EX — Executable Sprint: Full GS Tier Proof on RealWorld Conduit

Date: April 17, 2026
Duration: Single session (~8 hours)
Benchmark: RealWorld Conduit — TypeScript/Express/Prisma/PostgreSQL
Deployment: Railway (managed PostgreSQL, production Docker image)
Tooling: ForgeCraft MCP

What This Experiment Proves

EX demonstrates all four GS tiers on a publicly recognized benchmark — the RealWorld Conduit API — using the ForgeCraft Executable Sprint toolchain:

Tier Layer Evidence
L1 Specification artifacts PRD, use-cases (UC-001–UC-008), 4 ADRs, C4 diagrams, CLAUDE.md
L2 Behavioral contracts 13 hurl probes, 1,013 assertions — 13/13 passing against Railway
L3 Environment probes 3 bash probes (health+DB, JWT round-trip, env-vars) — 3/3 passing
L4 SLO / load test k6 ramp: 10 VUs sustained × 3 min — p95=350ms, errors=0.04%

All four tiers verified against the live production deployment on Railway — not localhost.

Results at a Glance 

L2 — Behavioral:   13/13 probes PASS  (1,013 assertions, ~24s total)
L3 — Environment:  3/3  probes PASS   (health DB connected, JWT configured, env-vars)
L4 — SLO ramp:     6/6  thresholds GREEN
  p95 overall:  350ms  (limit: 2000ms)
  p99 overall:  720ms  (limit: 4000ms)
  read p95:     181ms  (limit: 1000ms)
  write p95:    401ms  (limit: 2000ms)
  error rate:   0.04%  (limit: 1%)
  sustained:    10 VUs (limit: 10)

Directory Structure 

ex/
├── README.md           ← this file
├── spec/               ← L1 specification artifacts
│   ├── PRD.md
│   ├── use-cases.md
│   ├── CLAUDE.md
│   ├── adrs/           ← ADR-001 through ADR-004
│   └── diagrams/       ← C4 context + container (Mermaid)
├── harness/            ← L2/L3/L4 probe files
│   ├── *.hurl          ← behavioral probes (hurl HTTP assertions)
│   ├── hurl.env        ← target URL config
│   ├── slo-ramp.k6.js  ← L4 k6 ramp scenario
│   ├── env/            ← L3 bash probes
│   └── specs/          ← .forgecraft env/slo YAML contracts
├── evidence/           ← machine-readable run results
│   ├── harness-run.json     ← L2: 13/13 pass, 2026-04-17T20:30:31Z
│   ├── env-probe-run.json   ← L3: 3/3 pass, 2026-04-17T23:01:27Z
│   └── slo-ramp-summary.json ← L4: all thresholds green
└── project/            ← full Conduit source code
    ├── src/            ← TypeScript source (hexagonal architecture)
    ├── prisma/         ← schema + migrations
    ├── tests/          ← Jest unit tests + harness probes
    ├── Dockerfile      ← multi-stage Alpine build
    └── package.json

Reproducing This Experiment

Prerequisites

  • Node 18+, Docker, hurl, k6
  • Railway CLI (npm install -g @railway/cli) + account
  • PostgreSQL (Railway managed or local)

1 — Deploy to Railway 

cd project
railway login
railway init
railway add --name conduit-db --plugin postgresql
railway up --service conduit-api

Set env vars in Railway dashboard:

  • DATABASE_URL — auto-linked from the PostgreSQL plugin
  • JWT_SECRET — any strong secret (e.g. openssl rand -hex 32)
  • NODE_ENV=production

2 — Run L2 behavioral probes 

cd harness
API_URL=https://<your-railway-url> hurl --variables-file hurl.env \
  --variable host=https://<your-railway-url> \
  *.hurl --test

Expected: 13/13 files pass, 1,013 assertions.

3 — Run L3 environment probes 

cd harness/env
API_URL=https://<your-railway-url> bash health.sh
API_URL=https://<your-railway-url> bash jwt.sh
bash env-vars.sh  # skips gracefully outside deployment context

4 — Run L4 SLO ramp 

cd harness
k6 run --env API_URL=https://<your-railway-url> slo-ramp.k6.js

Expected: all thresholds green (p95<2000ms, errors<1%).

Implementation Notes

The Conduit implementation required 15 bug fixes during the session to achieve full compliance with the RealWorld API spec. The key fixes were:

  1. Error contract: each error class owns its toJSON() shape — { errors: { <field>: ["message"] } }
  2. Null semantics: bio/image use value || null not value ?? null (empty string → null)
  3. articlesCount: separate countAll()/countFeed() queries, not page-count
  4. ConflictError: 409 status for duplicate username/email (not 422)
  5. Rate limiting: disabled in non-production; raised to 500/min in production
  6. Prisma on Alpine: binaryTargets = ["native", "linux-musl-openssl-3.0.x"] + apk add openssl
  7. JWT middleware: error message "is missing" not "Authorization header missing"

All fixes are reflected in project/CLAUDE.md (the architectural constitution) and spec/adrs/.

ForgeCraft Toolchain

The Executable Sprint was driven by ForgeCraft MCP actions:

Action Role
generate_harness Scaffolds hurl probes from use-cases
run_harness Executes hurl probes, writes harness-run.json
generate_env_probe Scaffolds bash probes from .forgecraft/env/ specs
run_env_probe Executes bash probes, writes env-probe-run.json
generate_slo_probe Scaffolds k6 scenario from .forgecraft/slo/ spec
run_slo_probe Executes k6 ramp, reports threshold results
layer_status Shows L1–L4 coverage picture
close_cycle Evaluates all gates — returns READY when all tiers pass

Source: github.com/jghiringhelli/forgecraft-mcp