In my rerun, Headroom saved 0 tokens on its GSM8K eval prompts
A Source Audit of Headroom v0.39.1: its seeded savings table reproduces, but in my rerun its own GSM8K and TruthfulQA eval requests passed through unchanged.
Fodev JEO
LATEST INVESTIGATIONS
Source audits & field notes
Microsoft Agent Framework Harness Defaults Vary by Language
Compare Microsoft Agent Framework Harness defaults in Python and .NET, including Skills, file access, and configuration-dependent features.
D-073 Makes Each Proposal's Commit Gate Traceable
A source audit of D-073’s trace mirror: it rebuilds each proposal’s checks from committed state and reports gate states only when the hash-bound writer agrees on refusal codes.
Octop's Shell Guard Logs Risky Matches; POSIX Root Is `/`
The latest Octop default logs matching shell rules instead of blocking them, and its local-shell root is / on POSIX; virtual filesystem mode is not a process jail.
TypeSafe's routing verdict hinges on the price ratio
A Source Audit of Diogo Almeida’s TypeSafe coding-agent notes and a PDF built from them: the routing break-even, a withdrawn token statistic, and what the synthesis smoothed over.
Octop's shell guard flags curl | bash, then runs it anyway
A Source Audit of Octop: the README promises approval for risky shell commands, but a fresh install ships approval off and the command guard in log-only mode.
mini-AGI's forgetting fix is measured on reading, not on chat
A Source Audit of mini-AGI: the no-forgetting result comes from a 0.1x trunk rate on a corpus probe, while chat learning is on by default and unmeasured.
GameDevBench's 68.8% top score is the only strictly confined row
A Source Audit of GameDevBench’s leaderboard: the new top row ran inside the strict bubblewrap box added on September 1, while the other 17 rows predate that box.
Context Engineering Is the Missing Layer Between Your Agent and Its Memory
Why smart AI agents fail on game live-ops: separating retrieval, persistent memory, context assembly, and reasoning scaffolds with verifiable code graphs.
Jev shrinks the browser-agent decision loop for faster actions
A Source Audit of Jev Ultrafast and Jev Desktop: the indexed action space, the narrow matched speed result, and the boundary between faster decisions and unproven end-to-end gains.
인기 태그
Search article titles, categories, and tags. Full text is searched when needed.
Type to search articles.