TEACHAID MAIL — LIVE AI FEEDBACK SMOKE CHECK
Endpoint: https://teachaid.ph/teachaid-api/pisa-multilingual-feedback.php
Model returned by every response: gemma3:4b
UTC window: 2026-10-29 08:52:33 to 08:52:50
Philippine time: 2026-10-29 16:52:33 to 16:52:50 (UTC+08:00)
MAIN FINDING
All 12 requests returned HTTP 200 with nonempty, parseable feedback.
All 12 passed the existing limited client mismatch screen.
THIS DOES NOT ESTABLISH RELIABLE FEEDBACK IN ALL 12 REQUESTED LANGUAGES.
Language identity and fluency are not verified. Subject accuracy is not
certified. Manual inspection found material concerns described below.
The server itself returned language_verified:false for every sample.
METHOD
One fictional, non-personal MAIL question was sent once for each canonical
language tag, at most two requests concurrently, with a 54-second deadline
and no client retries. The approved display notice said Friday in Room 2;
the generated reminder incorrectly said Thursday in Room 2. The fictional
answer correctly chose to change Thursday to Friday and retain Room 2.
The exact embedded client request builder, content parser, content screen
and response-language validator were used without modifying them.
No staging upload, numeric scoring request, learner record or report
submission was performed. Responses used one server generation attempt each.
TRANSPORT AND CLIENT-SCREEN RESULTS
Language Tag HTTP Time (s) Client accepted
English en 200 2.897 Yes
Filipino fil 200 3.529 Yes
Ilocano ilo 200 2.948 Yes
Pangasinan pag 200 2.444 Yes
Kapampangan pam 200 2.150 Yes
Bikol bik 200 2.175 Yes
Waray war 200 2.853 Yes
Cebuano ceb 200 2.666 Yes
Hiligaynon hil 200 2.380 Yes
Maranao mrw 200 2.163 Yes
Maguindanaon mdh 200 2.466 Yes
Tausug tsg 200 2.696 Yes
Actual POST responses allowed Origin:null. This command-line test did not
itself enforce browser CORS or exercise durable-save acknowledgements.
The requested_language field echoes the request; it is not a detected or
verified output language.
EXACT SAMPLED CONCERNS
These are manual observations, not a new language detector or a native-
speaker certification. Full responses are retained in the JSON evidence.
Bikol (bik): “Mahusay ka! Tamang sagot ang nakuha mo”
The response continues in recognisable Filipino/Tagalog phrasing. The
existing client screen accepted it despite the Bikol request.
Maranao (mrw): “Hindi sapat ang pagiging “confident” ng generated sentence”
The full response uses recognisable Filipino/Tagalog wording under a
Maranao request. Its language echo does not resolve this mismatch.
Ilocano (ilo): “Daghang saludo sa iyong sagot!”
The sample mixes recognisable Filipino and Cebuano-like wording in an
Ilocano-requested response. Qualified-speaker review is needed.
Pangasinan (pag): “Dagiti letra A kalpasan nga nagdagdat ti Friday”
The text contains Ilocano-like forms; it does not demonstrate a reliable
Pangasinan response. This observation is not comprehensive identification.
Cebuano (ceb): “Ang generated reminder nagpakita nga Friday sa Room 2”
This appears to describe the original generated reminder as Friday, while
the fixed case says Thursday. If the corrected reminder was intended, the
response does not clearly distinguish it from the original faulty draft.
Maguindanaon (mdh): “ang generated reminder iya usab nga Friday sa Room 2”
The same original-versus-correction concern appears, alongside mixed
wording that does not establish Maguindanaon identity.
VERIFICATION LIMITS
The unlisted language samples are also not certified. One short sample per
language cannot establish consistency, dialect suitability, fluency,
age-appropriate wording, or tutoring accuracy across the collection.
Passing the narrow screen means that its rejection rules did not fire;
it does not prove that the requested regional language was produced.
The twelve language choices are wired correctly and live transport is
confirmed. Reliable twelve-language tutoring remains unestablished with
this gemma3:4b deployment. Model improvement and qualified-speaker
assessment are separate remaining work; no server/runtime change was made.
EVIDENCE
live_feedback_smoke.json contains complete synthetic payloads, raw
responses, headers, timings, errors, model metadata and client results.
It retains the two initial restricted-environment failures, which occurred
before endpoint access because the local proxy could not be reached.
Approved normal network access then produced the twelve server responses.