Known limits
This page names what Product Definition as Code cannot claim yet. It exists because a methodology about explicit, inspectable intent should apply the same standard to itself.
Maturity
Section titled “Maturity”There is one implementation, ProductShape, and the specification was extracted from it. Until an independent implementation passes the versioned conformance tests, conformance is self-conformance: useful regression evidence, not independent validation. There is one listed adopter. Governance is founder-led while the work is experimental. The full picture is in the maturity matrix.
What the tooling cannot enforce
Section titled “What the tooling cannot enforce”The rule that the definition changes only through an explicit Product Change, approved by a human and accepted through review, is enforced by branch protection and a CI validation gate. Those are repository configuration, not spec-enforceable invariants: in a repository without branch protection, a direct edit to the canonical branch followed by validation succeeds. Protecting the canonical branch is on the adopter.
Nor can the specification compel a modification to go through a Product Change in the first place. Structural validation runs against the resulting graph and has no way to ask whether a change record ever existed, so canonical model files edited directly still validate clean. Overlay validation constrains a change that exists; it cannot require one to exist. What you get is a place for the meaning of a change to live and a review boundary where a human accepts it, not a mechanism that makes bypassing it impossible.
What the model cannot see
Section titled “What the model cannot see”Structural validation proves the model is well-formed, not that it is true. The model records accepted intent; whether that intent is implemented, deployed or verified in a named environment is delivery evidence, which the methodology gives a place to land (citation statuses, coverage of cited scenarios) without claiming it. A green model and a divergent reality can coexist, and detecting that divergence is on you.
Topology
Section titled “Topology”The model repository may sit beside the code it describes or stand alone; both topologies conform, and the spec defines the pointer a consuming repository uses to name a dedicated one (Conformance, Topologies). What is not covered: verification runs against a local checkout of the model, so there is no remote model resolution, and one Product Definition cannot span several model repositories.
Ontology
Section titled “Ontology”The artifact vocabulary is a closed reference profile. Domain-specific artifact kinds and relationships (regulated obligations, safety cases, data contracts) do not fit yet. A namespaced profile mechanism is planned; the kernel is deliberately small and will stay small.
AI consumption
Section titled “AI consumption”Canonical Markdown that agents consume is also an attack surface. Digests prove content is unchanged, not that it is safe. A threat model covering prompt injection, provenance tiers and agent permissions is planned and does not exist today.
Cost and applicability
Section titled “Cost and applicability”Explicit models cost time to author and maintain, and semantic approval remains a human bottleneck by design. This methodology is most likely to pay off where product behaviour spans teams or repositories, knowledge has a long half-life, misinterpretation is expensive, agents or rotating teams repeatedly reconstruct context, or audit evidence matters. It is least likely to pay off for disposable prototypes, very small co-located teams, or products whose main uncertainty is market discovery rather than delivery interpretation. If that is you, do not adopt this; a good README and a short feedback loop will beat it.
Unknown in general. The testable hypothesis is that reduced re-derivation, ambiguity and untracked divergence exceed the authoring and maintenance cost for products with substantial change, risk or coordination complexity. External pilots must establish where that threshold lies, and their results will be published including failures.