T

Triangle IP

The models drive one real capability inside a workflow product that would run without them. Without machine learning, the TIP Tool is still a working board where invention disclosures are captured, scored, merged and moved through stages, with portfolio views, family trees, prosecution timelines, status updates and export. The models power the Predictor, which analyzes claim language against public USPTO prosecution data and returns ranked art units and CPC codes.

It is not a minor attachment either: it is sold as a product at five dollars a month, sits in the main navigation, and feeds the patentability score the platform uses to rank ideas. The examiner and counsel analytics beside it are statistics drawn from USPTO records, not model output.

The Predictor is never validated in public, though what it shows the user is substantial. Each prediction comes with a confidence level and its historical allowance rate, and the source is named as the USPTO's public prosecution data. No accuracy figure, test set or error rate is published, and nothing says how often a predicted art unit turns out to be the assigned one. The nearest statement is a case study claiming a 30 percent rise in allowance rate at an unnamed company.

The product advises and the person decides, and almost none of that is written down. The machine does not act: it returns predicted art units, codes, confidence levels and allowance rates, and a person chooses whether to amend the claims. Nothing published describes that division, a threshold below which a prediction should not be relied on, a review point between a prediction and a drafting decision, or what happens when a prediction proves wrong.

Permissions by portfolio govern who sees a case, not who checks the model. The closest published limit is a scope statement about a different feature: the tool is not a substitute for an attorney's formal docketing system.

One figure and one anonymous customer are published, with no method. A case study is headed as a 30 percent increase in patent allowance rates for a leading global tech company. It describes a firm that could not tell which innovations to back, used examiner allowance rates and typical rounds of prosecution to forecast cost and difficulty, and questioned whether its counsel had the right record in the relevant art unit.

It does not give the company, period, portfolio size, baseline, or how the increase was measured. No customer is named anywhere else on the site: there are no logos, named references or dated deployments, and the reviews the vendor points to sit on a third party software directory.

One precise commitment exists, with no agreement to read it against. It concerns the Predictor and answers the question a patent team asks first: invention text is used only while the prediction runs and then discarded, not stored or reused. Around it sit general assurances on encryption, US hosting, access by role and portfolio isolation. The product terms and privacy policy are reachable only after login and could not be read, so none of this can be checked against anything binding.

The published agreement governs the website only and treats material posted through its interactive features as nonconfidential.

A disclaimer of accuracy stands in for a position on advice. The Terms of Use say the information on the site is for general information only, that the vendor does not warrant its accuracy, completeness or usefulness, and that reliance is at the user's own risk. Nothing says the output is not legal advice or addresses an attorney client relationship. Nothing addresses the supervision question the product raises: a prediction about art unit assignment is used to reshape claim language before filing.

Patent practitioners and enterprise IP teams are addressed directly, and so are individual inventors and startups, the users least likely to have counsel reviewing the result.

No governance position is published about the model. Nobody is named as accountable for it, nothing describes what is tested before a change to the Predictor ships, no retraining cadence or data cutoff is stated, and no principles page exists. The bias question is concrete here. The model learns from historical USPTO assignment and allowance patterns, it is sold on the point that some art units grant at 1 percent and others at 98, and its stated purpose is to steer applications toward the favorable ones.

Whether its predictions hold evenly across technology areas, applicant sizes or drafting styles, and what happens when historical patterns encode something a user would not want to optimize toward, is not addressed. The Predictor page, FAQ, help center, Terms of Use and footer were checked.

Submitted text is addressed specifically, and the rest is generic. The Predictor discards invention text once a prediction is generated. Encryption, US hosting and access control by role are stated, along with secure document exchange inside the platform to keep drafts out of email. No retention period for anything held in the platform, subprocessor list, or incident response or breach notification practice is published, and the privacy policy that governs the product sits behind the login.

Deletion runs the other way, and the vendor says so: patent applications cannot be permanently deleted, only archived, and archived records stay in the system with their linked ideas, tags and analytics.

Liability is handled by a website disclaimer, and the document that would govern the product cannot be read. The published Terms of Use exclude damages of any kind under any legal theory arising from use of or inability to use the site, direct and indirect alike, except for the vendor's gross negligence or willful misconduct. Everything is supplied as is with all warranties disclaimed, and indemnity runs one way, from the user to the company.

Arbitration is available at the company's sole discretion, so it is optional for one side only, under Colorado law and in Colorado courts. Nothing addresses the loss this product could cause, claim language reshaped toward an art unit on a prediction that proves wrong. No indemnity, warranty, service credit, insurance or service level commitment is offered.

Real mechanisms are described, and no named system. A law firm or docketing system can send weekly asset reports or data feeds by email for processing, the vendor's team helps with manual imports at onboarding, and portfolios export to spreadsheet. USPTO status is pulled weekly by default, or daily where outside counsel grants access to a dedicated USPTO customer number through an attorney sponsorship arrangement. Single sign on is supported, and deep links let a case be dropped into an email or chat.

The vendor says it deliberately avoids Outlook, Adobe and DocuSign integrations to move teams away from email workflows. No document management, matter management, billing or electronic signature system is named as a connection, and no API or developer documentation is published.

The region is stated and the tenancy is not. The vendor says data sits on secure cloud hosting on servers in the United States, and the Terms of Use state that the owner is based in Colorado and the site is for use only by persons located in the United States. No statement covers how customers are separated from one another: shared or dedicated infrastructure, a private option, a region choice for buyers with data elsewhere, or a cloud provider.

Separation is described only inside an account, by portfolio. No processing location is stated for the prediction models as distinct from storage.

No independent security attestation is published, and none is claimed. There is no security page or trust center, no SOC 2, ISO or penetration test reference on any page including the Terms of Use, and nothing offered on request. The strongest published phrase is encryption described as enterprise grade, which describes practice rather than anything a third party has examined. The navigation and footer were read to the end, and one search for an attestation returned nothing from the site. This records what is published, not a finding about the vendor's actual security.

The training data is named, and the model is not. The vendor says its predictions come from its own machine learning models trained on the USPTO's public prosecution data, comparing claim language against public patent records to produce ranked art units and codes with confidence levels. Nothing on the site depends on an outside large language model, so questions about a third party model provider do not arise. Not published: architecture, model version, the date of the training data or how often the model is retrained, where inference runs, or any commitment to tell customers when the model changes.

Pricing can be read in full from the site, as set out in the pricing structure: a flat platform rate per organization, a separate Predictor rate, a free plan that never expires and a trial that needs no card. Only a custom plan for extra portfolios is quoted on request. Implementation costs nothing extra for annual subscribers.

Four customer segments have their own pages, and the functional limits are stated directly. The segments are startups, enterprise, individual inventors and patent practitioners, and the product is built for the corporate side of patent work: inventors, product and engineering teams, company IP staff and the outside counsel they invite in. Coverage is US patent prosecution. International filings can be cataloged, tagged and tracked, but allowance rate prediction and examiner analytics cover the US only.

Cost analysis does not work for foreign filings at all, and the vendor calls this a gap across the industry and says it does not oversell capabilities it cannot deliver accurately. Firm size and type on the practitioner side are not described, and government or university technology transfer use is not addressed.