Continuous Correctness Audit · for AI-native systems

A model writes the function. A model writes the test that agrees with it.

Neither knows what you meant. The test passes because it was derived from the same code it checks, so a green build measures agreement, and agreement was never the question. The question your board will eventually ask is “who vouches for this code, and can they show their work”, not “did CI pass”.

Request confidential scoping   The engagement

Hero job: state the circular-test problem in one screen and route straight to the buyer-level stake (board accountability), because this segment's champion often reports to someone who will be asked exactly this. The board question is this page's single permitted "not X, it is Y" contrast; nothing else on the page may use that shape. House bans held: none of the three forbidden AI-hype adjectives appear anywhere on this page, copy or notes.
§ 1 · The moment

Software does the work now. Someone still has to vouch for it.

Most teams reading this page are living the shift from the inside: production changes a model wrote and a reviewer skimmed. The wider industry is building entire service firms on the same premise, software doing the work with people supervising it.

The failure mode is already on record: audit-style paperwork generated at volume, conclusions that look alike from client to client, and no evidence anyone can re-run. When that work gets questioned, there is nothing underneath it.

Whichever side of that shift you are on, the scarce asset is the same: evidence someone outside the building can re-run.

Structural job: cash in the current wave honestly, then reframe its bottleneck. The hype says the constraint is headcount; this section says the constraint is verifiable evidence, which is the page's contrarian spine and the setup for everything below. Specifics deliberately kept off-page, VERIFIED against the public record 2026-08-02: YC Su26 RFS "AI-Native Service Companies" (Alströmer; names audit as a target vertical; "they just do the work"), Sp26 "AI-Native Agencies" (Aaron Epstein, not Migicovsky as an earlier digest said), Fall 2026 "AI-Native Compliance Infrastructure", and the Delve expulsion 2026-04-04 (fabricated SOC 2 evidence, removed from YC directory, YC silent). Pitch calibration confirmed; on-page copy stays nameless.
§ 2 · What the bar catches

Four things model-written code gets wrong, and where the audit stands.

Diagram placeholder: owner-approved spec at the top; model-written code and model-written tests each measured against the spec, so the two can no longer certify each other.

Signed intent, recovered first

We turn what you meant into requirements the engineers who own the component approve before any code is judged, about two hours of their time in week one. The spec exists outside the model that wrote the code. That is the point of it.

Tests measured against the spec

A generated test usually mirrors the code it was generated beside. We measure your suite against the approved requirements instead of the implementation, so a test that agrees with a wrong function gets counted as what it is: agreement.

MC/DC on the decision logic

Condition-level MC/DC, a strictly harder bar than line coverage, on the decision logic that code-writing tools touch most and test least.

Timeouts and error paths

The least-tested surface in model-written code, by construction: models learn from examples, and examples are mostly happy paths. The bar requires timeout and error branches to hold at condition level, with evidence.

Each finding is a red reproducer, which is exactly the target a code-writing model needs: point your tooling at it, and the gate grades the whole change. A fix isn't done because the model says so; it's done when the reproducer finally passes and the gate agrees. The coverage bar is the same in all eleven languages we measure, so a polyglot codebase does not dilute the standard. Findings arrive agent-ready: every known issue carries a copy-ready prompt, context and reproducer included, so your own coding agents can be pointed at the fix with the evidence as the acceptance test. And unlike a session, the corpus persists: every future model and every future engineer inherits the same written intent instead of re-guessing it. None of this is hypothetical: read a public corpus, browse its live dashboard, or the ledger it feeds.

Structural job: convert "we audit" into segment-specific mechanism. The closing paragraph is the fix-loop payoff (founder field material 2026-08-01), stated without "agentic" vocabulary; for this audience the red reproducer as a fix target for code-writing models is the sharpest operational hook on the page. The four cards map one-to-one to the known failure modes of model-written code, so a staff engineer can check them against their own incident history. The MC/DC card carries no uniqueness claim (founder ruling 2026-08-03: capability statements only for MC/DC site-wide; the former Go superlative is retired, see guide §8). Dark band per house style: this is the closest thing to an artifact section on this page.
§ 3 · Two numbers

Two numbers, and only two.

Stack Overflow's 2025 developer survey put the top frustration with AI tools at 45%: solutions that are “almost right, but not quite.” The same survey found 66% of developers spending more time fixing almost-right AI code (the survey).

Structural job: minimal external validation for a stat-skeptical audience that distrusts marketing more than it distrusts AI. Exactly two stats, dated and attributed, both from BRIEF's permitted pool; index.html's pair (SO 46%/33%, DORA 90%/30%) deliberately not reused per house rule. The closing sentence is the section's real move: it tells the reader we know stats are cheap.
§ 4 · The pilot

One component a model helps write. Four weeks. Evidence in your repo.

The same mechanics as Proof's flagship Continuous Correctness Audit. You choose one component with us, one where model-written changes land weekly and customers depend on the result. We scope it and quote a fixed fee, agreed before work starts. Roughly four weeks end to end.

Every finding ships with a runnable reproducer that fails on your current main, validated by a named reviewer, or it does not ship. The evidence corpus lands in your repo as plain YAML and executable tests, and the audit gate runs in your CI without us. Optionally continuous afterwards, a retainer on the cadence you choose, up to daily, so the claim gets re-earned on every release.

When the board asks who vouches for this code, the answer is an artifact: the signed register, and evidence you can re-run in front of them. Every finding is signed by name, so we take a limited number of engagements each quarter. The full engagement contract is on the flagship page →

Structural job: land the offer without restating the whole contract; this page sells the fit, the flagship page sells the terms. Every claim here is from the fact inventory: trust primitive, fixed fee after scoping, ~4 weeks, corpus-in-repo continuity, cadence-choice retainer billed flat monthly, limited capacity. No prices, no tiers, per honesty rules.
§ 5 · Fit

Sound familiar?

A fit if

  • Model-written code already runs in production paths your customers depend on.
  • You can name the component where a wrong merge costs real money.
  • You have a release cadence an audit gate can attach to.
Structural job: disqualification as the conversion path, segment-tuned. The honest "you do not need us yet" line is deliberate and load-bearing: this audience is allergic to being sold to, and the line is also true, which keeps the page inside the honesty doctrine. Mirrors index §6 structure without copying its lines.
§ 6 · Scoping

Request confidential scoping.

Your email is the only required field; the rest is optional context. We reply personally, with a scoping call or an honest no. Fees are quoted after scoping, fixed before work starts.

Private by default. Keep source out of the form; code is read only under NDA, after scoping.

Frozen form structure from index §9: the 7 field names (email/company/component/timeline/cadence/confidentiality/message) are locked by subscribe.js; only data-source ("icp-ai-native") and the form-note line differ, both per the page spec. ⚠ Same caveat as index: the deployed subscribe function persists only email+source until the richer function ships, so segment attribution works today but qualifier answers silently drop.