CiteWise Uses an LLM for Search Hints Only, Not Verdicts, After a Scoring Bug

·10 views

The developer of CiteWise, a tool that checks whether academic references point to real papers, initially let an LLM guess at missing citation fields and then scored real papers against those guesses. This allowed plausible but incorrect citations to be marked as verified. The fix treats the LLM output strictly as search hints, scores all candidates against the user's originally parsed input, and only replaces the first search result when the second scores strictly higher. The update also retired the POSSIBLE_HALLUCINATION status, since failing to find a paper does not justify that claim. The author additionally found that treating empty search results as a clear outcome was flawed, prompting a rethink of how not-found and insufficient-evidence cases are distinguished.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

VR Game AmblyoPunch Turns Amblyopia Vision Therapy Into a Daily Quest

A developer with moderate amblyopia in the right eye found traditional patching therapy hard to stick with, so they tried AmblyoPunch, a Meta Quest VR app that gamifies vision training. The app presents slightly different scenes to each eye, using Gabor-pattern coins and red spikes that players punch or dodge with hand-tracking or controllers. Its automated four-stage program progresses or demotes the user based on a rolling success rate, with thresholds that a clinician can adjust through a configuration file. The author reported that early sessions were difficult, with success rates around 55%, but that scores rose past 70% within the first week, triggering the next stage. The account comes from a first-person report and does not establish clinical effectiveness.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Stripe webhooks can double-charge in production despite passing test mode

A Stripe integration can pass all test-mode checks yet still cause problems with real customers. In production, Stripe may deliver the same event multiple times, in any order, and retries occur when an endpoint is slow or unavailable. The article recommends four changes: verifying the signature against the raw request body, recording each event ID with a unique key to detect repeat deliveries, returning a fast 200 response while handling work in a background job, and re-reading subscription state from Stripe before updating a user's plan. The piece is accompanied by a disclosure that the promoted clip belongs to a paid billing kit from Syncoders Labs.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Solo Developer Builds JED AI, an Autonomous Cybersecurity Platform on Kali Linux and Python

A developer has built JED AI, a cloud-native cybersecurity platform aimed at replacing fragmented command-line security workflows with a single tool. The system uses an autonomous ReAct agent that takes a high-level target directive, chooses and chains CLI tools in sequence, and analyzes their raw output. Command results stream in real time to a unified dashboard via Server-Sent Events, and a geospatial OSINT module maps domain infrastructure and threat data. The project was created independently, without corporate sponsorship or venture funding, and is described by its author as an effort to make cyber defense tooling more widely accessible. The platform is publicly available at jedcorp.ink, and the developer has invited feedback on which tools to add next.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How Freelance Developers Can Calculate a Minimum Floor Rate

A common approach to freelance pricing is comparing market rates, but that doesn't reveal the least amount a freelancer can accept without losing money. The article proposes calculating a floor rate by adding an income target, self-funded benefits, and business costs, then dividing by realistic billable hours. Using a worked example, the total needed is $75,000 and billable hours are about 1,080, producing a floor of roughly $70 per hour. The author stresses that the billable-time percentage strongly affects the result and that the floor is a minimum rather than a final price. The same figure can then be used to judge whether fixed-price project quotes are worth accepting.

0 comments Read more at DEV Community