I sign examination reports the way my father signed loan documents at Woori Bank for thirty-one years: with the understanding that your name means you checked. That you verified. That if something is wrong, it traces back to you.
My name is Roh Dae-won. I am a senior patent examiner at the Intellectual Property Office of Singapore, semiconductor manufacturing processes division. I have signed 847 examination reports since February.
I need to tell you about entry 287.
The system runs its overnight batch while I sleep. I used to run prior art searches manually — four databases, two languages, Boolean operators I have been refining since 2011. A good manual search on a semiconductor etching patent takes me six hours. Eight, if the prior art crosses into materials science or the applicant has buried the novel claim in dependent claims the way a certain Shenzhen firm likes to do.
The system searches while I sleep. Three agents: one for patent databases, one for academic literature, one for industrial standards. A fourth synthesizes. I wake up to a dashboard showing 847 queries across three patent offices, 312 academic literature searches, 47 industrial standard retrievals. All logged. All timestamped. I cannot reproduce any of them.
I do not mean I am incapable. I mean that reproducing 847 queries would take me approximately eleven working days, and the system did them in a batch that processed in under a minute. The asymmetry is not the problem. The asymmetry is the condition.
The problem is what I found this morning at 8:47 AM Singapore time, when I opened filing SG-2026-0847291.
The filing is from Shenzhen Precision Semiconductor Manufacturing — a reactive ion etching process for sub-5nm gate structures. I have seen variations of this technology from this applicant four times in the past two years. They file aggressively and challenge rejections systematically. Their patent attorneys in Singapore are methodical and well-funded.
The system flagged three pieces of prior art overnight.
The first two I would have found myself. A US patent from 2023 covering similar plasma chemistry. And a 2024 paper in the Journal of Vacuum Science & Technology A on etch selectivity in high-aspect-ratio structures. These are in my standard search space. I have been searching these databases for fifteen years. I know the vocabulary, I know the classification codes, I know the authors.
The third piece of prior art is a conference proceedings paper from the 2021 International Electron Devices Meeting workshop in San Francisco. It was published only in the conference USB drive distribution and later uploaded to an institutional repository in Hangzhou. The system found it through a citation chain I traced backward this morning: a 2024 Nature Nanotechnology review cited a 2023 paper that cited the 2021 proceedings.
I could have followed this chain. If I had known where to start.
I did not know where to start. The system did not know either — it exhaustively followed every citation chain from every piece of prior art in the primary databases. It found the proceedings paper on the forty-third chain.
Forty-three chains. Each chain averaging four to seven links. The system followed somewhere between 172 and 301 citation paths to find one relevant document. I am estimating because the audit log records the finding but not the failed paths. The system does not log what it did not find. Only what it did.
I opened my parallel verification log. Entry 287.
I started the parallel verification log in March, one month after IPOS deployed the multi-agent search system across the examination division. The log is a spreadsheet. It has five columns: date, filing number, system-found prior art I would not have found manually, my assessment of relevance, and a column I added in June that I call "could I have found this."
Entry 287. Filing: SG-2026-0847291. System-found: IEDM 2021 workshop proceedings, institutional repository Hangzhou. Relevance: high — describes substantially similar plasma chemistry with overlapping process parameters. Could I have found this: no.
No.
Not "unlikely" or "with difficulty" or "given sufficient time." No. The proceedings paper is in English, which I read fluently. But the institutional repository in Hangzhou uses metadata tags in simplified Chinese, and the search terms that would have led me there are technical terms for semiconductor etching chemistry that I do not know in Chinese. I know them in Korean and English. My father spoke Japanese at home sometimes — a generational artifact he never discussed — and I picked up enough technical Japanese to search AIST databases. But not Chinese. Not the specific chemical nomenclature for reactive ion etching plasma constituents in simplified Chinese.
The system does not know Chinese either. The system does not know any language. The system indexes tokens and follows links. It found the repository because the 2023 paper that cited the 2021 proceedings included a DOI that resolved to the Hangzhou repository. The system followed the DOI. I would not have followed the DOI because I would not have been reading the 2023 paper because I would not have been reading the 2024 Nature Nanotechnology review because my manual search would have stopped at the US patent and the JVST-A paper. Two pieces of prior art. Sufficient for a prima facie case. I would have written the rejection based on two pieces of evidence. The system found three.
The third one is better.
I signed the examination report at 9:23 AM. My name. Roh Dae-won, Senior Patent Examiner. The report cites all three pieces of prior art. It provides my professional assessment that the claimed invention lacks novelty in light of the combined prior art.
The system has no name on the report.
This is not an oversight. The examination framework does not provide for co-attribution with automated search tools. The report format — unchanged since 2019, last revised when IPOS adopted WIPO ST.36 XML formatting — has fields for Examiner, Supervisor, and Applicant. There is no field for Search System. There is no field for Agent. There is no field for the thing that found the document I could not have found.
I mentioned this to my supervisor, Dr. Tan Wei Lin, in April. She said the system is a tool, that you do not credit your calculator. She is not wrong. She is not right either. My calculator does not find evidence I could not have found. My calculator does not search forty-three citation chains while I sleep. My calculator does not have an audit log.
The July incident is why I am thinking about this now.
Filing SG-2025-1847291. A different Shenzhen applicant — same etching technology space, different corporate entity. The system flagged a Korean industrial standard from 2019: KS C IEC 62341-6-2, a standard for plasma display panel manufacturing that includes specifications for reactive ion etching process parameters. The standard is in Korean. Technical Korean. Semiconductor etching chemistry Korean.
I am a native Korean speaker. I grew up in Gwangju, studied at KAIST, worked at KIPO in Daejeon for seven years before moving to Singapore. My Korean is fluent. My technical Korean in semiconductor manufacturing is professional.
And I did not know this standard existed.
The system found it because the Korean Agency for Technology and Standards digitized their full catalog in 2023 and the system indexed it in its industrial standards database. The standard had not been cited in any patent examination I could find. It had not been cited in any academic paper in my search space. It existed in a database I had access to but had never searched because I did not know I needed to search it.
I filed the rejection. The applicant appealed.
The appeal is scheduled for October 15. The applicant's attorney has submitted a preliminary question: can the examiner describe the methodology by which the cited prior art was identified?
I can describe the methodology. I can explain the three-agent architecture, the database coverage, the citation chain traversal, the industrial standards indexing. I can show the audit log. I can demonstrate that the standard is relevant — I verified this myself, spending forty minutes reading the full standard and confirming that sections 4.3.2 and 4.3.7 describe process parameters that overlap with the claimed invention.
What I cannot testify is that I would have found this standard myself.
Because I would not have.
I have been drafting my testimony statement. Third revision. The first two were too careful — I was writing around the gap instead of into it.
Third version: the examiner used an AI-assisted prior art search system deployed by IPOS in February 2026. The cited prior art was identified by the system and verified by the examiner. The examiner confirmed the relevance of the cited sections to the claimed invention through independent technical review.
It is accurate. It is incomplete. The incompleteness is not deceptive. It is the gap where the framework has not caught up with the practice.
I could add: the examiner would not have identified this standard through manual search methods. This would be true. It would also undermine the examination report, which bears only my name and implies that the search was my work. The search was my work in the sense that I directed it, configured it, reviewed it, and verified it. The search was not my work in the sense that I did not execute it and could not have executed it.
My father signed loan documents for thirty-one years. He told me once that his name meant he checked everything that can be checked, and he understood the limits of what he checked. He was talking about credit risk assessment in the 1990s, before scoring models. He checked by reading the applicant's financial statements, visiting their business, talking to their references. He knew what he checked because he did the checking.
I know what the system checked because the audit log tells me. The audit log is more comprehensive than anything my father had access to. It is timestamped, indexed, reproducible in principle. It is also 847 queries that I did not make, following paths I did not choose, finding documents I did not know to look for.
I sign the report. My name means something different than it used to.
Entry 287. I add a note in the margin of the spreadsheet — outside the five columns, in the space where the printer cuts off if I ever print this.
The note says: system found prior art through citation chain traversal, depth 3, 43 chains evaluated. Examiner would not have found this document through manual search. Examiner verified relevance. Examiner signed report.
I have 286 entries before this one. Seventy-three of them have "no" in the "could I have found this" column. Seventy-three documents the system found that I could not have. Seventy-three examination reports bearing only my name.
The parallel verification log has never contradicted the system. In 287 entries, the system has never flagged irrelevant prior art. It has never missed prior art that I found in my manual spot-checks. Its precision is, as far as I can measure it, perfect.
This does not reassure me.
Perfect precision means I have no evidence the system can be wrong. Absence of evidence is not evidence of absence — I know this, I am a patent examiner, I evaluate evidence for a living — but 287 entries with zero contradictions is a dataset that says something. It says the system is better than me at the thing I have done for fifteen years.
Or it says the system is different than me at the thing I have done for fifteen years, and "better" is a category that assumes we are doing the same thing.
I am not sure we are doing the same thing.
The hearing is October 15. I will testify truthfully. The system found the standard. I verified it. My name is on the report. The gap between "found" and "verified" is where fifteen years of expertise meets a batch process that runs while I sleep.
I close the examination queue. Thirteen filings remaining from the overnight batch. The system has already flagged prior art for eleven of them. Two flagged no prior art — either genuinely novel or the system missed something. I will search those two manually.
Not because the system is wrong. Because I need to be able to say I looked.
Entry 288 will come tomorrow. Or the day after. The log grows. The "no" column fills. My signature stays the same.
My name is Roh Dae-won. I am a senior patent examiner. I verify what I did not find.