Proof

If you distrust demos, start here. This page is the position, the proof, and the honesty in one place — with the caveats left in. Everything below either links to a running product doc or embeds a recording captured from the live fleet. We claim a position, not market share.

The self-hosted governance and verification layer for autonomous coding: the factory that runs your agents' code, tests it for real, and refuses to overclaim. Trust-verified, not code-faster.

The gap we sit in

The 2026 data says buyers already have agents — what they lack is a way to trust them.

84% / 29%developers using AI coding tools vs those who trust the output
81% / 14.4%teams with agents deployed vs teams with security approval for them
53%of deployed agents run unmonitored
25–45%the gap between SWE-bench scores and real production reliability
Aug 2 2026EU AI Act high-risk rules: logging, human oversight and audit become mandatory

Two axes decide the market: where the code runs (someone else’s cloud vs your own infrastructure) and what the product optimizes for (writing code faster vs proving the code is trustworthy). Almost every vendor clusters in the SaaS + code-faster corner. The self-hosted and verification-first quadrant is nearly empty. That is where Factory sits.

quadrantChart title Where the tools cluster x-axis "SaaS / vendor cloud" --> "Self-hosted / your infra" y-axis "Optimize: code faster" --> "Optimize: trust and verify" quadrant-1 "Empty quadrant" quadrant-2 "Compliance suites" quadrant-3 "SaaS agent platforms" quadrant-4 "Self-hosted coding agents" "SWE-bench vendors": [0.22, 0.18] "SaaS agent platforms": [0.30, 0.34] "Local coding agents": [0.74, 0.28] "Factory": [0.80, 0.82]

The upper-right quadrant — self-hosted and verification-first — is the position. Dot placement is illustrative of category, not a benchmarked score.

Factory vs the alternatives

Scored on the dimensions a skeptical engineering buyer actually asks about. We only mark Yes where the fleet does the thing today.

Dimension Factory SWE-bench-score vendors SaaS agent platforms
Self-hosted Yes — runs on your own k8s; local models via Ollama No — hosted service Rarely — cloud-first, some enterprise VPC
Independent verification Yes — TFactory is a separate product that grades AIFactory's output; builder never signs its own homework No — vendor reports its own benchmark score No — the system that writes also declares "done"
Real test execution Yes — tests generated and run in a per-task sandbox (Nix env), graded on a 5-signal verdict Benchmark harness only — not your repo Varies — often lint/build, not a graded test verdict
Governance / HITL Yes — PFactory review gates with citations; human approval before code is emitted; approve/merge in the cockpit No — score in, patch out Some — PR review after the fact, not a gate before work
Audit / evidence Yes — HMAC-anchored audit log, completion-event records, screenshot evidence for browser tests No Limited — activity logs, rarely tamper-evident
Data-egress control Yes — self-hosted with local models; code and prompts need never leave your cluster No — code goes to the vendor No — cloud by default
Refuses to overclaim Yes — verification-core drift guard on the never-overclaim path; VAL assurance levels; false-failed builds fixed rather than hidden No — optimized to the leaderboard No — "done" is asserted, not proven

The proof, recorded

Captured from the running fleet. Nothing here is a mockup.

PARR pipeline deploying a service and then verifying it against the live endpoint
Flagship — deploy, then verify. One task driven through the whole PARR pipeline: planned and governed, built, deployed, then tested against the live endpoint. This is the core claim in motion — the factory runs the code it wrote and checks it for real, rather than declaring success on a green build.
TFactory generating and running Python unit tests
Real test execution. TFactory generating and running unit tests in a sandbox, then grading them — coverage delta, stability, mutation, lint and semantic relevance — not just a passing bar.
TFactory verifying multiple languages
Polyglot, same verdict. The same graded verification across languages — the test lane is language-aware, provisioned per task via Nix.
Browser test solving an MFA one-time-password challenge
Evidence, not assertion. A browser test logging in through a real MFA one-time-password challenge — proof the agent exercised the authenticated UI, not a stubbed page.
Authenticated account page reached after MFA login
The authenticated result. The account page the browser test reached after passing MFA — captured as a screenshot artifact attached to the run.
CFactory cockpit mission control view
You can watch it. The CFactory cockpit threads plan to code to test for every task — the observability the "53% run unmonitored" statistic is missing.
CFactory audit log view
Audit trail. Human-approval gates and completion events land in an HMAC-anchored audit log — the logging and oversight the EU AI Act asks for.

One task, end to end

PFactory plan detail
Prepare. PFactory decomposes the plan into acceptance criteria and child issues; review gates pass before any code is emitted.
AIFactory subtasks
Act. AIFactory builds it subtask by subtask against the signed plan.
TFactory verification report
Reflect. TFactory verifies the build and returns a graded verdict — with a bounded handback to the coder on failure.
CFactory execution DAG
Review. The whole run as a live execution DAG in the cockpit, one correlation key threading plan to code to test.

Proven on real clouds

Service deployed to Azure
Azure. A factory-built service running on real Azure infrastructure, then torn down.
Service deployed to GCP
GCP. The same pipeline deploying and verifying against live Google Cloud, provider-agnostic through the plan.

What we do NOT claim yet

The caps are the trust signal. We would rather show you the ceiling than let a demo imply we cleared it.

  • Verification assurance is capped at VAL-2. Higher assurance levels are designed but not yet the default — we do not claim VAL-3 evidence.
  • Deployment is dry-run first. The deploy lane plans and dry-runs real infrastructure; full unattended production rollout is deliberately gated, not "click once and ship."
  • Concurrency is bounded. Per-task Nix environments on a single node cap throughput at roughly a handful of concurrent tasks today — scaling is in progress, not finished.
  • It is an early, open project. No revenue, market-share or named-customer claims. The scenarios in these docs are illustrative of how a team would use Factory.

Go deeper