Kiarash Taherikiatah.comAll work ↗
03 / Language and evidence / recorded experiment

Reading a
regulation.

One recorded answer, checked against the page the prototype actually retrieved.

01 / QUESTION IN THE RECORDING

مشکلات نداشتن هوابندی چیست؟

What problems arise from a lack of airtightness?
02 / RETRIEVED PDF · PAGE 91

عدم هوابندی مناسب باعث هدررفت قابل توجه هوای تهویه شده داخلی و در نتیجه، افزایش مصرف انرژی ساختمان می‌شود.

Conditioned air lost → energy use rises
03 / CHECK THE ANSWER AGAINST THIS PAGE
✓ Air loss
✓ Energy use
? Air quality
? Health
Retained Persian retrieval prototype / inspected source: Mabhas19.pdf, viewer page 91
01 / Recorded exchange

What did it answer?

Crop from the original recording showing the Persian question, generated answer and collapsed source panel
Retained interface at 00:31. Watch the complete exchange ↓
Question · English translation

“What problems arise from a lack of airtightness?”

مشکلات نداشتن هوابندی چیست
Prototype answer · English translation

Air can enter or leave the building. The response also names poorer indoor air quality, increased energy use and health problems.

Generated response in the recording; checked against its retrieved page below.
02 / Source check

One page. Four claims.

The retrieved PDF supports two parts of the answer. The remaining claims need other evidence.

RETRIEVED SOURCE 01 / VIEWER PAGE 91 / PRINTED PAGE 72
Actual retrieved PDF crop from Section 19, viewer page 91, with the airtightness sentence marked
Actual retrieved PDF / the outline locates the sentence examined at right.
Printed Persian / normalized transcription

عدم هوابندی مناسب باعث هدررفت قابل توجه هوای تهویه شده داخلی و در نتیجه، افزایش مصرف انرژی ساختمان می‌شود.

English translation of that sentence

Inadequate airtightness causes a significant loss of conditioned indoor air and, consequently, increases a building’s energy use.

01 / STATEDConditioned air is lost
02 / STATEDEnergy use increases
03 / TERM ONLYInfiltration / exfiltrationTerms appear in footnotes; the answer’s elaboration is not checked by this sentence.
04 / NO SUPPORT HEREAir quality and healthThese generated claims need other evidence.
Recording frame showing Source 1 labelled Page 91 and damaged Persian extraction
The prototype’s source panel at 01:05. Its extracted Persian has direction and line-break damage.
What the recording returns
  1. Source 1 · page 91Selected PDF page; checked above.
  2. Source 2 · page 76Label visible in the recording; content not checked here.
  3. Source 3Returned in the recording; page label not reliably legible.
  4. Source 4 · page 50Label visible in the recording; content not checked here.
Original 90-second record

Watch the source check.

The subtitle option translates legible Persian shown in the recording. The timed notes beneath the video identify interface and source-page changes.

This subtitle track covers legible on-screen Persian and the inspected PDF sentence. It does not claim to transcribe all speech or the damaged source-panel extraction. The notes below the video are a separate viewing guide.

03 / Saved method

From page to answer.

The script describes how pages become searchable chunks. The recording shows why a returned page still needs a claim-by-claim check.

SAVED SCRIPT / DOCUMENT PREPARATIONOpen exact Python source ↗
01 / LOAD + REPAIRMabhas19.pdfPyMuPDF pages → Persian reshaping and bidirectional display repair
02 / SPLITOverlapping passagesRecursive splitter · 600 characters · 100 character overlap
03 / INDEXMultilingual E5 → FAISSintfloat/multilingual-e5-large · normalized CPU embeddings
RECORDED QUESTION / RETURNED CONTEXT / ANSWEROne observed exchange
QUESTION

What problems arise from a lack of airtightness?

TOP FOUR CHUNKS · SCRIPT k=4
01 PDF p.9102 p.7603 page unclear04 p.50
Only Source 1’s printed page was checked against the answer here.
LOCAL MODEL → SOURCE 1 CHECKLlama 3.1:8BSaved prompt: answer from context; abstain if absent.
✓ Air loss✓ Energy use? Air quality? Health
Where the chain can fail

The returned source-panel text is visibly fragmented. The page ID identifies a candidate source; it cannot, by itself, validate the model’s added air-quality or health claims.

04 / Separate fine-tuning experiment

One passage. Two targets.

Records 12 and 13 are generated from the same retained introduction. The corpus contains 912 question–target pairs but only 206 distinct source contexts by exact hash.

RETAINED CONTEXT / RECORDS 12 + 13SHA-256 a873c254…28ab
Same Persian source in both records · selected exact excerpts

ساختمانها با سهم بیش از 40% از کل مصرف انرژی سالانه کشور، میتوانند نقش قابل توجهی در مدیریت انرژی به ویژه در دوره اوج بار شبکه برق و گاز ایفا نمایند.

این قانون، افزایش بهرهوری انرژی در ساختمانها را از اهداف خود معرفی کرده است.

تلاش شده است مصرف انرژی در ساختمانهای کشور مدیریت و مطابق معیارهای مشخص، ساماندهی شود.

English reading The introduction attributes over 40% of annual national energy use to buildings. It also describes improved building energy efficiency and managing energy use against specified criteria as aims. The figure is a claim in this retained document, not a verified current statistic.

The displayed Persian preserves the extraction’s spelling and spacing. The full shared context is available in the records below.
12 / PURPOSE

هدف از مقررات مبحث نوزدهم چیست؟

What is the purpose of the Section 19 regulations?

GENERATED TARGET

افزایش بهره‌وری انرژی در ساختمان‌ها و ساماندهی مصرف انرژی مطابق معیارهای مشخص در مبحث نوزدهم.

Improving building energy efficiency and organising use against specified criteria.
13 / SHARE

سهم ساختمان‌ها از مصرف انرژی کل کشور چقدر است؟

What share of national energy use does the document attribute to buildings?

GENERATED TARGET

بیش از 40 درصد از کل مصرف انرژی سالانه کشور.

More than 40% of annual national energy use, according to this text.
QUESTION-LEVEL SPLIT · RISK
TRAIN context A / 12EVALUATE context A / 13
CONTEXT-LEVEL SPLIT · KEEP THE PASSAGE TOGETHER
TRAIN context A / 12 + 13EVALUATE different context

These are split rules illustrated with the retained pair, not a recovered record of how the saved run actually split its data.

912generated pairs206distinct contexts

Select a question.

The records above share an exact context hash. A split by question could put that same passage in both training and evaluation. The generated targets have not been independently judged. Inspect all retained messages ↗

05 / Training record

A saved run.

Three captures document a Qwen3-8B LoRA training run. Loss observations do not establish answer quality; adapter weights and an independently judged answer set were not recovered.

Retained training run 1 history and summary screenshot
Run 1 / retained capture
Retained training run 2 history and summary screenshot
Run 2 / retained capture
Retained training run 3 history and summary screenshot
Run 3 / retained capture