ADR 0002 — No GPL code; clean-room implementation from papers

Accepted (amended 2026-10-31: scope, the structure allowance, weak copyleft rejected; amended 2026-11-06: black-box benchmarking) · 2026-05-21 · cortex/docs/adr/0002-no-gpl-clean-room-policy.md

Status: Accepted (amended 2026-10-31: scope, the structure allowance, weak copyleft rejected; amended 2026-11-06: black-box benchmarking) Date: 2026-05-21 Source: docs/plan/bootstrap-plan.md, judgment call #3 Related: ADR-0001 (dependency licenses), ADR-0006

Context

The reference implementations for VIO and SLAM are mostly GPL-licensed: OpenVINS and MINS (GPL-3.0), ORB-SLAM3, DBoW3, and the GPL parts of the Sophus/g2o ecosystem. Vendoring OpenVINS as a test-only oracle was proposed, since a separate test target would not link GPL code into the SDK.

cortex is meant to be redistributed by OEM partners (AV makers, robotics vendors) inside their own certified products. Any GPL obligation, even one that seems confined to tests, is something those partners’ legal reviews would have to clear.

Decision

No GPL-licensed code anywhere, for any purpose. Not as an SDK dependency, not in a test target, not as a build-time tool in what cortex builds, and not as a ground-truth oracle. The scope is what cortex ships, the code that builds it and the tests; the docs site’s toolchain is not part of it (see the amendment below).

• Operators are implemented clean-room from papers and documentation, plus permissively licensed (MIT / BSD / Apache-2.0) references.
• Reading GPL source and then “writing from scratch” is not clean-room. Check a reference’s license before opening its code. A GPL project’s structure may be consulted, meaning its directory layout and how it splits the problem into modules (e.g. sensor-modular decomposition), and nothing else: never its code, algorithm details or data (tuning tables, gains, coefficients, vocabularies, weights).
• Test oracles are in-house: synthetic trajectories, landmarks, and IMU noise models (sdk/include/branes/sdk/eval/synthetic_world.hpp), plus published-paper numerical thresholds. No test, CI job or build step runs a GPL binary or reads its output.
• Black-box benchmarking is allowed (amendment 2026-11-06, below): an unmodified GPL estimator may be run outside every Branes repository, and the accuracy it scores may be recorded as a target.
• LGPL, MPL, AGPL and SSPL sources are rejected, as dependencies and as references to port from, as ADR-0001 rejects them. An exception would need an ADR of its own that supersedes this one.
• Amendment (2026-10-31): the three rules above replace looser wording (“no GPL contact”, “structure only”, LGPL / MPL / SSPL “flagged for review and default to rejected”), to match the same policy as recorded in synapse, reflex, anatomy, racing-drone and scout-drone. The scope rule matters for the docs site: Astro brings in sharp (Apache-2.0), and sharp brings prebuilt libvips binaries, the @img/sharp-libvips-* packages (LGPL-3.0-or-later). They only build the documentation, in CI and on developers’ machines, and are not part of anything cortex ships.

Amendment (2026-11-06): black-box benchmarking

The goal for the VIO pipeline is accuracy competitive with OpenVINS on EuRoC. Published tables don’t settle that: they use their own sequence subsets, start offsets, alignments and evaluators, so a gap between our number and a paper’s mixes the estimator with the protocol. Running the other estimator ourselves, through our own evaluator, removes the protocol from the comparison. GPL-3.0 governs distribution and modification, not running a program, so running one creates no obligation for cortex. Mixing what we learn from its internals into cortex is the risk this ADR exists to prevent, and the conditions below keep the two apart.

Allowed. Running the unmodified upstream binary of a GPL estimator on public datasets, built from its upstream release, outside every Branes repository (for example in a container on a developer machine). An in-house evaluator scores its output trajectories, cortex’s output by the same code, and the evaluator may live in cortex. The results may be recorded in cortex docs, issues and PRs as aggregate metrics: ATE, RPE and NEES per sequence, with the estimator’s name, release and the run mode (e.g. “OpenVINS v2.7, monocular, upstream defaults, upstream start offsets”). Those numbers are targets, in the same standing as a paper’s published table.

Still forbidden.

• Reading the estimator’s source, configuration files or tuning values. Running it with its shipped defaults needs none of them; only its launch arguments (dataset, bag, start offset, camera count, output path) may be set. Its settings are data in this ADR’s sense, so they are never read, quoted or copied into cortex.
• Taking algorithm details from it, its source, its runs or its outputs. Ideas come from published papers and documentation, as before, and are cited.
• Anything GPL in what cortex builds, ships or tests. The scripts that fetch, build or run the binary stay outside the Branes repositories. No test, CI job or build step runs it, and no test asserts against its output; a regression target in cortex is a recorded number, like a published threshold.
• Committing its raw outputs (trajectories, logs). Only the aggregate metrics are recorded.

Disclosure. The first baseline (2026-11-06) was run before this amendment, under the reading that an external binary was outside the rule. In setting it up, lines of OpenVINS’s EuRoC configuration file were read (initialization thresholds and feature counts), and so was the argument list of its launch file, including the start offset it suggests for each sequence. The offsets were used to run it; none of the configuration values is used in cortex. One suggestion that came from them, an initialization trigger, was removed from issue #591. That run’s comparison results are admitted under this amendment.

The other repositories that record this policy (synapse, reflex, anatomy, racing-drone, scout-drone) are not changed: none of them benchmarks against a GPL estimator.

Consequences

• Every operator costs more to build: the math has to come from the literature, and we have to build our own simulators to validate it.
• Each algorithm records its citations so provenance can be audited (docs site: Algorithms & Provenance → Clean-room Citations).
• OEMs can redistribute cortex under MIT terms with no copyleft review.

Enforcement

• PR review and the CodeRabbit path instructions flag forbidden references (OpenVINS, MINS, ORB-SLAM3, DBoW3, Sophus/g2o) and non-permissive FetchContent dependencies. A benchmark result that names one of them is allowed under the 2026-11-06 amendment; code, settings or algorithm details attributed to one are not.
• Issues touching provenance carry the clean-room / no-gpl labels.