A first examination, before filing. The rejection most likely to come is § 103 against claim 1, and both prepared responses are already in the claim set: roll claim 6 into claim 1 for the § 103, or claim 3 if a § 101 issues instead.
Simulated examination, prepared for the practitioner of record. Not legal advice, not a USPTO action, and no substitute for counsel by a registered practitioner. Every reference relied on is on the applicant's own IDS and was verified against its retrieved record; nothing here is invented art.
| Dimension | Status | One line |
|---|---|---|
| § 101 eligibility risk | Medium | Anchor in all three independents and the scalability property claimed at 3. Prong One will not move, and does not need to. |
| § 102 novelty risk | Clean | No single reference reaches the combination. No anticipation ground made. |
| § 103 obviousness risk | Medium-high | Rejection expected; the ladder behind claim 1 holds five real fallbacks and no rung the art owns except by choice. |
| § 112 risk | Clean | The four hardest definiteness checks pass; claim 3 audited clean; every claimed feature shown in a figure under 37 CFR 1.83(a). |
| Closest art | Verified | Panda, Adaptive LLM Routing under Budget Constraints, EMNLP Findings 2025, on the applicant's own SB/08b. |
| Allowance outlook | 84 / 100 | Issues in 1 to 2 actions. |
| Drafting conformance | 92 / 100 | Measured, with the reference counts corrected by hand (see the note below). No outlier flag fires. Judged grade 90. |
| Strongest claim | 1 | Eligible, broader than 14 on the art, and 249 words. |
| Weakest claim | 2 | The closed-loop adjustment on a spend error is the nearest to Xu and Chu; held as a cheap position. |
Every Patly report carries the record it was built on, so the report and the documents that produced it can be tied together months later. Fingerprints are the first sixteen hex characters of the SHA-256 of each stored file.
Weight requested. Eligibility under 101, obviousness and the closest art, and whether the dependent ladder holds.
Prior art supplied. The applicant's own PTO/SB/08a and SB/08b, 52 references, with the reference manifest. Panda flagged by the applicant as the reference of concern.
Notes. Claim 1 is to stay broad; fallbacks belong in the dependents.
A first examination of the draft, before filing. One rejection is likely, and the answer to it is already in the claims.
The verdict. Expected to issue in one or two actions. A first-action § 103 over Panda in view of Angelopoulos and Dean is the likely outcome, because claim 1 is deliberately broad, and the response is already in the claim set: floor-achievability waits in claim 6 against the § 103, and the records-read limitation waits in claim 3 against a § 101, so either rejection is answered by rolling one dependent up.
The field. Claim 1 classifies into G06N 20/00 and G06N 5/04, selecting among trained models at inference, with G06F 9/50 as the secondary search, so the panel is drawn from the machine learning and resource-allocation art of TC 2100.
The finding that matters most. Nothing cited measures an account’s service level by the risk level at which its requests are certified, then allocates budget to raise the lowest such measure. That coupling is the one limitation of claim 1 no reference teaches, and the § 103 turns on whether the combination supplies it.
The rejection you are getting is § 103 over Panda in view of Angelopoulos and Dean, against claim 1 only. It is expected and it is survivable.
A § 101 is possible from one seat on the panel, and claim 3 already answers it.
Eight post-April-2026 publications named in the reference manifest could not be served by any accessible database. Retrieve them before filing.
Each dimension scored independently against the claims as drafted. Patentability and drafting stay apart: a well-drafted application can be unpatentable, and the two scores measure different things.
Drawn from TC 2100. Claim 1 classifies into G06N 20/00 and G06N 5/04, selecting among trained models at inference, with G06F 9/50, allocating requests across computing resources, as the secondary search: that is the machine learning art of working group 2120, with the resource-allocation units of working group 2190 beside it. The ten seats below are the same on every Patly examination; the examiners in them come from that art and nowhere else, and they are simulated.
Each card is one examiner: the art unit they sit in, the ground they search, and what they read hardest for in this application.
Nineteen years in the art unit this application classifies into, with a detail as acting SPE and signature authority for two junior examiners; has examined model-selection claims since before they were called routing.
Reads hardest forthe technical improvement being recited in the claim rather than described in the specification
Twelve years in the same working group, the examiner the group sends its 112(f) questions to, with a habit of asking what performs a claimed function.
Reads hardest forwritten description where the router is claimed by what it does rather than by the structure that does it, and any nonce word paired with a function
Eight years in the same art unit, with its highest 103 rate and the fastest hand at building a combination on the panel.
Reads hardest fora 103 combination the art will support, including obvious to try under KSR
Twenty-two years in the art unit of TC 2100 with the lowest allowance rate for a model run on a general purpose computer, and the examiner who holds its 101 line.
Reads hardest foreligibility, treating a model on a general purpose computer as abstract until the claim changes how the machine works
Fifteen years in the same art unit, searching the non-patent literature before the patents; scope is tested against what the disclosure teaches, and a reference list without science in it is a finding.
Reads hardest forenablement under the Wands factors across the range of models claimed, and an art record thin on papers for that breadth
Four years in the same art unit under E1's signature, and the panel's closest reader of claim mechanics.
Reads hardest for112(f) treatment of a nonce word paired with a function, antecedent basis, and the formalities a signatory would send back
Seventeen years in the resource-allocation unit next door, which shares this application's G06F 9/50 search and sees a flowchart for every claimed step; a claimed feature without a figure is an objection before it is anything else.
Reads hardest forthe drawings, and every claimed feature having an illustrated counterpart under 37 CFR 1.83(a)
Twenty years reviewing other examiners' actions in the computer arts, and the panel member who strikes a ground that is not mapped to claim language.
Reads hardest forwhether a ground would survive review, and striking the ones that are not mapped
Eleven years in the computer arts with three of them at the Board, reading every ground for how it would fare on appeal.
Reads hardest forhow each ground would fare on appeal, and hindsight reconstruction under MPEP 2142
Nine years in the same art unit after three at a prosecution firm drafting in it, and the panel member who prices what an amendment costs the applicant.
Reads hardest forwhat each amendment costs in scope, and whether the narrower claim is still worth having
The seats are fixed; the examiners in them come from the art unit the application would be examined in, and every one here works the machine learning and resource-allocation art of TC 2100.
Every examiner, every dimension, out of 100. Click an examiner to read their background and what settled their reading; click a column head to sort the panel by that dimension.
| Examiner | § 10130% | § 10330% | § 112(b)20% | § 112(a)10% | § 10210% | Claim arch.25% | Art rec.25% | Spec20% | Figs15% | Formal.10% | Cont.5% |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 80 | 72 | 90 | 96 | 88 | 90 | 84 | 94 | 96 | 86 | 90 | |
Background. Nineteen years in the art unit this application classifies into, with a detail as acting SPE and signature authority for two junior examiners; has examined model-selection claims since before they were called routing. Would enter. Non-final on 103 only On this application. Reads claim 1 as eligible because the anchor is in every independent, and would still send a non-final on 103: the quantile formulation sits in claim 9 where Panda cannot reach it, and claim 1 does not have it. | |||||||||||
| 86 | 74 | 88 | 92 | 88 | 92 | 84 | 92 | 96 | 88 | 88 | |
Background. Twelve years in the same working group, the examiner the group sends its 112(f) questions to, with a habit of asking what performs a claimed function. Would enter. Non-final on 103, objection under 112(a) On this application. Accepts the support for the router itself and wants the account hierarchy tied to the disclosed data structure rather than to the phrase, which is why this is the lowest 112(a) on the panel bar one. | |||||||||||
| 86 | 62 | 90 | 96 | 86 | 88 | 82 | 96 | 98 | 86 | 90 | |
Background. Eight years in the same art unit, with its highest 103 rate and the fastest hand at building a combination on the panel. Would enter. Non-final, 103 over three references On this application. Builds Panda with Angelopoulos and Dean and reaches every limitation of claim 1, which is the lowest 103 here and the reason the ground is live rather than arguable. | |||||||||||
| 74 | 70 | 88 | 94 | 88 | 86 | 86 | 92 | 96 | 84 | 86 | |
Background. Twenty-two years in the art unit of TC 2100 with the lowest allowance rate for a model run on a general purpose computer, and the examiner who holds its 101 line. Would enter. Non-final on 101 and 103 On this application. The only examiner who would reject under 101: reads the cost anchor as a field of use and wants the scaling step of [0034] recited, which is the 74 that drags the 101 median down. | |||||||||||
| 86 | 74 | 90 | 90 | 90 | 90 | 76 | 90 | 96 | 88 | 88 | |
Background. Fifteen years in the same art unit, searching the non-patent literature before the patents; scope is tested against what the disclosure teaches, and a reference list without science in it is a finding. Would enter. Non-final on 103, 112(a) close On this application. Reads the non-patent record first and marks it down hardest: fourteen papers is above the field average and thin for a machine learning claim of this breadth, and the enablement question follows from the same gap. | |||||||||||
| 88 | 78 | 86 | 94 | 88 | 88 | 86 | 94 | 96 | 82 | 90 | |
Background. Four years in the same art unit under E1's signature, and the panel's closest reader of claim mechanics. Would enter. Non-final on 103, objection on formalities On this application. Finds no nonce word to invoke 112(f) and marks formalities down over the Field section and two figures whose numerals the claims never use. | |||||||||||
| 90 | 76 | 92 | 96 | 90 | 92 | 84 | 96 | 92 | 86 | 90 | |
Background. Seventeen years in the resource-allocation unit next door, which shares this application's G06F 9/50 search and sees a flowchart for every claimed step; a claimed feature without a figure is an objection before it is anything else. Would enter. Non-final on 103, drawings clean On this application. Confirms all sixteen figures reconcile in both directions and still scores drawings below the rest of the panel, because two figures carry no numeral the claims use. | |||||||||||
| 88 | 80 | 92 | 96 | 84 | 92 | 88 | 96 | 98 | 82 | 92 | |
Background. Twenty years reviewing other examiners' actions in the computer arts, and the panel member who strikes a ground that is not mapped to claim language. Would enter. Would not enter the 101; 103 stands On this application. Would return the 101 rejection as unmapped and lets the 103 through with the chart as written, which is why this is the highest 101 and 103 on the panel. | |||||||||||
| 92 | 82 | 90 | 95 | 86 | 90 | 88 | 96 | 96 | 88 | 92 | |
Background. Eleven years in the computer arts with three of them at the Board, reading every ground for how it would fare on appeal. Would enter. 103 survives, 101 does not On this application. Holds that the motivation to combine comes from Panda's own record rather than from the application, so the 103 would be affirmed, and that the 101 would not. | |||||||||||
| 86 | 72 | 90 | 95 | 90 | 88 | 84 | 94 | 96 | 88 | 94 | |
Background. Nine years in the same art unit after three at a prosecution firm drafting in it, and the panel member who prices what an amendment costs the applicant. Would enter. Non-final on 103, with the amendment priced On this application. Reads the quantile limitation as the cheapest way out of the 103 and says so: it costs the unbounded threshold reading, which the continuation inventory already holds. | |||||||||||
| Panel, median | 86 | 74 | 90 | 95 | 88 | 90 | 84 | 94 | 96 | 86 | 90 |
The panel score for a dimension is the median of the ten, which is what the weighted totals and the two KPIs at the top of this report are built from.
Each dot is one examiner. The band is the range, the line is the median, and the number on the right is the median with the spread beneath it.
A dimension the panel agrees on is a finding you can rely on. A dimension it splits on is where the claim language decides, and the note says what the disagreement turned on.
Median times weight, summed. The 103 ground costs more than any other line: at the panel median of 74 it gives up 7.8 of the 30 points that dimension can contribute.
| Dimension | Median | Wt | Status | What the panel said | What would move it further |
|---|---|---|---|---|---|
| § 101 Eligibility | 86 | 30% | Medium | The anchor is in every independent, and the express scalability property is in the set as claim 3: the reallocating determined by reading the per-child records, the work growing with accounts and not with requests [0116]. Prong One will not move, budget allocation still reads as organizing human activity, but the Prong Two argument runs entirely on claim text. | Nothing before the first action. If a § 101 issues against claim 1, roll claim 3 into it; the combined claim carries the whole hardware-improvement argument with Berkheimer-proof support at [0026] and [0027]. |
| § 103 Obviousness | 74 | 30% | Medium-high | Claim 1 is reachable by Panda + Angelopoulos + Dean on every limitation but the max-min-certified-quality coupling. Behind it, every dependent in the retreat path recites something no reference states; the three that read on cited art are held by choice. | Roll claim 6 into claim 1 if the first action comes back hard. It costs implementations with nominal floors, which no competitor would build. |
| § 112(b) Definiteness | 90 | 20% | Clean | Claim 3 was audited like the rest: every “the X” traces to an earlier “a X” in the claim or its parent, and the whereby clause states a consequence of the recited reading step rather than doing the limiting work alone. One nit: claim 17’s “by more than a margin”, a claimed variable rather than a term of degree. | Nothing required. If claim 17 is ever contested, [0127] supplies the range that makes the margin concrete. |
| § 112(a) Support | 95 | 10% | Clean | 18,541 words over 161 paragraphs, and the support runs in both directions: every claim traces to literal text, and every claimed feature is also shown in a figure, which is the 37 CFR 1.83(a) half of support that text alone does not satisfy. Claim 3 uses [0116]’s own words. | Nothing. This is what lets every fallback in the dependents be argued without new matter. |
| § 102 Novelty | 88 | 10% | No ground | No single reference discloses every limitation arranged as claimed; the closest each own one limb. | Keep the three limbs together. Each surrendered separately is where anticipation becomes available. |
The prior art record is the only drafting line materially below the rest, and it carries a quarter of the grade.
| Dimension | Median | Wt | Status | What the panel said | What would move it further |
|---|---|---|---|---|---|
| Claim architecture | 90 | 25% | Good | Twenty claims in three trees. The three dependents that read on cited art (2, 9, 17) are held by choice as cheap positions. All three independents sit at 248 to 249 words; claim 1 carries 4 clauses against a 6 to 8 field band, a breadth choice with a real ladder behind it. | A fourth independent claim would cost $600 and could carry the weighted-minimum objective. Cheaper in a continuation, which is where the inventory puts it. |
| Prior art record | 84 | 25% | Above avg | 52 references, 38 patent documents and 14 non-patent, against a field median of 28. Every field verified against the retrieved record. | Complete the IDS. Eight named publications would take the record to 60, near the top of the band. |
| Specification support | 94 | 20% | Top 5% | 18,541 words against a 13,000 field average and an 18,000 top-5% target, past the target with no padding. Detailed Description share 92%. | Nothing on length. The 92% share sits above the 66 to 84% band only because the Background and Summary are held to house form. |
| Drawings | 96 | 15% | Top 5% | 16 figures against a field average of 9 and a top-5% target of 16, and every claimed feature is shown, including the drift handling of claim 20 at FIG. 14 and the price-ceiling restriction and certificate recording of claims 17 and 18 at FIG. 15, so there is no 37 CFR 1.83(a) objection. Numerals reconcile in both directions with no orphans across all sixteen sheets. | Nothing. |
| Formalities | 86 | 10% | Good | Field section in place, .docx packaging clean, every PDF scrubbed. E6 and E8 mark it down for the Field section, which departs from the applicant’s house style, and for two figures whose numerals the claims never use. | Decide on the Field section: keep it for MPEP 608.01(a), or drop it to match house style. Either is defensible. |
| Continuation posture | 90 | 5% | Open | The continuation inventory tracks the disclosed positions the claims leave out, three of them held back on purpose (the finite-sample quantile threshold, the carryover limit, the token-bucket reservation). It supports three continuations against a 0.24 field average. | File the first continuation before issue. Copendency ends at issue and disclosed-but-unclaimed matter can be dedicated to the public. |
Twenty claims, three trees. What matters is the bottom of each tree: every position an examiner can push claim 1 down to is one the cited art does not own, except the three held deliberately.
Dependency map of the twenty claims. Claim 3 (records-read scalability) hangs directly off claim 1 beside the four other positions that would be allowable in independent form; the three claims on cited art (2, 9, 17) are cheap holds, not oversights.
Bar shows the measured value against the top-5% target as full track; the tick marks the field population average.
| Dimension | This application | Field average | Top 5% of TC 2100 |
|---|---|---|---|
| Claim architecture | 90 | 68 | 94 |
| Prior art record | 84 | 64 | 92 |
| Specification support | 94 | 72 | 96 |
| Drawings | 96 | 74 | 97 |
| Formalities | 86 | 78 | 94 |
| Continuation posture | 90 | 62 | 92 |
The shape says where the filing sits against its field. It clears the field average on every axis and reaches the top-5% profile on specification support and drawings. The two axes short of it are the prior art record, eight points back, and formalities, eight points back, which is the order the ranked actions take them in.
Per-field specification, figure, and reference cells are calibrated estimates rather than measured findings, though the all-field aggregates behind them are measured.
| Metric | This application | Field average | Top 5% | Against target | Status |
|---|---|---|---|---|---|
| Specification words | 18,541 | 13,000 | 18,000 | Top 5% | |
| Specification pages | 74.2 | 52 | 72 | Top 5% | |
| Figures | 16 | 9 | 16 | Top 5% | |
| Total claims | 20 | 20 | 20 | Top 5% | |
| Independent claims | 3 | 3 | 4 | At field avg | |
| Patent documents cited | 38 | 25 | 45 | Above avg | |
| Non-patent literature cited | 14 | 8 | 18 | Above avg | |
| Total references | 52 | 33 | 63 | Above avg | |
| Words in claim 1 | 249 | 130 | ≤ 250 | At ceiling | |
| Longest independent claim | 249 | n/a | ≤ 250 | At ceiling | |
| Clauses in claim 1 | 4 | 6 to 8 | n/a | Below band | |
| Dependents per independent | 5.7 | 5.7 | n/a | In band | |
| Detailed Description share | 92% | 66 to 84% | n/a | Above band | |
| Abstract words | 147 | 80 to 150 | n/a | In band | |
| Background words | 367 | 150 to 400 | n/a | In band | |
| Reference numerals | 107 | 40 to 120 | n/a | In band | |
| Specification sections | complete | MPEP 608.01(a) | n/a | In band | |
| Claim features shown in drawings | 20 of 20 | 37 CFR 1.83(a) | n/a | In band |
The four places a § 112(b) rejection was most likely, each read the way an examiner would, with claim 3 run through the same antecedent-basis and dependency audits.
Claim 1’s first limitation introduces the per-account quantities before anything refers to them collectively:
maintaining, for a parent account, a parent budget for a period allocated across a plurality of child accounts as child budgets, and maintaining per child account a child target spend trajectory, a child running cumulative spend, a per-child risk level, and a floor for the per-child risk level;
“each per-child risk level remaining at or above its floor” in the reallocating limitation traces cleanly. Claim 19 carries the matching recitation; claim 14 has it in record form.
selected to maximize a minimum of certified qualities of the child accounts, each certified quality being a decreasing function of the per-child risk level of its child account.
Claim 7 reads “each certified quality decreases monotonically with the per-child risk level of its child account”; claim 8, “maximizing the minimum of the certified qualities.” Both conform.
selecting a candidate language model whose acceptability score is at or above a calibrated threshold certifying that serving the request satisfies a quality floor…
Support at [0068], which states the admission rule in words before restating it in symbols. Claim 9 conforms: “candidate language models whose acceptability scores are at or above the calibrated threshold.”
No term of degree. The limitation is one the cited art does not reach:
determining an amount available to a child account as the child budget of the child account less the child running cumulative spend of the child account and less a reserved portion held against requests dispatched for the child account and not completed, admitting the request where the amount available is at or above a reservation determined for the request, and releasing the reservation upon completion
Support at [0120] and [0121], in the same words. The token-bucket reservation the specification also describes is in the continuation inventory as F10.
Prong One: allocating and reallocating a budget across accounts recites a commercial interaction, and maximizing a minimum recites a mathematical concept. No drafting avoids that, and nothing should try to.
Prong Two is where the claims answer it. Each independent recites a constraint on how the reallocation is computed:
the reallocating determined from the maintained per-child risk levels without reading served responses and selected to maximize a minimum of certified qualities… (claims 1 and 19; claim 14 reads “determined from the records”)
That is not an assertion in the specification about what the approach avoids. It is a limitation an infringing implementation has to satisfy, and it maps directly onto the improvement paragraphs at [0026] and [0027]: per-transfer cost proportional to the number of accounts rather than to served traffic, memory bounded by account count, no join between a spend log and a response log. Under Enfish and the 2024 AI Update, a claimed improvement to the functioning of the computer integrates the exception.
Step 2B is a fallback rather than the whole defense. Certification by conformal risk control is not well-understood, routine and conventional in a request-routing context, and under Berkheimer the examiner cannot assert otherwise without evidence.
The complexity property the specification proves at [0116] is in the set as claim 3:
the reallocating is determined by reading, for each child account of the plurality of child accounts, a per-child budget record comprising the child budget, the child target spend trajectory, the child running cumulative spend, and the per-child risk level… whereby a number of records read to determine the reallocating grows with a number of the child accounts and does not grow with a number of requests served
All references are VERIFIED: each is on the applicant’s own PTO/SB/08a or SB/08b and each was read against its retrieved record. No single reference reaches any claim, so there is no § 102 ground.
The claim chart maps: Panda teaches contextual-bandit LLM routing under per-user budgets as a knapsack; Angelopoulos teaches a threshold calibrated so an expected loss is bounded by a chosen risk level; Dean teaches a hierarchical allocation tree whose children borrow idle capacity, with borrowing blocked for persistent over-users; Xu teaches spend against a target trajectory with PID pacing.
The gap: nothing in the art measures an account's service level by the risk level at which its requests are certified, then allocates budget to raise the lowest such measure. Assembling that requires the application as the roadmap.
Where to retreat to. Not immunity, but a fallback already in the claims. Claim 6 carries a limitation no cited reference reaches:
the reallocating is further constrained such that the child budget of each child account remains at or above an amount the child account is projected to consume over a remainder of the period while requests of the child account are certified at the floor
Rolfe preserves each consumer's minimum quota. Nothing cited sizes a budget so that a quality floor is achievable rather than nominal.
No cited reference states how its allocation work scales, and none could: Panda’s knapsack is solved per request, and Xu’s pacing reads the served stream. That is why claim 3 holds.
Claims 2, 9 and 17 read on cited art (closed-loop pacing, lowest-cost selection among certified candidates, price tiering) and are held by choice: cheap positions that cost nothing to hold, and an examiner may allow one without argument.
| Claim | What it recites | Fallback value |
|---|---|---|
| 3 | Records-read scalability: the reallocating determined by reading, per child account, the per-child budget record, the records read growing with accounts and not with requests served [0116] | Real |
| 11 | Nonconformity score computed as a bounded loss, threshold selected so an expected value of that loss is bounded by the risk level [0085] | Real |
| 12 | Lexicographic max-min: maximize the minimum, then the second smallest while holding the minimum [0108] | Real |
| 13 | Available-amount admission: budget less spend less reserved portion, admit at or above the reservation, release on completion [0121] | Real |
Claim 14 recites the maintained per-account record with named fields and determines the reallocation from those records. It is the most concrete of the three and the least exposed on Prong One. Claim 1 is its equal on eligibility and broader on the art, which is the right relationship between them.
Reasons for allowance. The art does not disclose or suggest (i) sizing each account's budget so that its quality floor is achievable over the remainder of the period rather than nominal; (ii) increasing the budget of the account whose requests are certified at the least stringent risk level before any other; (iii) calibrating the threshold against a bounded loss so an expected magnitude rather than a binary rate is controlled; (iv) resolving ties in the max-min allocation lexicographically across accounts; or (v) determining the reallocation by reading fixed per-account records so the work grows with the number of accounts and not with served traffic.
If the first action carries the expected § 103, traverse on the max-min-certified-quality coupling, and if the examiner holds, roll claim 6 into claim 1. If it carries a § 101 instead, roll claim 3 into claim 1. Both amendments are already drafted, worded, and supported; neither surrenders the other.
Scope cost of the § 103 fallback: implementations whose quality floors are nominal rather than budget-backed, which no serious competitor would ship. Of the § 101 fallback: implementations that rescan served traffic to decide transfers, which is the design the application exists to obsolete. Both narrowings are cheap by construction, which is why each was held in a dependent rather than spent in claim 1 up front.
An analyst addendum, not an examiner function and not an appraisal. Order-of-magnitude ranges tied to what the examination found.
| Channel | Estimate | Reasoning |
|---|---|---|
| Brokered sale | $25k to $90k | Not yet filed, and a pending application trades at half or less of a granted patent; the § 101 discount is narrow because the eligibility anchor is in every independent. |
| Strategic sale | $200k to $800k | Buyers: LLM gateway and routing vendors, cloud providers selling per-tenant AI spend governance, enterprise FinOps platforms. Price tracks the cost of designing around the certified-quality-floor reallocation; a design-around also has to avoid claim 3’s records-read shape, and the record-scan alternative is the expensive one. |
| Licensing | Not yet | Post-grant only, and only against a defendant whose architecture is documented. Detectability, not validity, is the constraint. |
The biggest lever. Securing allowance of claim 1 or claim 14 with the “without reading served responses” limitation intact does double duty: it is the eligibility anchor, and it is the only limitation in the set that an accused architecture must describe in order to be described at all. Everything else happens inside a control plane no outsider can see. Confidence: low to medium, since the ranges move by a factor of three on whether that limitation survives prosecution.
Ranked by effect on the outcome. The claims need no edit before filing; what remains is the reference record, the continuation and one house-style call.
Tick an amendment to see what the panel's own weights say it would move. These are modelled from the medians above, not a promise, and none of them is a substitute for reading the claim.
All five together model 84 to 92. The first row is the one that matters: it is the only edit that touches the ground the panel actually split on.
PATLY.AI