APAC Regulatory Readiness
Europe's regulatory difficulty is depth. Asia-Pacific's is concurrency — and concurrency is a program-management problem before it is a legal one.
The problem this is built around
APAC has no DSA. No single instrument, no single regulator, no harmonised reporting calendar — instead roughly a dozen sovereign regimes converging on the same four themes from different legal traditions and at different speeds.
The operational consequence is that one product change can trigger obligations in eight jurisdictions with eight deadlines, eight evidentiary standards, and eight regulators who do not coordinate with each other. India runs a three-hour statutory takedown clock. Australia runs an effort standard where the evidence trail is the compliance position. Korea attaches personal liability to a company officer.
A three-hour clock is not a policy question. It is a staffing and tooling question.
You cannot convene a cross-functional review in three hours, so the work has to move upstream — pre-agreed decision rights, a named on-call owner per surface, and a pre-approved decision tree, so the person on shift at 2am is executing a decision that was already made rather than making one.
Four parts, because a tracker alone is a list
A structured dataset of the regulation; the reactive workflow that responds to regulators; a toolkit of AI prompts with the measurement attached; and a readiness framework that turns an obligation into a dated, owned, costed plan.
The parts are joined, not adjacent. An inquiry record references the tracker instrument it arises under. The readiness assessment pulls its likelihood scores from the tracker's recorded enforcement signals. The consistency-check prompt runs at a named stage of the response plan.
The finding
The models find the obligation and miss the exemption.
Across five open-weight models, F1 on obligation extraction runs 0.79–0.96 — which reads as solved. Carve-out recall on the same runs is 0.00–0.50.
And the failure is not random. It splits on the grammar of the carve-out:
| Carve-out expressed as | Example | Models capturing it |
|---|---|---|
| Explicit | "does not apply to…" | 4 / 5 |
| Exception clause | "…unless the direction has been varied" | 3 / 5 |
| Threshold condition | "more than 100,000 monthly users" | 0 / 5 |
| Transitional deferral | "periods commencing on or after 1 July 2027" | 0 / 5 |
Where the text announces a carve-out using carve-out vocabulary, models find it. Where the limit is carried by a threshold or a commencement date, no model in the set found it, at any parameter count. Scaling from 3B to 14B does not help, because this is not a capacity problem — the models are classifying text that looks like an exception, and treating everything else as scope-setting prose.
Both missed forms are the expensive ones. A missed threshold over-scopes compliance, funding remediation for a population the law excludes. A missed deferral puts a phantom 2027 deadline into a 2026 plan, where it competes for the same engineers as a real one.
A missed obligation gets caught in legal review, because counsel is reading for what the platform must do. A missed carve-out is far more likely to survive review, because the output is not wrong about the duty — only about its edges.
Separately, and despite an explicit instruction not to: 5 of 5 models extracted a duty imposed on a Minister as a platform obligation. Recall is unaffected, so the output looks complete — and the spurious row is a duty belonging to a government ministry.
What the prompt engineering actually bought
The superseded prompt ships alongside the current one and is runnable, so the comparison reproduces rather than has to be believed.
| Measure | v1 → v2 |
|---|---|
| Fabricated an obligation on text containing none | 4/5 models → 1/5 |
| F1 | improved on 3 models, regressed on 2 |
| Trigger fidelity — what rule 6 targets | improved on 2, fell on 3 |
One guardrail worked cleanly. A single sentence telling the model that an empty list is a correct and preferred answer removed the fabrication failure on every model but the smallest. The other did not: the rule written specifically to protect trigger fidelity made it worse on three of five models, including a fall from 0.89 to 0.50 on a top-scoring model.
mistral-nemo:12b was the best model under v1 and the guardrails
made it the second worst, taking its carve-out recall from 0.25 to 0.00.
Evaluated on one model, v2 would have been reported as a clean improvement. Evaluated on five, it is a trade — and the rule that failed is documented as failing rather than quietly retained as a success.
Two things the tooling found that a document would not
The weekend-arrival trap. The timeline generator plans backwards from a deadline in working hours rather than calendar days. A notice arriving Friday evening with a Tuesday deadline shows 95 calendar hours and 17 working hours — 18% of elapsed time. The calendar reads comfortable; the budget does not. The decision between requesting an extension and standing up out-of-hours cover has to be made Friday evening, when both options are still cheap.
Plans that are individually achievable and collectively impossible. In the worked readiness assessment, every obligation has positive float against the enforcement deadline. The framework still fails it — because two of the six workstreams share one owner, whose serialised workload is 20 weeks against 19 available. Every row looks fine, which is exactly why this failure mode survives review.
What it does not claim
The repository was created in August 2026 and does not claim a long history of sustained monitoring. What it has is the machinery that makes monitoring cheap — schema, validator, link checker, staleness warnings, changelog discipline — so the history accrues from that date and is checkable in the commit log.
Jurisdiction depth is deliberately uneven and documented as such: three depth jurisdictions, five working, two monitor-only. Claiming equal command of ten jurisdictions is a claim no individual can support. The validator enforces the same honesty at entry level — an entry's confidence rating may not exceed the strength of the sources actually cited, a rule which caught a genuine overclaim of mine during the build.
It is not legal advice, it is unaffiliated with any employer, and the worked examples are fictional and labelled.
Next project
A crisis comms playbook, published alongside the adversary built to attack it — and the finding that saying nothing scores a perfect 1.00 on credibility.
Crisis Comms War-Game