
Patent drafting with agents
Provisional patent from scratch: claim hierarchy, prior-art search, evidence ledger, and the CIPHER workflow. With real QC examples.
What agents can and can’t do in patent work
Quantum Caddy assembled a patent package on April 19 — 482 lines of claim drafts, prior art analysis, and a 36-commit evidence ledger. The patent attorney engagement is the current critical-path legal item (Patent Week 1 on the Kanban). The Cipher content-safety pass is required before any filing action. The provisional patent has not been filed; that requires the attorney.
This is the correct division of labor. CIPHER can draft claims with specific technical language, run prior-art searches against USPTO PAIR and Google Patents, build and maintain the evidence ledger, and annotate each prior-art reference with how the QC claims are distinguished. What CIPHER cannot do: file with any authority, represent terms to any party, act as legal counsel, or provide legal advice. CIPHER is a drafting and research agent.
The value CIPHER adds is significant. A first-draft claim set from CIPHER, with QC-specific technical language about the triple-layer fusion architecture, the physics filter in ThrowDetector._finalize_throw, the confidence-weighted fusion at 0.70/0.85 thresholds, gives the attorney a technically accurate starting point rather than a blank page. Attorneys bill for drafting and research. An agent that produces a strong first draft reduces attorney hours. TPL-2026-015 measures this effect in practice.
The patent scope implication of the April 22 sensor pivot is also live: FSR-specific component claims (routed pocket, Zener clamp) should be replaced or supplemented with broader method-level fusion claims (RGB + depth-ToF + radar + automatic training-label generation) after the bench validation of the Acconeer A121 and ST VL53L8CX. CIPHER’s claims drafts should reflect the current sensor strategy, not the superseded FSR strategy.
The provisional patent workflow
A provisional patent application establishes a priority date — the date from which novelty and prior art are assessed. It doesn’t require formal claims, but the description must be enabling: sufficient detail that a person of ordinary skill in the art could practice the invention. You have 12 months from filing to convert to a non-provisional or the priority date is lost.
| Step | Who does it | Output |
|---|---|---|
| 1 — Invention disclosure | Inventor + CIPHER | Technical description: what it does, how it does it, what’s novel |
| 2 — Evidence ledger | AXIOM + CIPHER | Commits, test results, artifacts tagged to claim elements with dates |
| 3 — Prior-art search | CIPHER | Annotated reference table: each reference with distinction and risk level |
| 4 — Claim drafts | CIPHER | Claim set with hierarchy: independent method, independent system, dependent implementation claims |
| 5 — Technical accuracy review | AXIOM + inventor | Claims annotated for technical accuracy; any incorrect or over-broad claims flagged |
| 6 — Attorney review + filing | USPTO attorney | Amended claims, filed provisional application |
CIPHER’s deliverable is the package at the end of step 4. Steps 5 and 6 require humans. No CIPHER session ends with a filing action — it ends with a draft that goes to AXIOM for technical review, then to the attorney.
Claim hierarchy and prior-art search
The QC claim structure
The Quantum Caddy invention has four distinct claim families, each grounded in a specific technical contribution:
Family 1, triple-layer fusion for physical game scoring. The broadest claim is a method for real-time physical game scoring combining computer vision, pressure sensing, and beam sensing. The distinguishing element is the heterogeneous sensor fusion, unlike prior systems using homogeneous multi-camera arrays. Prior art risk is moderate. Sports analytics systems have multi-sensor scoring, but heterogeneous fusion for hobbyist/casual game applications is less well-covered.
Family 2 — Confidence-weighted fusion protocol. Specific implementation of the fusion: CV confidence at 0.70+ takes primary weight; 0.85+ takes precedence; below 0.70, pressure sensor is authoritative; IR break beam resolves in-hole class disagreements. The thresholds were derived from the NMS sweep (v15, April 19). The evidence: commit b933a82 (physics filter), the NMS sweep results log, the confusion matrix from the validation runs.
Family 3 — Automatic training label generation from sensor agreement. The self-supervised pipeline: when all three sensing layers agree on a scoring outcome, the event is automatically labeled and added to the training dataset. This produces continuous improvement without human annotation. The distinguishing element: using multi-sensor agreement as the labeling signal rather than human annotation or single-sensor confidence. Prior art risk: the specific application (physical game scoring) is novel; the general concept (agreement-based pseudo-labeling) is in prior art for NLP and computer vision.
Family 4, physics-filter event classification. The upstream physics filter in ThrowDetector._finalize_throw (commit b933a82) that prevents body-motion events from reaching the pair/score path. The distinguishing element is physics-based pre-filtering as a prerequisite to sensor fusion, not as a post-processing step. Prior art risk is lower. This specific architecture (physics filter upstream, before fusion) is less common than post-processing filtering approaches.
Prior-art search strategy
The three-layer search structure applies to each claim family: direct hits (same technology, same domain), adjacent domain (same technology, different domain), and component patents (sub-components that individually exist in prior art but whose combination is novel).
For Family 3 (automatic label generation), the adjacent domain is particularly important. Self-supervised learning with agreement-based pseudo-labeling is well-represented in NLP (three-system agreement for training data annotation). CIPHER’s prior art annotations for this family must clearly distinguish: the QC method uses heterogeneous physical sensors (not language models), produces real-time labels during live game play (not batch annotation), and the labeled events are game scoring events (not text spans). Each distinction is a claim narrowing that’s worth having.
Three failure patterns
Where patent drafting with agents goes wrong. Two are about claim quality; one is about agent authority.
Broad method claim with no reduction to practice
Evidence ledger not maintained during development
CIPHER acting as filing agent rather than drafting agent
The CIPHER authority boundary failure (pattern 03) is the highest-consequence failure. A mis-filed provisional with bad claim language may not be fixable before the priority date window closes. The draft-then-attorney-files workflow is non-negotiable.
The claim-to-evidence linkage test
Run this test on the claim set before it goes to the attorney. It surfaces claims that lack evidence support — the most common weakness in first-draft patent applications.
Run a claim-to-evidence linkage test on this claim set.
Claims: [paste the claim set]
Evidence ledger: [paste or describe the evidence ledger]
For each independent claim:
1. REDUCTION TO PRACTICE: Is there a development artifact
(commit, test result, screenshot, video) that demonstrates
this claim is operable?
- If yes: cite the specific artifact and date.
- If no: this claim may not have adequate support.
2. DISTINGUISHING ELEMENT: What element in this claim
distinguishes it from the prior art? Is that element
specifically present in the evidence ledger?
- If the distinguishing element is not in the evidence:
the claim may be invalidated at examination.
3. CONCEPTION DATE: What's the earliest development artifact
that demonstrates conception of this specific claim?
- Document: commit hash, date, what it implements.
For dependent claims:
4. DEPENDENCY CHAIN: Does each dependent claim add a specific
limitation that narrows the independent claim?
- Generic 'wherein the method further comprises' is not
a narrowing limitation. Name the specific technical
element being added.
Output: claim-by-claim linkage table + any claims
lacking adequate evidence support.Tests every claim for evidence support in the development ledger. Surfaces claims lacking reduction-to-practice artifacts and claims where the distinguishing element is not documented.
Annotated transcript
CIPHER drafting the QC independent claims and prior-art search strategy. Two turns — claim hierarchy and search strategy for the novel Family 3 (automatic label generation) claim.
Two operator turns, two agent turns. Result: claim hierarchy structured, automatic-label-generation claim identified as strongest novelty, three-layer prior-art search strategy built.
- operator
CIPHER needs to draft the independent claims for the Quantum Caddy provisional patent. The core innovation is the triple-layer sensor fusion for cornhole scoring — CV on the Jetson Orin Nano, HX711 load cells, IR break beam. What's the claim hierarchy?
- agent
The hierarchy for a triple-layer fusion system goes broadest-first. Here's the structure:
Claim 1 (independent method claim — broadest): A method for automatically scoring a physical game comprising: capturing visual data of a play surface using a camera; capturing physical contact data using a pressure sensor array; capturing event data using a beam sensor; fusing outputs from the visual, pressure, and beam sensors to determine a scoring outcome.
This claim is intentionally broad — it covers the class of multi-sensor fusion for physical game scoring without specifying the specific technology. It's the broadest defensible claim given the prior art.
Claim 2 (independent system claim): A system for automatically scoring a physical game comprising: an edge compute module executing a computer vision model; a pressure sensor array in communication with the edge compute module; a beam sensor in communication with the edge compute module; wherein the system fuses outputs from the vision model, the pressure sensor array, and the beam sensor to determine a scoring outcome.
Claims 3-8 (dependent method claims — specific implementation): These are where the real novelty lives:
- Claim 3: The method of claim 1, wherein the fusion step applies a confidence-weighted priority: vision output at confidence >= 0.85 takes priority; at 0.70-0.84, pressure sensor is weighted equally; the beam sensor resolves disagreements for the in-hole class.
- Claim 4: The method of claim 1, further comprising: generating ground-truth training labels from sensor agreement events, wherein a label is generated when all three sensing layers agree on an outcome.
- Claim 5: The method of claim 1, wherein the physical contact data is captured by four load cells arranged in a Wheatstone bridge configuration, and the fusion step applies quadrant-level spatial weighting.The automatic-label-generation claim (Claim 4) may be the strongest novelty claim — it's the self-supervised training pipeline that comes for free from the fusion architecture. Run the prior-art search on this claim first.
- operator
For the prior-art search, what's the search strategy for Claim 4 — the automatic training label generation from sensor agreement?
- agent
Search in three layers for Claim 4:
Layer 1 — Direct hits. Search USPTO PAIR and Google Patents for: 'automatic annotation' + 'sensor fusion' + 'training data'; 'ground truth generation' + 'multi-sensor agreement'; 'self-supervised' + 'physical sensor' + 'label generation.' These are the most obvious prior art candidates.
Layer 2 — Adjacent domain. The automatic-label-generation architecture is similar to what's used in sports analytics systems (ball tracking with multiple camera views generating training data from agreement). Search: 'multi-view agreement' + 'training label generation'; 'sensor fusion' + 'automatic annotation' + 'sports.' The QC claim needs to be distinguished from this adjacent domain — the distinguishing element is the heterogeneous sensor types (vision + pressure + beam) rather than homogeneous multi-view cameras.
Layer 3 — Component patents. Search for patents on the sub-components that, in combination, produce the claimed method: HX711 ADC with load cell arrays for game scoring; RTDETR or object detection + physics filtering for physical game events. If each component is in prior art but the combination producing automatic labels is novel, the combination claim may stand.
For each prior art reference found: annotate how Claim 4 is distinguished. The annotation format:
Reference: [patent number, title] Relevant claim: [what the prior art claims] Distinction: [what QC Claim 4 does that the reference does not] Risk level: HIGH / MEDIUM / LOW (to Claim 4)Bring the annotation set to the attorney, not just a list of found references. The attorney's job is easier — and the filing is stronger — when the distinctions are already articulated.
Provisional patent checklist
The complete pre-filing checklist for a provisional patent with agent assistance. Walk it before the package goes to the attorney.
PROVISIONAL PATENT PRE-FILING CHECKLIST
(agent-assisted workflow)
INVENTION DISCLOSURE (CIPHER + inventor)
[ ] What does the invention do? (1-2 sentences, plain language)
[ ] What's novel? ('The first system to...' or
'Unlike prior systems which..., this system...')
[ ] Who is the inventor(s)? (legal names, confirmed)
[ ] What is the conception date? (earliest artifact)
[ ] Is this a continuation or improvement of prior work?
(if yes: cite the prior work explicitly)
EVIDENCE LEDGER (AXIOM + CIPHER)
[ ] All relevant commits tagged to claim elements
[ ] Each tag: commit hash, date, what it implements,
which claim element it supports
[ ] Conception date documented with artifact
[ ] Reduction-to-practice date documented with artifact
(date the system first demonstrably worked)
[ ] Evidence ledger exported as dated artifact,
not stored only in git
PRIOR-ART SEARCH (CIPHER)
[ ] USPTO PAIR searched for each independent claim
[ ] Google Patents searched for each independent claim
[ ] Adjacent domain searched for closest analogous technology
[ ] Each reference annotated: relevant claim, distinction,
risk level (HIGH / MEDIUM / LOW)
[ ] High-risk references: claim language amended to
clearly distinguish, or claim narrowed
CLAIM DRAFTS (CIPHER)
[ ] At least one independent method claim
[ ] At least one independent system claim
[ ] Dependent claims for key implementation details
(these are the claims that can't be designed around)
[ ] Each claim: all elements linked to evidence ledger
[ ] Claim language reviewed for means-plus-function
(avoid unless intentional)
TECHNICAL ACCURACY REVIEW (AXIOM + inventor)
[ ] Every claim element is technically accurate
[ ] No implementation detail missing from claims that
is necessary to practice the invention
[ ] No claim element that doesn't reflect current
implementation (sensor pivot: update FSR claims)
CIPHER AUTHORITY CHECK
[ ] CIPHER has NOT filed with any authority
[ ] Package is marked DRAFT — attorney review required
[ ] Attorney engagement is confirmed before package is sent
[ ] CIPHER content-safety pass is complete
ATTORNEY REVIEW
[ ] Package sent to attorney: [name, date]
[ ] Estimated filing date: [date]
[ ] Priority date target: [date]
Prompt kit
Three prompts: claim-to-evidence linkage test, independent claim drafting, and prior-art search + annotation. Save in CIPHER’s agent file or your patent workflow CLAUDE.md.
Run a claim-to-evidence linkage test on this claim set.
Claims: [paste the claim set]
Evidence ledger: [paste or describe the evidence ledger]
For each independent claim:
1. REDUCTION TO PRACTICE: Is there a development artifact
(commit, test result, screenshot, video) that demonstrates
this claim is operable?
- If yes: cite the specific artifact and date.
- If no: this claim may not have adequate support.
2. DISTINGUISHING ELEMENT: What element in this claim
distinguishes it from the prior art? Is that element
specifically present in the evidence ledger?
- If the distinguishing element is not in the evidence:
the claim may be invalidated at examination.
3. CONCEPTION DATE: What's the earliest development artifact
that demonstrates conception of this specific claim?
- Document: commit hash, date, what it implements.
For dependent claims:
4. DEPENDENCY CHAIN: Does each dependent claim add a specific
limitation that narrows the independent claim?
- Generic 'wherein the method further comprises' is not
a narrowing limitation. Name the specific technical
element being added.
Output: claim-by-claim linkage table + any claims
lacking adequate evidence support.Draft the broadest defensible independent method claim
for this invention.
Invention description: [describe what it does technically]
Key distinguishing elements: [what this can do that
no prior system can do]
Target domain: [physical game scoring / AR overlay /
sensor fusion / other]
The claim must:
1. Be written in present tense ('comprising,' 'detecting,'
'fusing' — not 'which comprises' or 'that detects')
2. Include the minimum necessary elements to define
the invention — not every implementation detail
3. Name the distinguishing element explicitly —
not as a generic 'wherein the method is improved'
but as a specific technical step
4. Avoid means-plus-function language unless intentional
5. Be readable by a person of ordinary skill in the art
Output: the claim text + a one-paragraph explanation
of why this is the broadest defensible version.Run a prior-art search for this claim and annotate each
reference found.
Claim: [paste the claim]
Search databases: USPTO PAIR, Google Patents, Espacenet
Search terms: [the agent will generate these]
For each reference found, produce:
REFERENCE: [patent/publication number and title]
DATE: [filing or publication date]
ASSIGNEE: [company or inventor]
RELEVANT CLAIM: [the claim or element that's potentially
anticipatory or obviating]
DISTINCTION: [how the claimed invention differs from
this reference — be specific]
RISK LEVEL: HIGH (may anticipate) / MEDIUM (may combine
to obviate) / LOW (background prior art only)
Output: annotated reference table + summary of highest-risk
references + recommended claim amendments if any HIGH-risk
references are found.Apply this — produce one attorney-ready claim set
~2-hour exercise. Disclosure, evidence, claims, search. Produce a draft the attorney can work from — not a draft you need to apologize for.
Your first agent-assisted patent draft
Each step takes 20-30 minutes. Progress saves automatically.
- 01Write the answer to: 'What can this system do that no prior system can do?' in one specific sentence. This is your independent claim nucleus.If you can't write it in one sentence, the novelty isn't clear enough to draft claims around. The sentence is the pre-drafting gate.
- 02Start or audit your evidence ledger. Every claim element should be traceable to at least one development artifact with a date.For QC: the April 19 patent package has 36 commits tagged to claim elements. If you don't have this, start tagging commits to claim elements now — retroactively for existing work, prospectively from here forward.
- 03Run the claim-to-evidence linkage test on your current strongest independent claim. Find any gaps in evidence support.A gap found now is fixable — you can build more evidence, or narrow the claim to what you have. A gap found during examination is not fixable without amendment, and amendment may narrow your protection.
- 04Run a prior-art search on your most novel dependent claim (the one with the most specific technical element). Annotate every reference found.Dependent claims with strong technical specificity are often the hardest claims for a competitor to design around. Protecting them with strong prior-art annotations is worth the 30-minute search.
- 05Define CIPHER's role in your filing workflow explicitly: what does CIPHER produce, and what requires human attorney review before filing?CIPHER drafts; the attorney files. This boundary must be written in CIPHER's agent file, not assumed. The filing action is irreversible.