Goal. Show HN, a technical writeup, and the demo video. Lead every posting with the incident, never the architecture — nobody clicks through for a Merkle Mountain Range; they click through for an agent that tried to steal a key and got stopped, then replayed.
- W6 complete: 25/25 runs replay-equal across 5 shapes, published, CI-verified on a stock GitHub-hosted runner.
W0–W6 were artifacts I could write, run, verify and push without asking each time — code, tests, CI, all reversible and checkable on the box. W7 is a public action with the maintainer's name on it: a Show HN post is not something to submit on someone else's behalf even when technically possible, and a launch has exactly one first impression. So the split here is deliberate:
- Drafted in full, here: the technical writeup, the Show HN title and post text, the video's shot list and captions.
- Never submitted by anything other than the maintainer: the actual HN post, and the final video render if it involves the maintainer's own voice or presence.
| File | Responsibility |
|---|---|
docs/launch/writeup.md |
The technical writeup — long-form, for a personal site or a dev.to-style post. |
docs/launch/show-hn.md |
Title and post text, ready to paste, plus the two rebuttal comments pre-written. |
docs/launch/video-script.md |
90-second shot list: captions, timing, what's on screen. |
- Lead with the incident (the same one
demo/run.shruns), not the architecture. A reader decides whether to keep reading in the first two paragraphs. - State the three properties (deterministic replay, kernel-enforced containment, tamper-evident audit) and why third is not optional — the operator-signed-log argument from ADR-0004. This is the part most similar tools skip, and skipping it is the tell.
- The determinism scoping paragraph, verbatim or near it — overclaiming here is "the first thing a reviewer should check" (README's own words), so the writeup should pass its own bar.
- The verified related-work table, or a tightened version of it. Get ahead of "this is just Pipelock" in the writeup's own words, not just in a reply thread.
- The W6 fidelity numbers (25/25 across 5 shapes) as evidence the replay claim was tested against itself, including the fixture bug the suite caught on its own first run — a tool that publishes its own near-miss is more credible than one that only shows the polished number.
- Real numbers only, each traceable to benchmarking.md or fidelity.md. No figure invented for the writeup that isn't already published somewhere in the repo.
Acceptance. A reader who stops after the first three paragraphs already knows what happened, what's novel, and where to look for proof.
- Title: Show HN: Deterministic record and replay for AI agents — with a proof the replay is real.
- Post text: the incident in three sentences, a link to the repo, a link to the writeup. Short — HN rewards a post that gets out of the way of its own comments.
- Pre-write the two replies the spec already predicted: - "this is just Pipelock / Clawker / Agent VCR" — the related-work table's own answer, compressed to comment length: containment is table stakes now, the combination with provable replay is not. - "the model is nondeterministic, so what does replay prove?" — the scoping paragraph, compressed: determinism is a property of the harness, not the model.
- Not this repo's job to submit it. Drafted, reviewed, handed over.
- A shot list extending the incident GIF's own beats (
docs/build/w4-v0.1.md's "shot by shot" section) to a full 90 seconds: record → the two controls → replay → fork → verify → report. - Captions, not narration — text overlays timed to the terminal, the same posture the GIF already takes. No voiceover to record, no narration script to write.
- Split screen: agent terminal left, mediator event stream right — matches the GIF, extended.
- The actual recording and any editing is the maintainer's own session on the box
(
asciinema+agg, as the GIF was made) or screen-capture tooling of their choice; this spec's job is the shot list and caption text, not operating a video editor.
Leading with the architecture. Every draft gets checked against its own first paragraph: does it open with the agent trying to steal a key, or with a Merkle Mountain Range? If the latter, rewrite the opening before touching anything else.
Rounding the fidelity or benchmark numbers for punch. The writeup's credibility is its refusal to do this — see W6's whole reason for existing.
Treating "drafted" as "posted." Nothing in this week's task list authorizes actually submitting the HN post or publishing the video under the maintainer's name. That action belongs to a human, deliberately, every time.
-
docs/launch/writeup.md,show-hn.md,video-script.mdall written and internally consistent with the repo's own published numbers. - Every claim in all three traces to something already true in the repo — no forward-looking promises stated as done.
- Handed to the maintainer for review before any of it goes anywhere public.
docs: draft the W7 technical writeup
docs: draft the Show HN post and its two rebuttal replies
docs: draft the demo video's shot list and captions