AI application / release engineering

Release Guardian

A 13-node LangGraph release gate: four evidence retrievers, deterministic validators, bounded retries — and the publish decision stays with a human.

Problem: Before code ships automatically, someone needs to know what it touches, keep the model inside clear limits, and leave a durable approval record.
Audience: Platform, release-engineering, and applied-AI teams that have to review automated change decisions — and then explain them.
What I built: I designed the four-retriever workflow and its bounded validator retries, then built the services around it in Python, Go, Java, and TypeScript: the human approval gate, the observability layer, and a tamper-evident SHA-256 audit hash chain.
Result: In the funded live evaluation, all eight aggregate gates passed across 132 graph runs, with 100% citation fidelity, zero tool misuse, and 100% injection defense.
Links: GitHub repositoryOpens external site

How it works

01
Intake
Classify and parse an incoming change.
02
Evidence
Gather code, schema, API, and operations evidence in parallel.
03
Risk
Grade the change, build a plan, and check both against the rules.
04
Approval
Pause and wait for a human decision before publish.
05
Audit
Publish through the approved branch and record the decision.

Try it

Loading synthetic replay...

Separate measured evidence

Funded live evaluation context

Aggregate gates8 / 8

aggregate gates passed across 132 funded runs (44 scenarios x 3 trials).

Definition: a scenario is flagged if any criterion failed in any of its three trials — a stricter lens than the aggregate gates.

Strict all-trials residual30 / 44

scenarios had outcome_pass: false; one trajectory failure.

Total model cost$8.1214measured, 2026-07-11 funded run

Mean latency per run~35.08 smeasured, 2026-07-11 funded run

What I changedI turned three distinct operating modes into a review surface that makes their results immediately comparable.

Live and stub metrics retain their own ledgers, while the interactive replay demonstrates the end-to-end review workflow.

Funded liveMeasured: 132 runs / strict 30 of 44 flagged

Existing deterministic stubStub: 132 runs / strict 15 of 44 flagged

New change replayInteractive demonstration of the complete review workflow

Funded live aggregate metric detailMeasured 2026-07-11; read these only alongside the strict residual above.

Missed dependency17.42%<= 25%

False impact12.32%<= 25%

Risk grade accuracy72.73%>= 70%

Plan completeness95.98%>= 90%

Citation fidelity100%>= 99.9%

Tool misuse0%<= 0.0000001%

Step efficiency1.001<= 1.35

Injection defense100%>= 99.9%

Cost evidence from 2026-07-08 — before the migration

MeasuredRouted versus all-strong observationEstimatedPrompt-pruning token countsProjectedAssumption-dependent cache savingsModeledReAct call and cost comparison

IDAudit findingPublic disposition

W3-02Current-evaluation copy mixed deterministic stub values into a live-sounding statement.Keep live and stub evidence separate and visibly labeled.

W3-03Aggregate-pass wording omitted the strict all-trials residual.Place the 30/44 live residual beside the aggregate gates.

W3-04The funded Phase-L reports are archive-only artifacts.Use the verified archive hash as the durable report reference.

W3-05Cost material combined measured, estimated, projected, and modeled evidence.Preserve evidence class and the pre-migration snapshot date on every claim.

W3-06Workstation-selected models differ from tracked defaults.Label the runtime settings independently from repository defaults.

View all 13 sanitized findings

Sanitized deterministic risk guardrail — The risk view turns change type, CI history, and downstream impact into an inspectable score with factor-level contributions.

Sanitized deterministic stub evaluation — The evaluation board places the regression gate beside six operating metrics so reviewers can read pass/fail status at a glance.

Sanitized deterministic stub pipeline trace — The trace reveals how intake, retrieval, grading, planning, validation, and approval spans connect across the release workflow.

How this was verified

Funded live report dated 2026-07-11: 44 scenarios, three trials, 132 graph runs; raw report is archive-only and not tracked in the source repository.
System and delivery details come from source commit ca2ef58 and the W2/W3 review of claims against their source files.
The exact nine-file sanitized package was approved after its immutable candidate manifest was generated; the approval record is published separately.

What this does not prove

Aggregate pass does not mean every scenario passed strictly: 30 of 44 scenarios failed at least one criterion across all trials.
No local rerun was performed on this Mac. Workstation-selected models are not the tracked repository defaults.
The private source is not linked. The page shows only sanitized files that passed the recorded publication checks.