Maxwell StudyClaim Gate

The description states that Maxwell StudyClaim Gate is a two-axis verification tool for scientific claims. It evaluates:

• Axis 1: Whether the claim is supported by the study’s content (e.g., endpoint, design-appropriate language, population, timeframe, comparator, outcome status, result direction).
• Axis 2: Whether the study's reported primary outcome matches its registered history, including registry version diffs.

It uses GPT-5.6 for reading and extracting unstructured text, and Codex-generated Python code to enforce deterministic rules and make decisions. The system is designed to separate model-based reading from decision-making, with a tamper-evident SHA-256 receipt binding each decision.

Evidence The author describes the tool as built with OpenAI Codex and GPT-5.6, and states that it includes 107 tests passing from a clean clone.

Inference The system is structured to avoid model-based unilateral verdicts by using deterministic logic for rule enforcement.

The author positions Maxwell StudyClaim Gate as a tool to speed up, improve reproducibility, and increase auditability of scientific claim verification, particularly in medical research. It is described as addressing a gap in current practices where outcome switches in registry data are not caught by standard checks.

It claims to solve a specific failure mode: registry primary/secondary outcome switching that occurs after publication but before the final registry version is updated, which can be invisible when comparing only against the latest registry version.

The tool is positioned as a verification layer, not a replacement for human judgment. It does not allege intent or make claims about the validity of the study itself — it simply states facts and routes unresolved concerns to methodologist review.

Evidence The author cites specific studies (Chan et al., Holst et al., Boutron et al.) to support the prevalence of outcome discrepancies, and references a real case (NCT01401283) where registry outcome roles were switched.

Inference The tool is built around a specific, non-obvious failure mode in scientific publishing — not a general-purpose claim-checker or fact-checking engine.

The description states that the tool is aimed at medical editors, research-integrity teams, scientific communicators, and regulatory and compliance reviewers who manually verify scientific claims today.

It is described as targeting users who need to verify that a claim earned its evidence, especially in cases where outcome switching may have occurred.

Evidence The author explicitly names these user groups and describes their current manual, slow, and inconsistent verification process.

Inference The tool is not intended for general public use or for automated scientific publishing workflows — it is a specialized verification tool for experts in research integrity.

Not evidenced.

The description does not contain any information about pricing, monetization, or business model. It is unclear whether the tool is intended to be sold, licensed, or offered as an open-source prototype.

• The system uses GPT-5.6 for reading and extracting unstructured text.
• Codex-generated Python code handles deterministic decision-making, including schema building, rule functions, registry version diffing, aggregation, and receipt binding.
• It enforces a two-axis separation to avoid conflating claim vs. study with study vs. registry.
• The system includes a tamper-evident SHA-256 receipt, where any change in bound facts invalidates verification.
• It has 107 tests passing from a clean clone, and the architecture is designed around four invariants:
  • No-source-no-clearance
  • Veto-only
  • Two-axis independence
  • Release-gating

Evidence The author describes the technical stack, logic, and testing framework.

Inference The tool is built with a deliberate separation of model-based reading and deterministic decision-making, which supports its stated goal of auditability and reproducibility.

Not evidenced.

There is no mention of revenue, customers, usage metrics, or adoption. The project is described as a hackathon submission and prototype, not a product in active use.

Not evidenced.

The description does not mention any competitors or existing tools in the space of scientific claim verification or registry outcome tracking.

• Prototype vs. scalable tool: The system is described as a hackathon prototype with 107 tests, but there is no evidence it has been scaled beyond a single replayed case.
• Dependency on curated data: The evaluation corpus was built using specific, verified identifiers and constraints — this raises questions about how easily the system can be applied to arbitrary studies.
• Limited scope of registry coverage: The tool is described as focusing on a specific failure mode (outcome switching) and not addressing broader claim-checking issues.
• No commercial or user feedback: There is no evidence of real-world usage, feedback, or iteration beyond the prototype.

1. What are the limitations of applying this system to arbitrary studies beyond the single replayed case?
2. How does the tool handle edge cases where registry data is incomplete or inconsistent?
3. Is there a plan to expand beyond the current focus on outcome switching to broader claim-checking?
4. What is the expected path from prototype to production use, and what are the key technical hurdles?
5. Are there any known issues with model confidence bleeding into verdicts despite the architecture design?