I have released five short papers on one idea: making AI-assisted work re-checkable. All five are archived on Zenodo with DOIs, and source and tests are public.
EMET — a byte-level integrity witness whose verdict {MATCH, DRIFT, UNVERIFIABLE} cannot, by construction, express "trusted." Four implementations, 44 conformance vectors. 10.5281/zenodo.21230267 · source #
BuildLang — a compiler that puts ambient capabilities in the function type and seals re-derivable receipts verified by re-execution. 10.5281/zenodo.21231253 · source #
Witnessed Independence — records whether a verifier graded its own work, and refuses to decide when independence is not positively witnessed. 10.5281/zenodo.21232206 · source #
Proof Packets — an envelope for one agent action whose verdict is derived from checks, so a claim can never vouch for itself. 10.5281/zenodo.21231837 #
Re-Perceived Effects — replaces an actuator's self-report with a re-perceived effect: on-disk hash against intended hash, return code, witnessed bytes. 10.5281/zenodo.21231747
The common thread: a verdict should be a re-derivable function of its inputs, should say plainly what it does not claim, and should fail to UNVERIFIABLE rather than to trust.
More at harperz9.github.io.
source & further reading
dev.to — original article
Claude Fable 5 for Agents: Tool-Call Refusals, Cost vs GLM 5.2
Tagging Environmental Sounds with YAMNet TFLite: An ESC-50 Evaluation over 2,000 Clips
OpenAI's Assistants API shuts down August 26 — but the silent failures hit weeks earlier, when you migrate