Appearance
Research: the revised poster (poster.pdf)
Question: What does poster.pdf change about the project, and which stage-0 findings and stage-1 decisions does it affect?
Source: poster.pdf — one page, one 1024×1536 Flate+DCT image, extracted by the same method as the first poster (see source-poster.md). Thai-language infographic.
Relationship to the first source. This is a later revision, not a second opinion. It is titled differently ("AI CHEMICAL SAFETY ASSISTANT"), restructured into seven numbered panels, and it tightens the earlier poster in every place they disagree — the objective list is shorter and de-duplicated, the audience is narrowed from settings to named roles, and two whole panels are new. Where the two conflict, this one wins and the earlier finding is marked superseded below.
Findings continue the R- sequence from source-poster.md; ids there are never reused.
Findings
R-21 This is explicitly a prototype, and the source now says so
The subtitle reads ต้นแบบแอปพลิเคชัน AI — "AI application prototype". Panel 7's final expected result is ได้ต้นแบบระบบที่สามารถพัฒนาต่อยอดให้ใช้งานจริงในอนาคต — "a system prototype that can be developed further for real use in the future".
D-01 was decided by inference from study verbs. The source now states it outright, in the same shape that was chosen: a prototype with a path to production, not a throwaway demo.
Source: p.1, subtitle and panel 7 · Confidence: high Effect on D-01: confirms it. No change needed.
R-22 The scope is one case study — an automotive parts manufacturing plant
A banner under the title reads กรณีศึกษา: โรงงานผลิตชิ้นส่วนยานยนต์ — "case study: automotive parts manufacturing plant". Panel 5 repeats it as the first scope bullet.
Source: p.1, title banner and panel 5 · Confidence: high Supersedes R-03, which listed four settings (factories, laboratories, chemical warehouses, educational institutions). The target is now one plant type. Panel 5's closing note preserves the generalisation as future work: the concept can be extended to other industry types by adapting the chemical database per plant context.
R-23 The audience is four named job roles, not four venues
Panel 3 กลุ่มเป้าหมาย names:
- เจ้าหน้าที่ความปลอดภัยในการทำงาน (จป.) — the statutory Thai workplace Safety Officer role
- หัวหน้างาน — supervisor
- เจ้าหน้าที่คลังสารเคมี — chemical store officer
- ผู้ปฏิบัติงานที่เกี่ยวข้องกับสารเคมี — workers who handle chemicals
plus a qualifier: ผู้ที่มีสมาร์ทโฟนหรือแท็บเล็ตในการใช้งานแอป — users who have a smartphone or tablet.
This is a materially better audience definition than R-03's venues: จป. is a regulated role with statutory duties, and a supervisor's needs differ from a line worker's. It also implies the app has more than one kind of user, which nothing in the first poster did.
Source: p.1, panel 3 · Confidence: high Supersedes R-03.
R-24 Three objectives, and the duplication is gone
Panel 2 วัตถุประสงค์:
- Consolidate chemical safety data into a system that is searchable and quick to access.
- Study the problems and needs of workers accessing chemical safety information.
- Develop an AI application for retrieving chemical safety information from SDS documents.
Source: p.1, panel 2 · Confidence: high Supersedes R-04, which had five objectives with two near-duplicate build verbs (4 and 5) and a second study objective about AI benefits. The duplication R-04 flagged as observed has been fixed in the revision, and one study objective has been dropped.
Effect on D-01: the objective that ADR-0001 leaned on — "study the benefits of using AI" — no longer exists. The surviving study objective is about worker needs, not about AI efficacy. ADR-0001's reasoning cited the deleted objective, so its rationale is now partly stale even though its conclusion is confirmed by R-21.
R-25 The named chemicals are the plant's actual process chemicals
Panels 1 and 5 both name Acetone, IPA (isopropyl alcohol), Toluene, น้ำยาล้างชิ้นงาน (parts-cleaning solution) and "others", described as สารเคมีที่ใช้จริงในกระบวนการผลิต — chemicals actually used in the production process.
IPA and parts-cleaning solution are new. Parts-cleaning solution is notable because it is a formulated mixture, not a pure substance — it has a product name and a supplier, not a CAS number.
Source: p.1, panels 1 and 5 · Confidence: high Extends R-16. Combined with the history list (Acetone, Toluene, HCl, NaOH, Xylene) the known corpus is at least eight entries across pure substances and at least one mixture.
R-26 The corpus is the plant's own SDS, integrated with the plant's system
Panel 6's comparison table states the product ใช้ข้อมูลจากฐานข้อมูลสารเคมีและ SDS ของโรงงาน — "uses data from the chemical database and the plant's SDS" — against general apps that อ้างอิงข้อมูลทั่วไปจากอินเทอร์เน็ต ("reference general information from the internet"). A further row claims เชื่อมโยงกับฐานข้อมูลสารเคมีของโรงงานโดยตรง — "links directly to the plant's chemical database" — where general apps ไม่เชื่อมโยงกับระบบโรงงาน.
Panel 5 adds ข้อมูลอ้างอิงจากเอกสาร SDS ฉบับจริง และมาตรฐานที่เกี่ยวข้อง — referenced from actual SDS documents and relevant standards.
Source: p.1, panels 5 and 6 · Confidence: high for the claims Conflicts with D-04, which chose a curated set of manufacturer-published SDS PDFs gathered by the project. The source is describing the plant's own SDS holdings and a direct integration with a plant system. These are different corpora with different provenance, different licensing positions and different ingest work. Unresolved — raised to the owner.
R-27 The source claims AI summarisation of SDS content
Panel 6: AI สรุปข้อมูลสำคัญจาก SDS ให้เข้าใจง่ายและนำไปใช้ได้จริง — "AI summarises the important information from the SDS so it is easy to understand and genuinely usable" — positioned against general apps whose ข้อมูลยาวอ่านเข้าใจยาก ("long, hard-to-read information"). Panel 1 likewise promises AI ช่วยสรุปและตอบคำถามได้อย่างรวดเร็ว ถูกต้อง.
Source: p.1, panels 1 and 6 · Confidence: high Conflicts with D-02/D-03, which confined safety-critical content to extractive spans with no paraphrase. "Summarise so it is easy to understand" is generative paraphrase of exactly the content D-02 protects, and the poster makes it a headline differentiator rather than an incidental feature. Unresolved — raised to the owner.
R-28 Offline operation is now an explicit differentiator
Panel 6's final row: general apps ต้องใช้อินเทอร์เน็ต ("require the internet"); this product รองรับการใช้งาน… ("supports [offline] use", the qualifier partly illegible at this resolution).
Source: p.1, panel 6 · Confidence: high that offline is claimed as a differentiator; medium on the exact wording of the qualifier Effect on D-06: confirms it, and promotes it from an inference about plant-floor signal to a stated product claim.
R-29 The scan target is "label or QR code"
Screen 3's viewfinder caption now reads วางฉลากหรือ QR Code ให้อยู่ในกรอบ — "place the label or QR code inside the frame". The first poster's caption did not say this.
Source: p.1, panel 4, screen 3 · Confidence: high Effect on D-05: confirms it directly. The revision resolves the R-10 contradiction the same way the grill did — label reading primary, QR supported — so D-05 now rests on the source rather than on reasoning about it.
R-30 First Aid is a top-level tile in the phone mockup
The title-block phone renders six tiles: สแกนสารเคมี (Scan Chemical) · ค้นหา SDS (Search / SDS) · แนะนำ PPE (PPE Recommendation) · การปฐมพยาบาล (First Aid) · การจัดการหกรั่วไหล (Spill Response) · ถาม AI (AI Chat).
Source: p.1, title-block phone mockup · Confidence: high Effect on D-10: confirms First Aid is a first-class destination. Note the tile set differs from the first poster's home screen (R-09): "SDS สรุปย่อ" has become "ค้นหา SDS (Search / SDS)", which reads as search rather than summary.
R-31 Screen 7 is still missing
Panel 4 again numbers screens 1–8 and again renders 1, 2, 3, 4, 5, 6, 8. Two independent revisions of the poster have the same hole.
Source: p.1, panel 4 · Confidence: high Effect on R-07 and D-10: the gap is reproduced, not corrected. Since First Aid now appears as its own tile (R-30) while still having no numbered panel, D-10's reconstruction looks right, but the source still never shows the screen.
R-32 The problem statement is sharper and names concrete failure modes
Panel 1 หลักการและเหตุผล: an automotive parts plant uses many chemicals; if safety information is slow to reach, the consequences named are การเลือกใช้ PPE ไม่เหมาะสม (choosing unsuitable PPE), การจัดเก็บสารเคมีไม่ถูกต้อง (storing chemicals incorrectly) and การตอบโต้เหตุฉุกเฉินล่าช้า (delayed emergency response). It adds that SDS data today มีอยู่แต่กระจัดกระจาย ค้นหายาก และเข้าใจยาก — "exists but is scattered, hard to search and hard to understand".
Source: p.1, panel 1 · Confidence: high Sharpens R-02. The bottleneck is now three named failure modes rather than generic "accidents", and the diagnosis is explicitly scattered and hard to understand, not absent. "Hard to understand" is the phrase that motivates R-27's summarisation claim.
R-33 Six expected results, still unquantified
Panel 7 ผลที่คาดว่าจะได้รับ: reduced SDS lookup time · increased workplace safety and reduced chemical risk · better PPE and procedure selection · improved emergency response effectiveness · raised standards of industrial chemical management · a prototype that can be developed for real use.
Source: p.1, panel 7 · Confidence: high for the claims; zero for any threshold Supersedes R-19's list. Still no numbers, so measurable outcomes remain something to agree with the owner rather than read off the source.
R-34 The revised poster carries no QR code
The "SCAN ME" block present in the first poster is absent here.
Source: p.1, whole page · Confidence: high Effect on D-13: supports it. The later revision dropped the QR entirely, consistent with it having been decorative artwork rather than a pointer to a real document.
Open questions
Not findings. Each needs an owner decision or investigation.
- Whose SDS is the corpus — the plant's or the manufacturers'? (R-26 vs D-04.) Decides provenance, licensing, ingest work, and whether a plant integration exists at all. This is the largest open item in the project.
- Does "AI summarises the SDS" override the no-paraphrase rule? (R-27 vs D-02/D-03.) The source makes summarisation a headline differentiator; the grill ruled it out for safety-critical content. One of the two must give.
- Does the app have per-role behaviour? (R-23.) A จป. with statutory duties, a supervisor and a line worker are different users. The source names them but shows one undifferentiated interface.
- How is a formulated mixture handled? (R-25.) น้ำยาล้างชิ้นงาน has a product name and a supplier but no CAS number, and D-05 makes CAS the identification key and R-11 makes it the primary key.
- Is there a real plant, and is it reachable? (R-22, R-26.) A case study implies a specific site whose SDS and staff are accessible. Objective 2 in R-24 — studying worker needs — cannot be done without one.
- What replaced "SDS สรุปย่อ" with "ค้นหา SDS"? (R-30 vs R-09.) Summary versus search is a real difference in what the tile does, and it points the opposite way to R-27.
Coverage
Stage 0 has no upstream artifact. Supersession of earlier findings is recorded inline above, not here — a superseded finding still has to land in a downstream decision, and stage 1 remains accountable for it.
| Upstream | Landed in | Evidence | Note |
|---|