Research: what the source document actually specifies

Question: What does chemical-safety-assisant-docs.pdf commit the product to?

Source: chemical-safety-assisant-docs.pdf — one page, a single 1024×1536 Flate+DCT image. It is a Thai-language concept infographic / pitch poster, not a requirements document. Every citation below is p.1 plus the panel it appears in.

Extraction method (reproducible, no poppler): the page is one embedded JPEG under a [/FlateDecode /DCTDecode] filter chain — zlib.decompress the stream between >> stream and endstream, and the result is a valid JPEG. Text-layer tools return binary garbage; do not retry them.

Standing caveat on every finding below. A poster states intent, not requirements. No finding here carries an acceptance threshold, a data schema, a regulatory basis, or a non-functional target, because the source contains none. Confidence values refer to what the poster says, never to whether it is the right thing to build.

Findings

R-01 The product is a mobile AI assistant for chemical safety information

"CHEMICAL SAFETY ASSISTANT / AI ผู้ช่วยอัจฉริยะด้านความปลอดภัยสารเคมี" — an AI application giving access to chemical safety information, promising three qualities: รวดเร็ว (fast) · แม่นยำ (accurate) · ปลอดภัย (safe).

Source: p.1, title block · Confidence: high

R-02 The stated problem is speed of access, not absence of data

ที่มาและความสำคัญ: factories and laboratories use many chemicals with differing hazards and handling methods; when workers cannot access safety information quickly, accidents follow, harming life, property and the environment. The remedy proposed is consolidating SDS data into one searchable system with QR scanning for immediate retrieval.

The bottleneck named is retrieval latency under field conditions, not data that does not exist. That framing should survive into the spec — it decides whether the product is a database or an access layer over one.

Source: p.1, ที่มาและความสำคัญ panel · Confidence: high

R-03 Four deployment settings are named

Superseded by: R-22, R-23 — the revised poster split this finding in two: R-22 narrowed the setting to one plant, R-23 recast the four venues as four job roles. The text below is left as the record of what the first poster said; do not cite it as a live upstream.

เหมาะสำหรับ: โรงงานอุตสาหกรรม (industrial factories) · ห้องปฏิบัติการ (laboratories) · คลังเก็บสารเคมี (chemical storage warehouses) · สถานศึกษาและหน่วยงานต่าง ๆ (educational institutions and other agencies).

Source: p.1, เหมาะสำหรับ panel · Confidence: high

R-04 Five objectives, two of which are research, not build

Superseded by: R-24 — three objectives, with the near-duplicate build verbs gone. R-24 also deleted the objective formerly numbered R-04.2, "study the benefits of using AI"; ADR-0018's D-33 lists the five ADRs whose reasoning cited it. Do not cite R-04.2.

วัตถุประสงค์, numbered in the source:

1. Consolidate chemical safety data into a system that is searchable and quick to access.
2. Study the benefits of using AI to deliver hazard information, PPE selection and basic first aid.
3. Study the problems and needs of workers accessing chemical safety information.
4. Develop an AI application for retrieving chemical safety information from SDS documents.
5. Develop an AI application for retrieving and presenting information from SDS documents.

Objectives 2 and 3 are study objectives — this reads as an academic or capstone project framing, not a purely commercial one. Objectives 4 and 5 are near-duplicates.

Source: p.1, วัตถุประสงค์ panel · Confidence: high for the text; the duplication of 4/5 is observed, not inferred

R-05 Scope is five capabilities

ขอบเขต:

1. Search and display chemical data from an SDS database.
2. Summarise safety information — hazards, storage method, first aid.
3. Recommend PPE appropriate to each chemical.
4. Give emergency-management guidance, e.g. chemical spills and basic first aid.
5. Answer chemical-safety questions via an AI chatbot.

Source: p.1, ขอบเขต panel · Confidence: high

R-06 SDS is the authoritative data source, and it is presupposed

Every retrieval capability is described as reading from SDS documents or an "ฐานข้อมูล SDS Cloud Database". The poster never says where that corpus comes from, who maintains it, or under whose authority its content is correct.

Source: p.1, ขอบเขต and เทคโนโลยีที่ใช้ panels · Confidence: high that SDS is the source; the provenance gap is an absence, tracked as an open question

R-07 Eight screens are numbered; only seven are shown

ระบบการทำงานของแอปพลิเคชัน numbers its screens 1–8, but the strip renders 1, 2, 3, 4, 5, 6, 8. There is no panel 7. Given screen 2's menu offers a การปฐมพยาบาล (First Aid) tile that no rendered screen satisfies, First Aid is the likely missing screen — but that is inference, not source.

Source: p.1, ระบบการทำงานของแอปพลิเคชัน strip · Confidence: high that 7 is absent; low as to what it was

R-08 Login is optional — anonymous use is explicitly supported

Screen 1 offers เข้าสู่ระบบ (log in) and สมัครสมาชิก (register), with the line ใช้งานได้โดยไม่ต้องเข้าสู่ระบบ — "usable without logging in". The same line appears on the phone mockup in the title block, so it is stated twice.

This is load-bearing for the whole design: in an emergency nobody logs in. It constrains what may sit behind an account, and it means search history (R-13) must work for an anonymous user or be understood as a signed-in-only feature.

Source: p.1, screen 1 and title-block mockup · Confidence: high

R-09 The home screen defines six functional entry points

Screen 2 (หน้าแรก) has a search field plus six tiles: สแกนสารเคมี (Scan Chemical) · ถาม AI (Ask AI) · SDS สรุปย่อ (SDS Summary) · แนะนำ PPE (PPE Recommendation) · การรับมือสารเคมีรั่วไหล (Spill Response) · การปฐมพยาบาล (First Aid). A bottom navigation bar carries four items including หน้าหลัก (home) and ประวัติ (history).

These six tiles, not the five scope bullets, are the closest thing the source has to a feature list.

Source: p.1, screen 2 · Confidence: high for the tiles; the bottom-bar labels are partly illegible at this resolution

R-10 Identification is by scanning a physical container label

Screen 3 (หน้าสแกนสารเคมี) shows a camera viewfinder framing a bottle labelled ACETONE (CH₃)₂CO with a DANGER banner and a barcode, with ยกเลิก / สแกน buttons. The technology panel names a QR Code Scanner, but the object in the viewfinder is a supplier label bearing a barcode.

Whether the product scans QR codes it generates itself, existing barcodes, or reads the label text visually is not settled by the source, and the poster's two halves point at different answers.

Source: p.1, screen 3 + เทคโนโลยีที่ใช้ panel · Confidence: high that scanning is the primary identification route; the code type is contradictory

R-11 The chemical record shape, as exemplified by Acetone

Screen 4 (หน้าสรุปข้อมูลสารเคมี) is the richest panel and the best evidence of the data model:

• Identity: Acetone, formula (CH₃)₂CO, CAS No. 67-64-1
• Three tabs: ข้อมูลทั่วไป (general) · SDS · การใช้งาน (usage)
• ความเป็นอันตราย (hazard): ไวไฟสูง (Highly Flammable), flagged in red
• Three GHS-style hazard pictogram diamonds
• อันตรายต่อสุขภาพ (health hazards) as a bullet list — eye and skin irritation; inhalation may cause dizziness
• การป้องกัน (protection) as a row of PPE icons
• แนวทางการจัดเก็บ (storage guidance) — dry, ventilated, away from heat and sparks

CAS number is the only stable external identifier the poster shows, which makes it the natural primary key.

Source: p.1, screen 4 · Confidence: high for the fields shown; the tab contents beyond ข้อมูลทั่วไป are not rendered

R-12 PPE advice is per-chemical, itemised, and carries a reason per item

Screen 5 lists PPE "ที่แนะนำสำหรับ Acetone": ถุงมือไนไตรล์ (Nitrile Gloves) · แว่นนิรภัย (Safety Goggles) · หน้ากากกรองไอระเหย (vapour-filter Respirator) · เสื้อกาวน์ (Lab Coat) · รองเท้านิรภัย (Safety Shoes). Each carries a one-line rationale beneath it. A closing note instructs the user to check PPE condition before every use.

Two things follow: the recommendation is specific to the chemical (nitrile is named, not "gloves"), and each line is justified, not merely listed.

Source: p.1, screen 5 · Confidence: high

R-13 Spill response is an ordered procedure, not reference text

Screen 6 (การรับมือสารเคมีรั่วไหล) presents "แนวทางการจัดการ Acetone หกรั่วไหล" as five numbered steps: put on full PPE → eliminate ignition sources and ventilate → absorb with an absorbent material such as sand → collect waste into a closed, labelled container → clean the area. A red warning box follows, then an emergency hotline row.

Ordering is semantic here. Ventilating before donning PPE is a different instruction from the one given, so this content cannot be stored as an unordered bag of tips.

Source: p.1, screen 6 · Confidence: high for the sequence; step 5's wording is partly illegible

R-14 Emergency hotline 1669 is surfaced in the spill flow

Screen 6 ends with สายด่วนฉุกเฉิน 1669 — Thailand's national EMS number. The product therefore has a hand-off point to real emergency services inside a hazard flow.

Source: p.1, screen 6 · Confidence: high

R-15 Search history is per-item and timestamped

Screen 8 (หน้าประวัติการค้นหา) lists Acetone · Toluene · Hydrochloric Acid · Sodium Hydroxide · Xylene, each with a date and time (e.g. 20 พ.ค. 2567 10:30). พ.ศ. 2567 is Buddhist Era — CE 2024. Dates in this product are Thai-locale by default.

Source: p.1, screen 8 · Confidence: high

R-16 Five chemicals appear as the worked corpus

Acetone (worked in full across screens 3–6) plus Toluene, Hydrochloric Acid, Sodium Hydroxide and Xylene in history. These are the only chemicals the source names, and they span flammable solvents, a strong acid and a strong base.

Source: p.1, screens 3–6 and 8 · Confidence: high that these are named; nothing in the source says the corpus is limited to them

R-17 Four technologies are named

เทคโนโลยีที่ใช้: AI Chatbot (ประมวลผลภาษาอัตโนมัติ — automatic language processing) · QR Code Scanner · ฐานข้อมูล SDS Cloud Database · Mobile App (ใช้งานง่ายทุกที่).

Named as categories, with no vendor, model, framework or hosting commitment. Nothing here constrains implementation, and nothing here should be read as an architectural decision already made.

Source: p.1, เทคโนโลยีที่ใช้ panel · Confidence: high for the list; it carries no technical commitment

R-18 The interface language is Thai

Every rendered screen is in Thai, with English retained for chemical names, CAS numbers, formulae and PPE terms. No language switcher appears anywhere.

Thai is therefore the default and only evidenced locale, and chemical identity stays in English inside a Thai interface.

Source: p.1, all screens · Confidence: high for Thai-first; the absence of a switcher is an absence, not a stated decision

R-19 Five benefits are claimed, and they are the implicit success criteria

Amended by: R-33 — the benefit list was replaced. What survives, and is still correctly cited by ADR-0001 and ADR-0017, is that the claims carry no thresholds. An amendment, not a supersession: R-19 remains live and citable for that half.

ประโยชน์ที่ได้รับ: fast access to information · reduced risk and accident prevention · correct, standardised information · AI answering questions at any time · reduced environmental impact.

None is quantified. If the spec needs measurable outcomes, they are derived from these five and confirmed with the owner — they cannot be read off the source.

Source: p.1, ประโยชน์ที่ได้รับ panel · Confidence: high for the claims; zero for any threshold

R-20 A QR code on the poster points to further material

The bottom-left carries a "SCAN ME" QR block labelled สแกนเพื่อดูข้อมูลเพิ่มเติม — "scan to see more information". It has not been decoded or fetched: it is an external destination and resolving it needs the owner's say-so.

If it resolves to a fuller specification, that document supersedes this poster as the stage-0 source and this note should be re-run against it.

Source: p.1, footer · Confidence: high that it exists; unknown as to target

Open questions

Not findings — nothing here is settled, so nothing here gets an R- id. Each is a candidate for a /grill-with-docs question or a /wayfinder ticket.

1. Where does the SDS corpus come from? Licensed feed, supplier PDFs, manual entry, or scraped? This decides the legal position, the ingest pipeline and the update cadence. It is the single largest unknown in the project (R-06).
2. Who is accountable when the app's advice is wrong? The product tells people what PPE to wear and how to handle a spill. Liability, disclaimers, and whether output is "information" or "instruction" are unresolved, and safety-critical.
3. Does the AI generate the safety content, or only retrieve and summarise it? R-05 says summarise; R-01 says AI assistant. A model that paraphrases an SDS can paraphrase it wrongly. This is the highest-risk ambiguity in the source.
4. What is actually scanned — a self-generated QR code, an existing supplier barcode, or the label text (R-10)? The source contradicts itself.
5. What was screen 7 (R-07), and is First Aid a screen the build owes?
6. Does the product work offline? A warehouse, a plant floor and a spill are all places with poor connectivity, yet the data is described as a cloud database (R-17). Never mentioned in the source.
7. Is search history anonymous-capable (R-08 vs R-15), and does it sync across devices?
8. How large is the chemical corpus at launch — the five named (R-16), or a full catalogue?
9. Is this a real product or a capstone deliverable? Two of five objectives are study objectives (R-04). The answer changes the definition of done for the whole effort.
10. Where does the poster's QR code lead (R-20), and is there a fuller source document that should replace this one?

Coverage

Stage 0 is the head of the pipeline — it has no upstream artifact to account for. This table is intentionally empty; the first real Coverage table is written by stage 1, which must account for every R-nn above.

Upstream	Landed in	Evidence	Note