The idea is that once the sequence of events is provably intact, you can attach the decision context to it — things like policy snapshots, inputs/prompts (or hashes of them), and state transitions.
Then the evidence layer proves the history wasn’t altered, and analysis tools can reconstruct why the system made a particular decision from that preserved context.
The demo focuses on the integrity layer because without that everything else turns into “trust our dashboard.” Interpretability tools can sit on top of the same evidence
https://guardianreplay.pages.dev/
The demo loads two runs directly in your browser — no signup, no uploads, no network calls after page load.Frank: a conservative agent. Verification returns VALID. Phil: an aggressive agent with tampered evidence. Verification returns INVALID and points to the exact line where the chain breaks.The problem I was solving: when an AI agent does something unexpected in production, the post-mortem...
https://github.com/yupme-bot/kernel-ndjson-proofs
A few days ago I shared a repo with deterministic NDJSON event logs and explicit gap handling (v1.1) and asked for a sanity check.I’ve now updated this with the same two scenarios generated under a newer model (v1.2).This repo still intentionally contains only NDJSON files — no code, no UI:• a clean run • a run with intentional missing dataIn v1.1, missing data was represented via an explicit...
The idea was to see whether the artifacts themselves (two NDJSON logs) are convincing without explanation. If they’re not, I’d really like to know why.
https://github.com/yupme-bot/kernel-v1.1-ndjson-proof
I’ve been working on a small deterministic event-capture kernel and wanted an external sanity check.This repo intentionally contains only two NDJSON files — no code, no UI:a clean runa run where I intentionally removed a persisted segment before exportIn the second file, the system emits an explicit gap marker instead of crashing or silently truncating data, then continues exporting...
I’ve been running the same test multiple times with the same seed and network profile and diffing the JSON. The timings move, but the event structure doesn’t, which helped me reason about cause vs effect more clearly.
If there are existing tools or workflows that already give you this kind of repeatable structure, I’d genuinely love pointers.
https://deterministic-stream-demo1.pages.dev/
I built a small reference demo that streams data in the browser, intentionally stalls the UI, and exports a structured trace you can compare across runs.This isn’t a benchmark or a product. I mostly built it to sanity-check my own understanding of a common problem: is this slowdown coming from the network or from the UI thread?What the demo does:Uses real SSE streaming (no simulated network...