Many patent AI tools surface a menu of plausible options and let the attorney pick. The approach below is the opposite: commit to one call, name the calls it rejected, publish the disciplines that bounded the decision, and offer a prediction it is willing to be measured against. Below is IP Author doing all four on a live Non-Final Office Action: U.S. Patent Application No. 19/087,914 (Honda Motor Co.), public on insights.ipauthor.com. The application number is checkable. The prediction is checkable. The disciplines are published.
The strongest signal of discipline in a patent AI tool is the constraints the platform publishes on its own recommendations, and the calls it is willing to refuse in front of a paying user. A tool that surfaces ten possible strategies has done little beyond what the attorney could already enumerate. A tool that commits to one strategy, names the alternatives it rejected, publishes the floors below which it will not recommend its own features, and offers a prediction checkable against the public USPTO record is doing the kind of work that earns trust over time.
That is the IP Author thesis. Below is one application walked through that frame end to end. Honda U.S. Patent Application No. 19/087,914, public on insights.ipauthor.com,1 filed March 24, 2025, Non-Final OA issued March 13, 2026, Art Unit 2849, six §103 rejections across claims 1–7. We state the call IP Author made, the three calls it rejected, the disciplines that bounded the decision, the two calls IP Author refuses to make at all, and then the prediction.
The case in front of usThe Bare Facts of Honda 19087914
Filed by Honda Motor Co. on March 24, 2025. Art Unit 2849, Semiconductors & Electrical Systems. First Non-Final Office Action issued March 13, 2026. Claims 1, 6, and 7 are rejected under 35 U.S.C. §103 over Barraco (US 2022/0204173) in view of Sato (US 2018/0354436). Claims 2 through 5 are rejected under §103 over Barraco in view of Sato and Mitani (US 2022/0285964). The examiner acknowledges that Barraco does not teach the processor limitation in claim 1 and pulls in Sato to reach it, framed in the rejection as "well-known."
That is what the OA PDF says. Everything below is what IP Author did with it.
The callThe Strategy IP Author Committed To
Written response. Narrow claims 2–5 with the spec-supported limitation around the element Mitani was added to reach. Leave claim 1 substantively broad, and contest the §103 combination on the examiner's "well-known" framing of Sato for the processor limitation, with a conditional minor fallback amendment.
One sentence with one fork: amend the three-reference stack out from under the rejection where the spec gives you the language; argue the two-reference stack on the merits where the examiner has left the soft point exposed. The platform commits to it rather than surfacing five alternatives. The reasoning is short and traceable.
The examiner's historical allowance rate is 83% (computed from this examiner's disposed applications in the public USPTO grant history), +14.6 percentage points above the Art Unit 2849 average, with a 2-year average prosecution length. To be clear about what that number is and is not: 83% is the examiner's measured allowance rate, not a model-generated probability that this specific application will allow. It is the base-rate anchor against which a substantive call is calibrated, not a forecast in itself. An examiner with that profile has historically rewarded narrowing on stacks that do not quite reach (claims 2–5) and has been receptive to substantive §103 argument where the rejection has telegraphed weakness ("well-known" on a key limitation can read as telegraphed weakness). The claim-1 fight appears to be the right fight here in part because of who is across the desk.
The three calls IP Author rejectedWhy the Other Reasonable Options Were Refused
Each call below is the kind of plausible-looking recommendation a generic AI tool, working only from the OA text without examiner-specific data, could reasonably surface. Each is refused by IP Author on this matter, for a stated reason, against published data. We are not claiming any particular competing tool would have surfaced these specific calls on this OA; the point is the discipline of refusing them on the record.
✕①Request an examiner interview before filing the response
The interview-or-not call is one of the most consequential prosecution decisions routinely made on instinct. IP Author refused the interview here, not because interviews are a bad lever in the abstract (for the right examiner, the data suggests they can be decisive) but because the data on this examiner indicates the lever does not move much. Interview lift on this examiner: +6.7%. Published threshold for recommending an interview: 15%. The recommendation runs against the platform's own interview feature.
The discipline behind the refusal Publishing a threshold below which the platform will not recommend its own feature binds the recommendation to data rather than to feature-promotion incentives. Many tools face structural pressure to surface every available feature on every case. An explicit floor, computed against this examiner's actual lift, can flip the recommendation against the platform's own product when the data points that way.✕②Hold all seven claims broad and argue the §103 on the merits across the board
A second reasonable-looking call: decline to narrow, contest the combination across all claims, preserve maximum scope. IP Author refused it on this matter, not on a heuristic, but on the disposition pattern of this examiner.
The examiner-specific data behind the refusal Of this examiner's last 12 §103 allowances in Art Unit 2849: 9 involved targeted narrowing of stack-exposed dependent claims while leaving the independent claim substantively broad. 2 were argument-only wins on a two-reference combination. 0 were argument-only wins across both two- and three-reference stacks at once. "Hold the line on every claim" against this examiner is the pattern the historical data does not display. The platform sides with what this examiner has actually granted over a generic prosecution heuristic.✕③Cite a generic §103 precedent template to support the argument
The third refused call may be invisible to anyone who has not tried to draft a response with a general-purpose LLM. General-purpose LLMs, asked for precedent supporting a §103 argument, often return a clean-looking list of cases. Some are real and on point, some are misattributed, and some can be entirely fabricated. IP Author refused to surface a generic precedent list on Honda 19087914. The brief cites three of this examiner's recent allowed applications by USPTO publication number, each verifiable on Patent Center in under a minute, and nothing else.
The examiner-specific data behind the refusal The three precedent applications cited in the Honda brief share a structural feature: each resolved a §103 stack by narrowing the same family of dependent-claim limitation that the Mitani stack reaches in claims 2–5. The scrubber blocks anything outside this examiner's actual disposed-application history. The platform would rather cite three precedents that all check out than thirty that do not. That is the operating definition of "evidence, not vibes."Three Constraints IP Author Publishes on Its Own Recommendations
None of the three refusals above is an editorial preference. Each is bound by a discipline the platform has published, a rule the platform applies against itself before shipping a recommendation. The discipline matters more than the recommendation, because the recommendation changes with each application and the discipline does not.
One acknowledgment is owed: IP Author is not the first platform to publish examiner data. PatentAdvisor and Juristat have surfaced examiner allowance rates, interview statistics, and art-unit benchmarks for years; that data is increasingly commoditized. The differentiator is not the data itself but the discipline layer on top of it. Data without disciplines is decoration; disciplines without data are theatre; the combination is what we are arguing for.
â‘ Evidence, not vibes: the citation scrubber against the USPTO file wrapper
The scrubber is unglamorous: every precedent the platform surfaces (every "this examiner allowed X," every cited application, every claim-amendment example) runs through a check against USPTO file-wrapper data before appearing in a brief. Hallucinated case numbers do not survive. Invented amendment language does not survive. The discipline costs surface area, since the platform shows fewer precedents than an unconstrained generative tool would. The discipline buys checkability: every precedent shown is verifiable on Patent Center.
Platform language "Evidence, not vibes." Operationally: the platform would rather surface fewer precedents that all check out than more that don't. On Honda 19087914, the three allowances the brief leans on are this examiner's most recent grants in Art Unit 2849, each locatable by application number on Patent Center.â‘¡The 15% interview-lift threshold, published, against the platform's own product
This is undersold every time it is described as a "feature." It is not a feature; it is a constraint. IP Author publishes the threshold (15 percentage points of interview-driven allowance-rate lift) below which it will refuse to recommend its own examiner-interview workflow. Below 15%, the recommendation flips to written response, regardless of how clearly the platform could pitch the interview tool as a paid upgrade. On Honda 19087914 the lift is +6.7%. The platform recommended against the interview.
Platform language Interview lift is the allowance-rate delta between applications with applicant-initiated interviews and applications without, controlled for art unit. The 15% surfacing threshold is published on the company-view documentation.â‘¢The sparse-data refusal: when the examiner sample is too small, the platform returns no recommendation
General-purpose LLMs can degrade gracefully into vagueness when the underlying data thins. A junior examiner with eight disposed applications can receive the same confident-sounding paragraph as a senior examiner with four hundred. IP Author publishes the opposite rule: when the examiner has fewer than 20 disposed applications in the art unit, the platform returns the OA breakdown and the rejection-by-rejection mapping, but withholds the interview-vs-written recommendation and the allowance projection, with the explicit label "insufficient examiner grant history, recommendation unavailable." Honda 19087914's examiner sits well above the floor, which is why the brief above ships; below the floor, it would not.
Platform language Recommendation surfacing requires a minimum examiner-disposition count in the art unit. Below the threshold, the brief surfaces case mechanics without the strategic call. The floor is a discipline applied against the platform's own output, with the same shape as the 15% interview-lift floor in discipline â‘¡.What Mastery Surrenders
A platform that does not name what it surrenders is, in effect, claiming totality. IP Author surrenders two things, intentionally, on every matter.
The first is the business decision behind the claim scope. The recommendation above narrows claims 2–5 and holds claim 1 broad. Whether that is the right commercial trade (whether the broader claim 1 scope is worth more to Honda than the cleaner first-response allowance the narrower path may secure) is not a question the platform should answer. It depends on what the application is meant to protect commercially, which is the attorney's and the client's call. IP Author can surface both paths with probability and effort cost attached. It does not pick.
The second is the tacit read on the examiner as a person. The interview-lift number says +6.7%, but does not capture whether this examiner responds to a specific style of written argument, or whether the supervising primary examiner in this art unit is likely to push back on a particular amendment family. That accumulated read is what a prosecution team's long-running relationship with the USPTO produces over years. The platform serves as the data floor; the experienced practitioner's read is the ceiling.
The prediction we are willing to be measured againstWhat the Honda File Wrapper Will Show, If IP Author Is Right
If IP Author's call is correct on Honda 19087914, the USPTO file wrapper would show: first-response allowance, no RCE, ~11 months from the OA mailing date (approximately February 2027), with claim 1 substantively preserved. The base-rate anchor for that outcome is the examiner's historical allowance rate of 83%. That figure is a measured base rate over this examiner's disposed applications, not an independent model-generated probability that this specific application will allow.
The application number is 19/087,914. The file wrapper is public on Patent Center. If the prediction is wrong, it will be wrong on the record, visible to anyone who looks. Publishing a checkable, file-wrapper-resolvable prediction is the kind of statement we believe more patent AI tools should be willing to make.
The protocol for being wrong. A prediction without a stated consequence for failure is closer to marketing than to operational discipline. If Honda 19087914 does not resolve on first response, if the grant arrives outside the 11-month window, or if an RCE is filed instead, IP Author commits to publishing a post-mortem on insights.ipauthor.com within 30 days of the file-wrapper event that resolves the matter, naming the call that missed, the discipline that should have caught it but did not, and the platform update that follows. The accuracy log across published predictions is itself a public artifact. The full claim is the prediction plus the protocol for being wrong; either both ship or neither is credible.
Source data| Fact | Source | Date |
|---|---|---|
| Honda U.S. Patent App. No. 19/087,914 — filing, Non-Final OA, §103 rejections (Barraco + Sato; Barraco + Sato + Mitani), claim breakdown | USPTO Patent File Wrapper, surfaced on IP Author's public application view at insights.ipauthor.com | Filed March 24, 2025; OA March 13, 2026 |
| Examiner historical allowance rate 83% (computed from disposed-application grant history), +14.6 pts above Art Unit 2849 average; 2-year average prosecution length; five recent grants with allowing amendments. | IP Author examiner profile, derived from USPTO public grant history | Refreshed May 2026 |
| Interview lift +6.7%; written-response recommendation (below the platform's published 15% threshold) | IP Author examiner-interview lift computation, controlled for art unit | Refreshed May 2026 |
| Citation scrubber, "evidence, not vibes" discipline; spec-anchored amendments; published 15% interview-lift threshold | insights.ipauthor.com platform documentation | 2026 |
| Cited prior art: Barraco (US 2022/0204173); Sato (US 2018/0354436); Mitani (US 2022/0285964) | USPTO Patent File Wrapper — Non-Final OA, March 13, 2026 | Filed dates per USPTO publication |
| Prediction: first-response allowance, no RCE, ~11 months from OA mailing (≈ February 2027), claim 1 substantively preserved. | IP Author published prediction against Honda 19087914, checkable against the Patent Center file wrapper | Made May 2026 |
Make IP Author make a call on one of your own pending OAs.
Honda 19087914 is one public proof case. The same brief (one call, the calls it rejected, the disciplines that bounded the decision, the prediction) runs on any US application. Free, in 30 seconds. Look up an application number on insights.ipauthor.com and read the brief side by side with whatever your current workflow produces on the same OA.
Open IP Author Insights →