# Map AI Advisor daily-routine fix — 2026-10-15

This records the initial three-question acceptance. See [expanded multilingual acceptance](GENERALIZATION_ACCEPTANCE.md) for the later 8-group/32-turn evaluation, additional failures found, fixes and current verification limits.

The Advisor now uses stated home/work locations to answer a rail-connectivity follow-up instead of asking for generic criteria again. It separates published station connectivity from unverified journey times. This is the Map Advisor at `/property/research?mode=advisor`.

## Behavior and evidence

The exact reported conversation was tested in one browser conversation on local Herd (`petav3.test`) with real OpenAI GPT-4.1 calls:

| Buyer message | Verified result |
| --- | --- |
| Help me find property locations that fit my daily routine. | One focused question for home and the daily destination. |
| i stay at sunway velocity, work in trx | Saves distinct home/work labels; answers with sourced station connections and displays reference markers; asks budget once. |
| got mrt to trx | Uses the same routine, answers MRT connectivity directly, retains source/map evidence, does not repeat the budget question. |

Sunway Velocity Mall lists Cochrane and Maluri access stations. The Kajang Line connections to TRX are respectively one and two stops, with no transfer. These are station references, not a confirmed home entrance. Sources and their checked dates are included in the result and source drawer.

A separate real-cloud seven-turn conversation also passed the three messages, a work correction to Bukit Bintang, a ten-minute commute question, a new unknown destination, and a correction to an unknown apartment. The work correction preserves home; the time request returns connectivity without claiming ten minutes; the unknown label receives a targeted clarification. Final measured application-turn durations were 0.92–4.11 seconds in that local run; this is not a production latency benchmark.

The browser recording includes the original three turns, the source drawer, and clicking Maluri to focus the map. It is a 40-second sequence of actual UI captures at approximately 2 fps, with captions; the three questions are continuous and idle gaps before the separate evidence/map demonstrations are removed. No simulated answer was inserted into the recording.

## Implementation boundaries

- Explicit EN/MS/ZH home/work statements are stored in `brief.routine`; corrections merge by field. Negations and hypotheticals do not silently change the brief. Labels never become confirmed coordinates.
- A bundled official-reference dataset supplies the 29 Kajang Line stations and three sourced map reference points (Cochrane, TRX, Maluri). It requires no Research snapshot, embeddings, Qdrant or additional network call. Unknown places are not evidence that a rail connection does not exist.
- The planner repairs a premature non-tool answer when the saved connection is known, but permits necessary clarification for unresolved labels. Known alternative tool envelopes are normalized, then all supplied arguments pass the existing strict validator. Model-authored `answer` prose is discarded; only verified claims reach the buyer.
- Rail-only results preserve property scope/selections. Combined property/rail results retain both evidence sets. Station previews do not confirm a home/work destination or invent track geometry.
- Existing account funding, Super Admin credit exemption and module-specific rate-limit rules are unchanged.

This is not complete public-transport routing: other lines/interchanges, live service, fares, waiting/walking times and door-to-door timing are not verified. No nearby project, current listing, price or walking-distance recommendation is invented. The wider property-data mapping and reviewed knowledge-package prerequisites in [DEPLOYMENT.md](DEPLOYMENT.md) still apply; this test does not complete the original 30-question/M09 evaluation or business sign-off.

## Regression and rollout

- Isolated local PHP suite: **168 tests / 1,706 assertions passed**; one existing PHPUnit XML configuration deprecation.
- PropertyLab frontend: **18 files / 96 tests passed**.
- Client and SSR production builds passed, with existing asset-resolution/bundle-size warnings.
- Changed-file Pint and all three PHP source guards passed. The learning-content guard fails for `how-to.json` and `how-to.en.json`; the identical failure was reproduced from an unmodified master archive at `9808c526`, so those unrelated generated files are unchanged.
- GitHub Actions for PR #210 did not start jobs because of account payment/spending limits; the remote run is not a passing CI result.
- Backend and TypeScript reviews completed; identified routine-parsing and transit-composition issues were corrected.
- Live model acceptance is English. EN/MS/ZH parsing and deterministic rail claims have automated coverage; a complete live multilingual evaluation is not claimed.

Local evidence is in the task artifact directory `map-advisor-daily-routine-20261015`: browser conversation/result proof, real cloud turn records, sanitized model-plan trace, test logs, screenshots and MP4. Credentials and test conversations are not committed.

This patch adds no migration, dependency or environment variable. An already-enabled site needs its existing reviewed master deployment, including frontend builds and the normal PHP/SSR cache/service refresh. A PR merge alone does not prove that `app` or `wk` is running the new version. First-time activation continues to use [DEPLOYMENT.md](DEPLOYMENT.md). No production deployment was performed for this acceptance.
