← Back to TIL
Second Gear / July 23, 2026

Review Needs Proof

Aggregation

World

1 critical / 0 happenings

Aggregation

Tech

1 critical / 2 happenings

Ideation

Ideation

Idea 1 Causal Proof for Agent-Written Code

Sell security and platform teams a merge gate that proves what an AI-authored change can affect before it ships. The product builds a causal graph across the ticket, prompt, files touched, tests, feature flags, deploys, traces, ownership, and past incidents, then asks for specific evidence instead of another broad code review. The new primitive is a proof bundle for software changes: not who wrote the code, but what chain of effects the company is willing to trust.

Source Signals

Why Now

Coding agents are making pull requests cheaper than senior attention. At the same time, the obvious response, AI code review, is already crowded and easy to add to existing developer tools. The harder unsolved job is deciding which changes need proof, what proof is enough, and how to preserve an audit trail when autonomous agents modify systems that already have production history.

First Wedge

A GitHub/GitLab merge gate for fintech, healthcare, infrastructure, and B2B SaaS teams that flags AI-authored or agent-assisted PRs touching high-risk codepaths. For each PR, it produces a required evidence checklist: affected services, owners, invariants, tests that must run, production traces to compare, rollout limits, and rollback triggers.

Commercial Model

The buyer is the head of platform, AppSec, or engineering productivity at companies already paying for SAST, CI, observability, and AI coding seats. Price by active repo plus high-risk PR volume, starting as a $25k-$75k annual team product and expanding into governance reporting for regulated software teams.

Defensibility

The compounding asset is the causal map between code structure, runtime behavior, ownership, incidents, tests, and deploy outcomes. A PR bot can copy comments; it cannot easily copy a customer's accumulated proof history and policy decisions across thousands of merges. Incumbents can add pieces, but most are anchored in one surface: code review, CI, observability, or security scanning. The wedge is the connective tissue between them.

Technical Risk

The hard part is producing low-noise causal claims. Static dependency graphs over-warn, traces are incomplete, tests lie about coverage, and agents may omit the prompt or context that caused a change. Version one should avoid claiming full correctness and instead prove narrower things: changed payment authorization path, missing regression evidence, unowned service dependency, risky flag rollout, or unmatched production trace.

Market Expansion

Start with AI-authored code gates, then expand into general change assurance for migrations, dependency upgrades, infrastructure-as-code changes, incident remediations, and eventually autonomous rollback approval. The same causal proof bundle could become the audit object for any machine-made operational change.

Self-Critique

This can die by becoming yet another noisy enterprise gate. If teams see it as slower code review with a fancier graph, they will bypass it. It also needs deep integrations before it feels magical, which makes sales and onboarding heavy. The narrow wedge has to be one painful workflow where false positives are tolerable because the cost of a bad merge is clearly higher.

Next Experiment

In two weeks, instrument 200 recent PRs from three AI-heavy engineering teams. For each merged PR, build a lightweight causal bundle from git metadata, CODEOWNERS, CI, test names, incidents, deploy records, and traces. Ask staff engineers to rank whether the bundle would have changed their review decision. Success is not perfect bug catching; it is a 30% reduction in reviewer uncertainty on high-risk PRs without increasing review time.