A prosecution co-pilot for patent attorneys. Office action responses brainstormed, applications drafted, your own drafts reviewed — first-pass work product a senior attorney edits instead of redoes.
Drafting and review run with zero external calls; nothing trains a model; every outbound query is logged. Read the security architecture →
| Claim | Rejections | Recommended path | Basis | Confidence |
|---|---|---|---|---|
| 5 | §102 | Amend — fallback | — | low |
| 6 | §112(b) · §112(a) · §102 | Argue + prepare amendment | Claim construction (#1) | moderate |
What it does
Five prosecution jobs. Documents in, first-pass work product out.
Drag in the office action, the disclosure, or your own draft. Each job returns the pass an associate would hand you — with the data posture stated on the card, so the confidentiality question is answered before you start.
Invalidity Search
Build the strongest §102/§103 case against a target claim set, with the closest art charted limitation by limitation.
Respond to an Office Action
The associate-pass before you draft the response.
It never argues against art it hasn’t read. A reference it can’t retrieve is marked unverifiable and taken as given — the analysis gets a verify-first posture instead of inventing a distinction over art it never saw.
Your stipulations are taken as given. Tell it what to concede and it scopes every argument to what’s left contestable — it doesn’t relitigate what you’ve already settled.
You write the response. This is the associate-pass before you start.
Whether §112(f) is invoked at all — 'filtering' may connote sufficient structure
MPEP § 2181 · § 2173.02
Identify the disclosed filter as the §112(f) corresponding structure — defeats both §112(b) and §112(a)
MPEP § 2181 · § 2163.03
Point to the disclosed structure showing possession, not to redundancy
MPEP § 2163
Reasonable-certainty rebuttal to the §112(b) indefiniteness finding
MPEP § 2173 · § 2173.06(II)
“Assume hypothetical reference X teaches every limitation the examiner maps to it (Fig. 1 elements 1–4, Fig. 2, page 5 lines 10–25). Do not contest X’s teachings. Focus the analysis on the §112 issues and claim construction.”
Provided by the attorney and taken as settled — the analysis scopes its arguments to the remaining contestable points.
Review My Draft
Read your own draft the way an examiner will.
Nothing about your draft leaves the boundary — zero external calls. You keep authorship; this is a reader, not a rewriter.
The lowest-friction way in — try it Monday morning on a draft already on your desk.
'means for filtering vibrational energy' invokes §112(f), but the spec links no structure to the filtering function.
MPEP § 2181
'said fastener' has no antecedent — parent claim 1 never introduces a fastener (the spec swaps 'connector' and 'fastener').
MPEP § 2173
Depends from 'claim 5' — which does not exist — and recites 'the ceramic coating,' unsupported anywhere in the spec.
MPEP § 2163
'well known and widely used' concedes identified art as admitted prior art an examiner can use under §103.
MPEP § 2129
'means for filtering vibrational energy…'
Missing — no structure expressly linked to a filtering function
'an enclosure'
Implicit — spec uses 'housing,' not 'enclosure'
'a plurality of folded-beam flexures'
Explicit — 'Four folded-beam flexures connect the housing…'
✓ explicit support · ◐ implicit · ⚠ §112 gap — add spec support before filing.
Draft an Application
The discipline, not just the generation.
Material it can’t trace to the disclosure goes to an inventor-confirmation list — never silently into your specification.
It audits its own draft and repairs only what the audit flagged, every edit shown before-and-after — the rest of your draft left alone.
Dual export when you’re done — a clean filing draft, and an attorney memo carrying every flag and its §112 basis.
⚠ 4 items below are not traceable to the disclosure and require inventor confirmation before filing.
a plurality of conductive layers
support: explicit · disclosure: generalized
arranged in a stack along a thickness direction
support: explicit · disclosure: disclosed
conductivity increases monotonically incident → opposite face
support: explicit · disclosure: disclosed
a mounting structure securing the stack (claim 7)
support: explicit · disclosure: confirm with inventor
a backing structure affixed to the opposite face (claim 7)
support: explicit · disclosure: confirm with inventor
Material that could strengthen the application but isn’t anchored in the disclosure — kept out of the specification prose until confirmed.
Before — “…wherein the increasing electrical conductivity of the conductive layers is configured to provide impedance matching…”
After — “…wherein the monotonically increasing electrical conductivity of the conductive layers is configured to provide impedance matching…”
The repair pass revised only the flagged claims and paragraphs — all other content untouched. Verify each edit before filing.
Why it doesn’t behave like a chatbot
Built like an associate. Checked like a machine.
A general chat answers in one pass, from memory. This runs the way a careful associate would — in stages, against the manual, adversarially — and then a machine checks the parts a person would miss.
A staged pipeline, not one long chat
The work moves through discrete stages — rejection map, claim chart, argument bank, decision table, stress test — each building on a checked prior step, not improvised in a single reply.
Doctrine from the manual, not from memory
MPEP sections come from a curated, edition-stamped corpus and are cited by number — §2181, §2173, §2163 — not recalled from model training, where citations drift or get invented.
An adversarial examiner pass
Before you see an argument, a separate pass plays examiner: it writes the strongest comeback and rates whether the argument survives — so the fragile ones are marked fragile, not sold to you.
Consistency enforced in code
An amendment can only target a claim the examiner actually rejected. The renderer checks that in code and flags anything that does not line up — a rule a chat thread has no way to guarantee.
A repair loop with diffs
When the audit flags a defect, a repair pass fixes only what was flagged and shows the edit before-and-after — the rest of the document left untouched.
Run your next office action through it and read the work product — free to start, no client data required to evaluate.
Start freeConfidentiality architecture
Built for privileged work — checkable line by line.
In patent work a careless disclosure isn’t a privacy incident — it’s prior art against your own client. Here is exactly what leaves the boundary, per job, stated specifically enough to check.
What leaves the boundary, per job
No training, either layer
Your inputs never train a model — not at Sparlo, and not at Anthropic under its Commercial Terms of Service.
Per-report isolation
Each analysis is pinned to your account at the database layer, not in application code. No cross-matter context, ever.
Encrypted job queue
Background-job payloads are application-layer encrypted and carry only record IDs — your document text never enters the queue or its logs.
A Data Egress Log on every run
Every report ends with a log of exactly what left the boundary — usually, in the product’s own words, zero external requests of any kind.
Passcode-only sharing
No share link exists until you make one; it serves nothing until a separate passcode is entered, and renders the finished work product only.
Browser-local PDF export
Privileged reports print through an in-boundary route your browser renders — the external PDF path refuses them, so the bytes never leave your machine.
Model calls go directly to Anthropic — SOC 2 Type II, ISO 27001:2022, ISO/IEC 42001:2023 — with its trust documentation and subprocessor list at trust.anthropic.com. No resold model marketplace, no failover chain: provider failover is disabled for these jobs, so your content is never rerouted to an alternate AI provider.
Answers for your firm’s diligence file.
Direct answers to the questions a Rule 1.6 diligence review will ask, stated specifically enough to be checked — and how the tool sits under ABA Formal Op. 512.
The path a disclosure takes
Your browser → Sparlo
Your document — office action, disclosure, or draft — travels over TLS to Sparlo and is stored in your private account in the United States (US-West). Row-level security at the database layer — not application code — means only your login can read it.
Locked to you alone
IP analyses are pinned to your personal account by a database trigger. They are never visible to teammates and never co-mingled with other firms. Nothing leaves your account unless you explicitly create a share link — and a shared page is passcode-gated and serves the finished analysis only, never the disclosure text.
Analysis on the Anthropic API
Model calls go directly to Anthropic under its Commercial Terms: no training on customer content, automatic deletion of API inputs and outputs within 30 days. Provider failover is disabled for IP analyses — your content is never rerouted to an alternate AI provider.
The report — and the delete button
The finished analysis lives in your account until you remove it. Deleting is immediate and hard: your document text, the analysis, and any uploaded file are purged from the database and file storage the moment you click — not on a schedule.
What your GC will ask
Is our client’s material used to train AI models?
No — at both layers. Sparlo does not train or fine-tune any model on customer content. Model calls run on the Anthropic API under Anthropic’s Commercial Terms of Service, which state that Anthropic "may not train models on Customer Content from Services." Anthropic also automatically deletes API inputs and outputs within 30 days.
Do our documents get sent to a search engine?
Your source documents — office action, specification, disclosure, or draft — are never sent to any search provider. Draft and review jobs make zero external calls at all. For the jobs that do search (office-action reference checks, patentability analysis), outbound queries are restricted to two providers — Exa (technical literature) and Google Patents — and general web search (Perplexity, Tavily) is disabled in code. Those queries are built from generalized invention terms so prior-art recall is maximized, and every run ends with a Data Egress Log that shows exactly what left. A stricter scrubbed mode also exists in code — queries restricted to search-safe framings, with the protected terms you confirm blocked by a deny-list gate before anything leaves — and your firm can elect it instead.
Could it leak through a share link?
Not unless you create one — and even then, two factors stand between the link and the content. No share link exists until you, the account owner, explicitly generate one. Links for IP analyses use unguessable 122-bit tokens plus a separate passcode: the page serves nothing until the passcode is entered, it is meant to travel separately from the link, and Sparlo stores only a salted hash of it — never the passcode itself. Entry is attempt-limited, links expire after 30 days, regenerating rotates the passcode, and revocation is immediate. Even unlocked, a shared page renders the finished analysis only; the underlying disclosure text is never served through the public path.
Is it encrypted?
In transit, TLS for every connection. At rest, AES-256 at the storage layer. Background-job payloads are additionally application-layer encrypted and carry only record IDs — the disclosure text itself never enters the job queue or its logs.
How long do you keep it — and who controls that?
You do. The analysis stays in your account so your work product persists; deletion is user-controlled, immediate, and hard — database row, disclosure text, and uploaded files are purged at once. Upstream, Anthropic deletes API inputs and outputs within 30 days of processing (except where law requires retention or content is flagged under its usage policy).
Which model is this? Are there hidden subprocessors?
Claude, by Anthropic, called directly — no resold model marketplace, no failover chain. Anthropic holds SOC 2 Type II, ISO 27001:2022, and ISO/IEC 42001:2023 certifications, and publishes its trust documentation and subprocessor list at trust.anthropic.com.
For your ethics review
Rule 1.6 — confidentiality
ABA Formal Op. 512 makes "how does this tool handle client data" a mandatory diligence question. This page is the answer: no training, restricted egress, account isolation, user-controlled deletion — stated specifically so your ethics review can verify rather than trust.
Op. 512 — no self-learning system
Opinion 512 attaches an informed-consent burden to self-learning generative AI that retains and reuses prompts across matters. Sparlo is not a self-learning system: your inputs do not train models or carry over into anyone else’s analysis.
37 CFR 11.18 — you remain the practitioner
Sparlo is a co-pilot, not an autopilot. Findings cite verifiable sources — patents and literature you can pull and check — because USPTO guidance is explicit that relying on a tool’s accuracy does not satisfy your reasonable-inquiry obligation. The product is built around your review, not in place of it.
Duty of candor — no new disclosure burden
Under the USPTO’s April 2024 guidance, there is no general duty to disclose AI assistance in a submission unless it is material to patentability. Sparlo produces analysis and strategy support — it draws no legal conclusions, and the judgment calls stay yours.
Sources your reviewer can verify: Anthropic Commercial Terms of Service and certifications (trust.anthropic.com) · ABA Formal Opinion 512 (July 2024) · USPTO, Guidance on Use of AI-Based Tools in Practice, 89 FR 25609 (April 2024). Questions this page doesn’t answer — ask us directly and we’ll answer in writing.
Need this in a form your evaluator can file? We’ll send a one-pager for your executive committee — the egress postures, the no-training terms, and the subprocessor list on a single page.
Get the one-pager for your executive committeeEconomics
An office-action first pass in about ten minutes, for single-digit dollars.
~10 minutes
An office-action first pass runs while you get coffee — not an afternoon of associate time.
Single-digit dollars
The compute behind that pass costs less than a sheet of the letterhead you print it on.
The margin is yours
Under a flat fee, the hours you no longer spend on the first pass stay with you — and senior time moves to judgment.
Example Library
See it on a real patent.
Every analysis below ran on a real, public patent — the same pipeline your disclosure would. Pick one and see the actual work-product: the applications it maps, the prior art it pulls, the claims it stress-tests.
Patent Moat Analysis · Biotech
Nucleic-Acid Sequence Amplification by Primer Extension
Prior-art & non-obviousness analysis — nucleic-acid amplification
8-Section deliverable
Retrieved prior art · sample
- US20220033891A1Amplification with primers of limited nucleotide composition
- JP2004526442A5
- US11466315B2Fast PCR for STR genotyping
- CN114250278BMethods, compositions and kits for capturing, detecting and quantifying small …
- CN116121349BA method for amplifying a nucleic acid sample containing ethanol
+ 23 more references retrieved and screened
Pricing
Priced per report, not per seat.
Start free — your first report is included, no credit card required. Every plan runs every job.
billed monthly
4 reports / mo
For the solo practitioner running the occasional moat analysis.
Start freebilled monthly
10 reports / mo
For attorneys building patent moats as a regular part of practice.
Start freebilled monthly
20 reports / mo
For teams with steady filing volume across matters.
Start freeEvery plan includes
- The full §102/§103 moat report — application map, prior-art differentiation, claim architecture, and an adversarial stress test
- Grounded prior-art citations with one-click verify links
- The disclosure never leaves for outside search and is never used to train a model
- Passcode-gated share links — the shared page serves the finished analysis only, never the disclosure text
- Locked to your account, deleted on demand
Run your next office action through it. Free to start; no client data required to evaluate — use any published application.