Weakness axis #4

Intent drift

When the code stops doing what the spec said it should — caught before it ships.

Run on your repo →See pricing

What this axis covers

The most ambitious of the repair axes. Securie reads ADRs + grok-extracted L1 symbols + spec docs + actual execution traces, then semantically compares against current code. When intent and code diverge, Securie opens a fix PR on whichever side rotted (usually code; sometimes docs when grok detects doc-rot).

Why now

AI agents iterate on code at a pace that makes documentation-update a chronic loser. Spec said the auth check goes here; six PRs later the auth check is gone but the spec didn't notice. Intent drift is silent — until it isn't.

Where it hides in your codebase

  • Code that no longer matches the ADR that motivated it
  • Test names that lie about what they test (after the test was refactored)
  • API surfaces that shipped a 'private' helper that consumers now depend on
  • Wired-but-inert code (function called but does nothing — see /features/inert-surfaces)
  • Comments that contradict the surrounding code (the comment said one thing, the AI replaced the code without updating the comment)

How Securie handles it

/align skill productized

Securie's intent-vs-code-aligner runs nightly across the Ring 0 Codebase Knowledge Base + ADR set + execution traces. Mismatches escalate to a fix PR or a doc-update PR depending on which side rotted.

Intent-contract substrate (ADR-072 §1)

Every component declares a machine-checkable contract (effects, capabilities, dataflows, auth boundaries). Code that diverges from its contract is a first-class finding, not a soft-warning.

Cross-PR pattern detection

Architectural auditor specialist (Business+) tracks intent across multiple PRs — a single PR may look fine; three PRs in sequence may collectively drift the auth model.

What this axis is NOT

Not a doc-rot detector

Documentation drift is one symptom of intent drift, not the whole axis. Securie also catches code drift FROM stable documentation.

Not a copilot review

AI review tools comment on PRs without a model of declared intent. Securie's intent-vs-code-aligner has explicit machine-checkable contracts to verify against.