EAS config, privacy-policy screen, review prompt, OTA update check
—CI pipeline (GitHub Actions) scaffolded with test + lint stages
—README includes run-locally instructions that actually work

PRD fidelity

Does the generated app match the PRD you uploaded?

25 points

—Every extracted functional requirement maps to at least one generated file
—Every extracted user story has a screen that implements its flow
—Every extracted entity exists as a model with the declared fields
—Every extracted state machine is enforced at the DB layer via migration
—Orphan FRs (declared but not implemented) are surfaced explicitly, not hidden

Two gates, one scorer

Delivery gate

Hard block — the ZIP does not ship

—Overall score < 80 (production stacks) or < 30 (preview stacks)
—CRITICAL security issues remain after auto-repair
—Any must-priority FR has zero generated files
—File count < 30 (template rendering collapsed)
—SBT sandbox boot failed (app crashed on startup during synthetic boot test)

Marketplace publish gate

Higher bar for the Archiet marketplace

—Overall score ≥ 80
—No single dimension < 15 (prevents a perfect-in-three, terrible-in-one passing)
—Passes delivery gate
—Non-trivial feature depth (not just CRUD boilerplate)

Why this exists

Over 300 marketplace solutions have calibrated the score today; new template additions slot into that distribution so you know how your output compares.