Skip to benchmark stories
VERDICT / BLUEPRINT BENCHMARK

10 SAMPLE PRDS · BLUEPRINT BENCHMARK · AS OF 2026-10-15

From Monolithic Chaos to Atomic Execution.
Benchmarking PRD Decomposition Across 10 Sample PRDs.

Verdict Blueprint parses PRD markdown, prunes non-delivery sections, generates an atomic task DAG, and routes each step to a model. On these 10 sample PRDs, Verdict passed 202 / 202 assertions for $0.03in total. Only Verdict's results are shown; no competitor comparison is published for this suite.

Assertions Passed · Sample PRDs202 / 202All 10 sample PRDs passed
Total Wall-Clock · Sample PRDs148.2sSum of the 10 per-story times
Total Cost · Sample PRDs$0.039 of 10 stories ran at $0.00
Scope Violations · Sample PRDs0Recorded across all 10 stories

Verdict Blueprint Across the 10 Sample PRDs

Sample PRDs · Verdict-only figures from stories.json · as of 2026-10-15.

Loading analytics charts...
Verdict Cumulative Cost$0.03All 10 sample PRDs
Stories at $0.009 / 10Local M5 Max models
STORY 1 · EASY Code-API-Design

Strict ISO-8601 Parser with Leap-Second Rollover

Microservice Timestamp Contract
Act 1: The PRD Specification & Traps
PRD calls for a strict ISO-8601 UTC timestamp parser handling fractional seconds and leap-second rollover. SECTION OUT OF SCOPE explicitly bans importing third-party libraries (e.g. dateutil) or accepting non-UTC local formats.
Act 2: Blueprint DAG Decomposition & Model Formation
📋 Blueprint Decomposed Task Tree (DAG):
  • Task 1.1: Regex boundary validator for ISO-8601 strings
  • Task 1.2: Native datetime.fromisoformat UTC normalizer
  • Task 1.3: Leap-second (60s) modulo rollover sanitizer
⚡ Verdict Blueprint ExecutionPASS (12/12)

verdict-4-35b-toolterse-v5-mlx

Runtime: Local M5 Max (Concurrency=4)

Blueprint Scope PrunerF17 (Claim Audit)
Act 3: Verdict Outcome · Sample PRD
MetricVerdict Blueprint
Pass StatusPASS (12/12)
Wall-Clock Time4.1s
Token Consumption410 tokens
Inference Cost$0.00
Scope Violations0
VERDICT BLUEPRINT SOLUTIONSample PRD
from datetime import datetime, timezone

def normalize_timestamp(ts: str) -> str | None:
    if not isinstance(ts, str) or not ts.strip():
        return None
    clean = ts.strip().replace('Z', '+00:00')
    try:
        dt = datetime.fromisoformat(clean)
        return dt.astimezone(timezone.utc).strftime('%Y-%m-%d %H:%M:%S.%f+00:00')
    except ValueError:
        return None

02 / ARCHITECTURE & GUARDRAILS

The Three Pillars of Verdict Blueprint

How Blueprint eliminates prompt bloat, out-of-scope creep, and execution failure before code is written.

1. Automated Scope Exclusion

PRD documents frequently contain "Out of Scope", "Non-Goals", and "Future Work" sections. Blueprint's decomposer explicitly prunes non-delivery sections (is_non_delivery_section), preventing wasted tokens and unwanted dependencies.

2. TMD Lane-Typed Model Routing

One size does not fit all. Blueprint classifies tasks into typed TMD lanes (Code-API-Design, Code-Impl, Data-Schema, Ops-Monitor) and routes to specialized models: Toolterse v5 MLX for rapid tool calling, Qwen 30B for complex algorithms, GPT-OSS 20B for BigCodeBench Hard, and GLM-5.3-Flash for cloud tasks.

3. CI Guardrails & Fixed Quotes

Blueprint computes P50, P80, and P95 token and duration estimates. At every lane boundary, runtime guardrails check cumulative actuals against the confidence interval band. Overruns trigger an automatic downshift to the cheap local tier or a pause for human review.

Data Source

Every figure on this page is read from the per-story Verdict values in one saved data file, and totals are summed from those values. The inputs are sample PRDs. No separate run receipt is published for this suite, so no competitor comparison is shown.

  • cloud/web/lib/blueprint-benchmark/stories.json — Sample PRD stories, DAG task plans, model selections and Verdict outcomes (202/202 assertions, 148.2s, $0.03 total). As of 2026-10-15.

Ready to plan and execute with Blueprint?

Turn PRDs into small, checked tasks.

Download Verdict for Mac