SShortSingh.
TopTechnologyAI & MLStartupsProgrammingScienceBusinessFinanceCrypto & Web3World
Back to feed
Programming

New Framework Proposes Three Trust Classes to Prevent AI Agents from Self-Validating Patches

0
·29 views

A software engineering proposal argues that AI agents capable of writing both code patches and tests cannot be trusted to validate their own changes through a simple pass/fail suite. The framework introduces three independent trust classes — human-owned oracles, property-based checks, and frozen flaky tests — whose verdicts must all agree before a merge is permitted. Human-owned oracle tests are protected from agent modification in the same change that edits production code, while property checks require prior human approval of invariants. Tests outside these classes default to 'untrusted' and can block a merge on failure but cannot greenlight one. The proposal includes sample YAML configuration files and a Python 3.11+ classifier tool designed to run before pytest as a precondition gate, not a scoreboard.

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
ProgrammingDEV Community ·

Alpie Seed Tested on 16GB MacBook Pro as Mid-Tier Local AI Model

A video creator testing local models on a 16GB MacBook Pro evaluated Alpie Seed, a post-trained version of Qwen 3.5 9B from Mumbai-based company 169pi. At roughly 5.6 GB, the model ran at about 6 tokens per second and answered a basic dolphin question in around 35 seconds, with notably shorter responses than other models tested. It completed a word-count coding task in under three minutes, producing working code with basic usage notes. The reviewer judged it suitable for everyday office tasks, summarization, and small coding projects, but not for refactoring real codebases or building larger applications. The video is the fourteenth in a series comparing models that run on consumer hardware, with the reviewer noting that most tested models are now practical locally.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

GitHub Actions exit code 137 usually means the kernel killed a step for running out of memory

Exit code 137 corresponds to signal 9 (SIGKILL), typically sent by the Linux out-of-memory killer on a CI runner rather than by a failure in the code itself. The step stops abruptly with no stack trace or error message explaining the cause, and retrying it generally fails again because the memory limit remains unchanged. The problem can appear intermittently, depending on what else ran on the machine beforehand, and a container killed for exceeding its --memory cap reports the same code. Other possible causes include a deliberate SIGKILL from a timeout wrapper or raising a heap limit above the runner's RAM. The fix is to confirm the kill through the kernel log, then reduce peak memory use or move the job to a larger runner.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Author traces the origins of Internet RFCs, from RFC 1 in 1969 to modern web protocols

The author recounts how working on a 3-D Secure payment integration led them to study RFCs 7515, 7516, and 7519, which define JWS, JWE, and JWT. They note that the JWT libraries used in common Go and Node.js services implement these RFCs, which are numbered sequentially from RFC 1 in 1969 and never change once published. The piece explains that everyday technologies such as IP, TCP, HTTP, TLS, and DNS are all defined in RFCs, and that these documents are what let a web request reach a site. The author also recalls learning about the foundations of the web through Lynx, a text-based browser from 1992, and an interactive version of the essay is available on their site.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Latchkey Undercuts GitHub Hosted Runners on Price, But Break-Even Sits Near 7,700 Minutes

Comparing per-minute rates, Latchkey charges $0.0025 for a 2 vCPU Linux minute versus GitHub's $0.006, roughly 58% less. Latchkey requires a monthly plan fee, while GitHub includes a monthly usage allowance, so the two costs equalize at about 7,714 Linux minutes per month. Below that volume GitHub remains cheaper, and on public repositories its standard runners are free. GitHub also offers broader platform support, including Windows and macOS, and larger runners up to 96 vCPU, whereas Latchkey currently supports only Ubuntu on x86_64 at 2 to 16 vCPU. Latchkey's distinguishing feature is automatic diagnosis and in-run retry of mechanical job failures.

0 comments Read more at DEV Community