Skip to content

performance-report.md

Latest commit

 

History

History
101 lines (79 loc) · 4.32 KB

performance-report.md

File metadata and controls

101 lines (79 loc) · 4.32 KB
Raw

M5 Performance Acceptance Report

Verdict

PASS WITH LIVE SCOPE CUT

Subject main SHA: 76b64eb59d1cd4e9e3af9f2b0e9be0b37d0c20bd

Execution date: 2026-08-09 (Asia/Hong_Kong)

This is a single-operator benchmark. No independent-review claim is made. The production Fixture, browser, verification, export, timeout, and byte-limit checks pass their targets and hard-fail thresholds. Live quote and successful Live preflight latency remain unmeasured because the baseline has no configured RPC and the Web route intentionally has no Live session. The five Live attempts below are recorded as deterministic, visible LIVE_UNAVAILABLE failures, not as successful Live performance samples.

Environment

  • Machine: Apple M5 Pro, macOS, darwin/arm64, 51,539,607,552 bytes memory
  • Node.js: v22.23.1
  • pnpm: 11.16.0
  • Browser: Playwright Chromium 148 via Playwright 1.60.0
  • Clear402: enabled for the production build and verification samples
  • RPC/network: not configured (network.configured=false)
  • Chrome DevTools MCP: unavailable in this execution environment; the benchmark uses Playwright Chromium and the browser Performance API instead

Commands

CLEAR402_ENABLED=true pnpm test:performance: PASS
  - production build
  - 2 performance guard tests
  - production server benchmark

The benchmark starts next start on a loopback port, runs against that production build, and terminates the process after the sample matrix.

Acceptance Matrix

Metric Samples Min Median P95 Max Target Hard fail Result
Fixture response (ms) 10 3.62 4.47 6.14 6.14 <300 >1,000 PASS
First interactive (ms) 10 63.70 71.20 107.30 107.30 <2,000 >4,000 PASS
Credential verify (ms) 10 2.20 2.52 3.47 3.47 <100 >500 PASS
Export JSON (bytes) 1 4,710 4,710 4,710 4,710 <2,000,000 >5,000,000 PASS
Large raw evidence render (ms) 1 55.56 55.56 55.56 55.56 <2,000 >4,000 PASS
Large raw evidence drawer (ms) 1 69.80 69.80 69.80 69.80 visible without freeze >1,000 PASS

Raw timing samples from the exact run:

Fixture response ms:       4.96, 5.48, 6.07, 4.32, 3.94, 6.14, 4.60, 4.47, 3.98, 3.62
First interactive ms:    107.30, 68.00, 71.10, 71.70, 63.70, 71.20, 72.10, 71.60, 72.20, 70.50
Credential verify ms:      3.47, 2.52, 2.63, 2.26, 2.33, 2.56, 2.20, 2.22, 2.56, 2.52
Fixture response bytes:    6,419 for all 10 samples
Verify response bytes:     31 for all 10 samples
Large raw payload bytes:   1,500,000 injected; 1,502,991 serialized response

Live Boundary

Five production Live attempts were made with the fixed valid intent contract:

count:              5
status:             LIVE_UNAVAILABLE for all attempts
response ms:        2.80, 2.29, 0.91, 2.22, 2.32
p95 / max:          2.80 / 2.80
response bytes:     183 for all attempts
run IDs:            5 unique v4 run IDs

The response code and UI error state were both checked for LIVE_UNAVAILABLE. A separate browser route test returned the real PREFLIGHT_TIMEOUT 504 contract and verified the code/message became visible in the Live error panel in 50.81ms. Since there is no Live session, no claim is made for the <8s Live Quote or <15s successful Live preflight targets.

Size And Timeout Guards

  • Fixture response cap: 2,097,152 bytes; observed response: 6,419 bytes.
  • Clear402 verify request cap: 2,097,152 bytes; observed response: 31 bytes.
  • A preflight body beyond 65,536 bytes returned 413 REQUEST_TOO_LARGE.
  • A verify body beyond 2,097,152 bytes returned 413 REQUEST_TOO_LARGE.
  • A 1,500,000 byte raw evidence payload rendered and opened without UI freezing or exceeding the 4 second hard bound.
  • Orchestrator budgets remain 8 seconds for quote and 25 seconds total; the performance guard tests enforce both remain below the issue hard-fail limits.

Scope Decision

The Fixture and local product loop meet the performance acceptance criteria. Live latency is an explicit scope cut for this baseline because /api/health reports network.configured=false and the production route does not resolve a Live session. Enabling Live measurements requires a separately authorized RPC environment and a fresh run; it is not inferred from fail-fast unavailable responses.