Wisp Science

The README states that approval gates are enabled by default, Full Permission is opt-in, credentials use the OS keyring, encrypted synchronization is manual, turns can be undone, and explorations are isolated. UI tests also show diff or confirmation stages for promotion, discard, and abandonment, justifying full marks for user confirmation. The product can still read and write project files, execute shells, access SSH hosts, invoke MCP services, and contact model providers; no fine-grained sandbox or per-tool permission policy is shown, so least privilege is 2. Data-flow and sensitive-data claims are clear at a high level, but the evidence lacks a network destination inventory, provider-boundary explanation, log-redaction rules, or credential-lifecycle implementation. External effects and rollback benefit from approvals, diffs, branches, and undo, although recovery after permanent discard is unclear. Dependency controls include an exact ACP version, locked tool installation, and hashed vendored assets, but many Rust requirements are broad, Actions use mutable tags, and no vulnerability-scanning or remediation policy is shown, limiting dependency security to 1. The author, repository, contributors, DOI, signing services, and third-party-notice route are attributed, but the publisher is unverified and the accountable maintenance body is not fully identified, so source attribution is not complete.

The README, Cargo workspace, build workflows, and UI tests describe a coherent desktop research workbench. Tests cover isolated build outputs, vendored-file hashes, and exploration and branch state transitions. Self-consistency is reduced to 2 because the citation advertises v1.2.0 while the Cargo workspace reports 1.3.0. Platform packages, Rust requirements, build tooling, and model configuration paths are reasonably documented, but no offline dependency mirror, compatibility matrix, or prerequisite diagnostic is shown, so dependency availability is 2. The supplied tests check script exit behavior and some warnings, but there is little evidence of user-facing errors, retries, or handling for network, SSH, MCP, and model-provider failures; failure messages score 1.

The audience and scenarios are unusually specific: literature research, Python and R computing, scientific databases, publication evidence, remote execution, and long-running research workflows, earning full marks. Capabilities, BYOM support, approval modes, and isolated explorations are described, but unsupported tasks, model constraints, and hazardous-operation boundaries are not comprehensively stated, so capability boundaries score 2. Explicit @, #, and / interactions, approval gates, and distinct branch and exploration states reduce accidental triggering, with tests confirming several state guards; however, the evidence does not establish natural-language router or tool-selection precision, so trigger precision is 2. Windows, macOS, Linux, WSL, SSH, GPU, Python, and R coverage is explicit and partly backed by build configuration, justifying full environment-fit marks.

The README has strong navigation across startup, research, projects, compute, extensions, and development, earning full information-architecture marks. It provides release packages, platform formats, and configuration links, but the supplied material lacks detailed system requirements, source-build instructions, and troubleshooting, so install notes score 2. Wisp/WISP, Runs, Skills, and Explorations are used consistently, although the citation/Cargo version mismatch prevents full naming-stability and versioning scores. A bundled demo, screenshot, case-study route, and extensive test scenarios provide useful examples, but no visible FAQ is included. Some operational limits are stated, including Full Permission and platform behavior, yet there is no consolidated known-limitations section, so that criterion scores 1. AGPL-3.0-only is consistent across README, Cargo, and the full license text, earning full marks. Releases and version metadata exist, but no changelog content is supplied. The repository owner and contributors are visible, while no explicit maintenance team, support commitment, security contact, or governance responsibility is established; maintenance responsibility therefore scores 1. Unknown publisher identity is not treated as suspicious.

The workbench turns conversations and computation into reusable project files, persistent kernels, run logs, figures, branches, evidence capsules, and publication artifacts. Tests substantiate diff, merge-summary, and evidence-file workflows, supporting full output-usability marks. Combining scientific databases, Python/R, remote runtimes, MCP, tracked history, and publication evidence in one local-first workbench provides clear marginal value beyond a generic chat agent, also earning full marks. Open source, BYOM support, local storage, and a keyless demo reduce adoption cost, but users still bear model charges, environment setup, remote infrastructure, and AGPL compliance obligations; the evidence does not quantify resource use or cost, so cost-benefit scores 2.

Many feature claims point to named documentation, Cargo modules, build workflows, or concrete UI and script tests, giving reasonable traceability. The linked documentation bodies are not included here, however, and the roughly 80 databases, local-storage behavior, and encryption claims lack complete implementation evidence in the supplied files, so claim traceability is 2. README claims are partially corroborated by Cargo, CI, and tests, notably for platform builds, exploration behavior, and dependency integrity, but privacy and security statements remain chiefly README assertions; cross-source corroboration is therefore 2. Product claims, operational guidance, and test assertions are mostly distinguishable, yet broad statements such as data and credentials remaining on the user's machines are not qualified by implementation-level exceptions in this evidence, preventing full fact/inference separation.