SShortSingh.
Back to feed

A pipeline that blocks invented statistics in AI-drafted posts about research

0
·4 views

A developer's Research-to-LinkedIn project turns new academic papers into social media posts while guarding against fabricated numbers. The system requires every stated claim to carry a verbatim quote from the paper, and code discards any important number that lacks one. Confidence scores are calculated from verification results rather than self-reported by the model. Drafts must pass deterministic checks that trace every numeral back to the source and flag causal language attached to correlational findings. The author admits the approach cannot confirm that a correct number is attached to the right finding, which is why an LLM fact-checker still runs alongside the code checks.

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)

Log in to join the discussion and vote.

Log in

Related stories

0
ProgrammingHacker News ·

Immigration Attorney Peter Roberts Opens Hacker News AMA on Startup and YC Immigration Work

Immigration attorney Peter Roberts, who represents Y Combinator and startups, has opened an Ask Me Anything thread on Hacker News. He says he will answer questions throughout the day and will be guided by the topics readers raise. Roberts notes he cannot give legal advice on specific cases because he lacks access to all the relevant facts. He asks participants to keep the discussion factual, and the post had 6 points and 3 comments at the time of writing.

0
ProgrammingDEV Community ·

How Vault and DLP keep credentials out of AI agent prompts and runs

AI agents that deploy code or call APIs need credentials, but shortcuts like .env files or pasted keys expose them in prompts, tool calls, logs and command output. The article describes six places a credential can leak and notes that no single control covers all of them. Sandscale's Vault stores secrets once, injects them into a run in memory without writing them to disk, and invalidates them when the run ends, with the value kept out of API responses, logs and audit entries. Its DLP addresses what leaves in prompts and tool calls, but neither control inspects what an agent's own code does inside the sandbox. A secret injected as an environment variable remains readable by any process in that sandbox, so a command like printenv could still reveal it.

0
ProgrammingDEV Community ·

Breaking Dependency Cycles in a Modular Laravel Monolith Without a Rewrite

Modular Laravel apps often start clean but accumulate circular imports as teams add cross-module calls, such as Billing and Orders depending on each other. The author argues this makes tests bloated, changes ripple unexpectedly, and modules impossible to extract. The recommended approach is to measure dependencies with a tool like deptrac, choose a single direction for each relationship, and replace reverse calls with interfaces, domain events, or anti-corruption layers. Teams should record a baseline of existing violations and then enforce the rules in CI so new cycles cannot be introduced. The goal is a one-directional dependency graph rather than zero dependencies.

0
ProgrammingDEV Community ·

SCIM Deprovisioning: Why Disabling a User Must Also Revoke Live Access

An enterprise customer's offboarded employee can keep using a SaaS app after their identity provider account is disabled, because existing session cookies and personal API tokens remain valid. The article argues that SSO only blocks new logins and does not end current access, which is why enterprise customers request SCIM provisioning. It outlines the core SCIM endpoints an identity provider calls, such as creating, updating, and deactivating users via PATCH or DELETE requests. The author stresses that deactivation must block login, revoke all sessions, disable API tokens and pending invites, and log the change in one transaction-safe step. The piece also recommends issuing a separate, hashed bearer token per tenant so each token can only affect its own tenant's users.