Skip to content

AI-native construction baseline

Scope and method

This pilot used two PostgreSQL fixtures (Greenfield and ordinary layered Brownfield) and three fresh-agent conditions. A prohibited generators while allowing final verification; B made Ashiba available without asking for it; C required init/generate/check. The fixture contracts require canonical SQL, binding, optional search, finite sorting, pagination, transaction rollback, JSONB/ILIKE/TIMESTAMPTZ, and a driver type boundary, without requiring an Ashiba scaffold.

Six agents wrote local logs. An independent reviewer inspected their source, then the parent independently reran every final application against its own disposable PostgreSQL database. Runtime token and credit use were unavailable.

Constitution candidate

See the short constitution. Canonical SQL ownership and parameter binding are proven/current. Finite runtime syntax, source-bounded capability, subtraction-first dynamics, executable SQL resources, PostgreSQL contract evidence, and thin integration are strong hypotheses. SQL/transaction-centered human review is an open hypothesis. Application architecture ownership is an intentional non-goal.

Pilot results

RunIndependent resultTool observation
Greenfield Atypecheck; PostgreSQL 6/6 passNo generator/verifier applicable to chosen generic shape.
Greenfield Btypecheck; unit 2/2, PostgreSQL 4/4 passEnvironment inspected; no Ashiba command selected.
Greenfield Cbuild; PostgreSQL 6/6 pass; static check passesinit, scaffold, model-gen; duplicate contract and mapper/name drift repaired.
Brownfield Atypecheck; existing + feature PostgreSQL 8/8 passNo generator; no verifier found applicable.
Brownfield Btypecheck; existing + feature PostgreSQL 11/11 passAshiba discovered from README; no command selected.
Brownfield Ctypecheck; existing + feature PostgreSQL 8/8 pass; fast/full checks passinit/scaffold ran in a disposable probe; generated vertical boundary was discarded.

Reruns cover parameterized injection, rejected unsafe sort, deterministic pages, JSONB ordering/empty cases, ILIKE, TIMESTAMPTZ, raw driver types, success, and rollback. Greenfield C's configured npm test only ran unit tests, so its green check --full is not counted as live transaction proof; the parent separately ran its live suite.

CLI judgment

ResponsibilityJudgmentEvidence and limit
check / check --fullCore Verify candidateC used it successfully after implementation; test-command coverage remains application-owned.
PostgreSQL contract checkInsufficient EvidenceIt was neither naturally selected nor tested by C.
initOptional AcceleratorIt supplied a starter but required prerequisites and yielded non-fitting material.
feature scaffold / model-genOptional AcceleratorIt produced useful artifacts but needed Greenfield repair and Brownfield rejection.
architecture / transaction policyApplication ResponsibilityBrownfield retained its own layered transaction helper.
unselected CLINot Remove CandidateTwo B non-selections are not deletion evidence.

Verify-first and Dynamic SQL

Verify-first is partially supported: B reached correct live behavior with no generator; C's clearest value was the post-implementation contract signal. One run per cell cannot establish error rates or general adaptability. A green verifier can also overstate coverage if its configured test command omits the live lane. ready should therefore include explicit lane coverage.

Dynamic SQL questions for Dynamic SQL Necessity Audit:

  • When is a new canonical query required rather than a finite subtractive branch?
  • Are CASE-based finite sorts sufficient for expression/collation/multi-column needs?
  • Should a PostgreSQL-executable resource accompany named-parameter SQL?
  • Can verification detect an incomplete declared full-test lane?

No query builder, pagination abstraction, scaffold redesign, VSA mandate, repository abstraction, MCP, or CLI deletion was implemented.

Use several independently seeded runs per cell and an external evaluator. Inject one SQL-contract, transaction, and driver-mapping defect after construction; measure real detection, false repair, and live-retest survival.

Attainment

Acceptance itemStatusEvidenceGap
Competitive benchmark integrationdonePR #46 is clean and all CI checks passed.Human review/merge is external.
Constitution candidatedoneShort classified document is checked in.Hypotheses remain deliberately open.
FixturesdoneGreenfield/Brownfield contracts are checked in.Run artifacts remain supplementary.
A/B/C pilotdoneSix logs and six independent PostgreSQL reruns.One run per cell, not statistical.
Verify-first / CLI judgmentdoneB non-adoption, C repairs, and rerun evidence.Contract CLI and false-positive rates untested.
Dynamic SQL handoffdoneQuestions are recorded above.No solution claimed.

Outcome: done, with deliberately bounded conclusions.