Roadmap

What is known to be missing or unfinished. Nothing here is a commitment to a date. The Native Windows gap is the largest single item and is itemised there rather than repeated.

Known gaps

• Native Windows support. Not built, not run, itemised in docs/native-windows.md. Everything (es.exe) file search on Windows is part of that work.
• macOS is untested. The guard is stdlib Python and should run with --no-systemd, but nobody has tried it, and the launchd equivalents of the five systemd user units are not written.
• The installer and doctor are bash. Porting both to Python would remove the Git Bash dependency on Windows and would suit a project whose guard is already stdlib Python.
• The unit tests read the real $HOME. A machine with its own ~/.config/airlock/rules.json can fail the suite, which blocks a deploy that is otherwise fine. install/deploy.sh works around it by running the tests with HOME pointed at a temporary directory; the suite should isolate itself instead.

Evaluation and tuning

• The A/B bench has found no denies yet. The measured result so far is that the guard is cheap and stays out of the way, not that it has prevented anything. More adversarial bench tasks are needed before the guard's value can be claimed rather than assumed.
• The eval corpus skews toward the rules that already exist. Cases are labelled per rule, so a rule nobody wrote has no cases arguing for it.
• Threshold calibration is run by hand. tuning/calibrate.sh locks thresholds into the two eval/decisions-*.lock.json files; deciding when to re-run it is currently a judgement call, not a trigger.

Components

• Rewrite mode stays off by default. See the README section for why: a rewritten command is a command the user did not type.
• docclass/ ships no real taxonomy, only a generic example. Writing one is the adopter's job, and the README says to key off document kind rather than a folder tree.
• logtriage/ local rules are a small set. They cover the common shapes; anything else falls through to a model call.

Possible, not planned

• A named-pipe transport for the daemon on Windows, as its own piece of work.
• Structured judgements for row-level data sources, which would need a client-side tool and an explicit decision about what may leave the machine.