SDI session 2026-04-28: durables to promote into posts

coilysiren commented

2026-05-23 20:55:41 +00:00

Owner

Originally filed by @coilysiren on 2026-04-29T06:42:01Z - https://github.com/coilysiren/website/issues/1330

🤖 Filed by Claude Code on Kai's behalf.

Source: coilyco-vault/inbox/2026-04-28-sdi-session-notes.md. Durables to fold into stochastic-design-iteration.md or sibling posts.

Distill targets

Burn-down between rounds is its own move. Mid-session technique not yet in the SDI post. Between iteration rounds, compress the WIP - strip scaffolding (status banners, conversation-orientation prose, sprawling TODO lists), leave only load-bearing substrate. Burn-down is a third operation in the loop, sitting between human filtering and machine variance generation. Open question: who does it. In session, Kai called for it and Claude executed; plausible split is "human picks the moment, machine does the compression." Worth testing whether the LLM can burn down on its own initiative. Target: new section in stochastic-design-iteration.md or its own short post.
"Where did you learn this" as an SDI grounding technique. When the topic is something the human coined, asking the LLM where it learned the concept reveals whether it is regurgitating prior corpus or actually working from the human's intent. Generalizes to any niche or coined topic. Cheap, fast, diagnostic. Target: SDI techniques sub-section.
Title-carries-intent, applied iteratively. First title (Agentic Pillar Velocity Framework / APVF) failed a guess-the-meaning test. Iterating with a human ("not quite, branch off pillar = structural metaphor") converged on Agent Launch Pillars (ALP). Process shape: human hints at the neighborhood, LLM proposes 3-5 candidates, human picks closest, repeat. LLM cannot converge alone because converging requires intent only the human holds. SDI applied to titles specifically. Target: SDI sub-section or sibling post on titling.

Side observations (not necessarily for posts)

Recursive SDI. The session itself was an SDI loop about a post that names a different SDI-shaped pattern. Interesting for the meta-collection issue (#1327), not load-bearing for the ALP post.
"I do not see the file" failure mode. A markdown link rendered in chat is not filesystem evidence. When the human reports not seeing a file, run ls and git status and report concretely. Operational rule, probably a CLAUDE.md or AGENTS.md note rather than a post.

🤖 Filed by Claude Code on Kai's behalf.

_Originally filed by @coilysiren on 2026-04-29T06:42:01Z - [https://github.com/coilysiren/website/issues/1330](https://github.com/coilysiren/website/issues/1330)_ > 🤖 Filed by Claude Code on Kai's behalf. Source: `coilyco-vault/inbox/2026-04-28-sdi-session-notes.md`. Durables to fold into [stochastic-design-iteration.md](src/pages/posts/stochastic-design-iteration.md) or sibling posts. ## Distill targets - [ ] **Burn-down between rounds is its own move.** Mid-session technique not yet in the SDI post. Between iteration rounds, compress the WIP - strip scaffolding (status banners, conversation-orientation prose, sprawling TODO lists), leave only load-bearing substrate. Burn-down is a *third* operation in the loop, sitting between human filtering and machine variance generation. Open question: who does it. In session, Kai called for it and Claude executed; plausible split is "human picks the moment, machine does the compression." Worth testing whether the LLM can burn down on its own initiative. Target: new section in stochastic-design-iteration.md or its own short post. - [ ] **"Where did you learn this" as an SDI grounding technique.** When the topic is something the human coined, asking the LLM where it learned the concept reveals whether it is regurgitating prior corpus or actually working from the human's intent. Generalizes to any niche or coined topic. Cheap, fast, diagnostic. Target: SDI techniques sub-section. - [ ] **Title-carries-intent, applied iteratively.** First title (Agentic Pillar Velocity Framework / APVF) failed a guess-the-meaning test. Iterating with a human ("not quite, branch off pillar = structural metaphor") converged on Agent Launch Pillars (ALP). Process shape: human hints at the *neighborhood*, LLM proposes 3-5 candidates, human picks closest, repeat. LLM cannot converge alone because converging requires intent only the human holds. SDI applied to titles specifically. Target: SDI sub-section or sibling post on titling. ## Side observations (not necessarily for posts) - **Recursive SDI.** The session itself was an SDI loop about a post that names a different SDI-shaped pattern. Interesting for the meta-collection issue (#1327), not load-bearing for the ALP post. - **"I do not see the file" failure mode.** A markdown link rendered in chat is not filesystem evidence. When the human reports not seeing a file, run `ls` and `git status` and report concretely. Operational rule, probably a CLAUDE.md or AGENTS.md note rather than a post. > 🤖 Filed by Claude Code on Kai's behalf.

coilysiren commented

2026-05-30 05:43:00 +00:00

Author

Owner

Iceboxed in the 2026-05-29 backlog burn-down: Distill-targets brain dump from session notes. Reopen anytime if it becomes real.

coilysiren

2026-05-30 05:43:00 +00:00

closed this issue
added the
icebox
label

Rows
Columns

SDI session 2026-04-28: durables to promote into posts #22

Distill targets

Side observations (not necessarily for posts)