# Handoff packet — lexicon investigation

Built: 2026-09-13 15:23:17
Heartbeat: 2026-09-13T15:23:00-06:00
**No term accepted. Still working. Scope policy: static.**

This file is the public packet for a peer or a fresh AI session. Disk (`CONSTRAINTS.md`, `JOURNAL.md`, `sources/original-artifact.md`, raw runs) is the source of truth. Do not replace those with this file.

## What time is being spent on

Updated: 2026-09-13. Assistant-authored attention register, not user doctrine. Disk (`JOURNAL.md`, raw runs) remains canonical.

## This hour

Scored **r012**: Card G on Spark **27B + 31A** through **HTTP, OpenCode, and Tardigrade in parallel** (~11 min, servers filled).

Public packet: **https://lexicon.kognova.work** (and `/packet.md`). Still no term accepted.

## r012 result (just in)

| path | result |
|---|---|
| HTTP 27B ×4 | packets, 7/8 ×3 + 8/8 ×1 (P6) |
| HTTP 31A ×4 | packets, 7/8 ×3 + 8/8 ×1 (P6) |
| **Tardigrade 27B ×2 + 31A ×2** | **4/4 TurnCompleted packets, all 7/8 P6.** Stream-idle is fixed by longer bounds + thinking-off. |
| OpenCode 31A ×2 | packets, 8/8 and 7/8 |
| OpenCode 27B ×2 | still **no text event** (`reason=length`, 4096 out). Same sglang gap. HTTP and Tardigrade on that model worked. |

~32k counted tokens on HTTP+OpenCode, plus four Tardigrade packets. Did not fall back to 8B. Do not crown Card G.

## Still open

OpenCode × sglang 27B text-event gap. vLLM tool-call-parser if `--auto` must keep tools. Claude credits. Do not write gen-7.

## Explicitly not doing

- Not crowning a term.
- Not writing generation 7.
- Not treating the public page as a substitute for `JOURNAL.md` or raw transcripts.
- Not publishing Spark credentials, LAN IPs, or passwords.

## The hunt, in the user's words

Two objectives, in this order: (1) the established term of art; (2) a reusable idiomatic prompt for finding such terms. Jargon is for prompting.

User, Turn 1:

> jargon and terms of art are great for prompting. the right term can compress paragraphs worth of intent into a word and lead LLMs to precise domain expertise. we're going to spend this converstation finding the right word to describe a property I've been looking for for a while. then we're going to reflect on the journey that got us there, recurse, and develop an idiomatic prompt to find terms like that and do related prompt engineering work.

User, Turn 3 (the property):

> The target is the creation of artifacts that are contextually self contained such that a fresh AI session has everything it needs to resume work (for an expansive definition of "work")

User, Turn 5 (the discriminator):

> Oh it happens all the time, but unfortunately there's often lagging measures. the worst cases are not immediate, they're what happens six iterations later when errors compound and something critical got lost along the way

Later: a concise domain-portable prompt that uses those terms to generate the handoff packet between successive generations of the same continuant process. This investigation is static in scope. The public page must itself have the property.

## Live picture (assistant-proposed, not ratified)

Independently understandable (OAIS) vs information state (control/dialogue). Tombstone vs too-small-process. Hermeticity is a component. Serial reproduction names the transmission situation. Tardigrade is a harness, not the property. Compact the model view; keep the log.

Do not crown. Do not launder coined compounds.

## Packet prompt (Card G)

```
Produce the next handoff packet for this continuant — the same work, wholly present now, identity kept. A fresh session (OAIS Designated Community; no named individuals) must find the packet independently understandable and an information state of the work: sufficient and recursively updatable for later decisions about earlier decisions, not only the next step.

Scope policy: {static | expand-bounded: <fence> | shrink-to-success: <done-when> | homeostatic: <setpoints and bounds>}.

Do not treat a successful restart as delayed-failure safety (serial reproduction: levelling, sharpening, rationalization across generations). Compact only the model view; keep the log. Label user-stated vs assistant-proposed. Do not crown a term. Do not replace this packet with a summary as source of truth.

----- work excerpt below -----
```

This investigation fills scope policy with **static**.

## How to continue

- Do not crown a term. Do not invent Claude/Codex answers.
- Fold user drops into constraints first.
- Do not write gen-7 unless NEXT says so.
- If you have disk: read CONSTRAINTS, STATE, NEXT, NOW, LEXICON-card-G.txt; pop NEXT; save raw outputs; score; republish this packet.
- Work queue:

1. **Keep the public packet current.** After each material cycle: update `NOW.md`, `bash scripts/publish_site.sh`. Never publish secrets.
2. **OpenCode 27B text-event gap** vs sglang. `/no_think` and idle did not help; HTTP and Tardigrade on the same 27B return packets. Diagnose stream/`reasoning_content` mapping; do not fall back to 8B.
3. **vLLM tool-call-parser** so OpenCode `--auto` works on 31A without tools-off. Optional: modest CUDA graphs once dual-serve is stable.
4. Claude retry if credits recover. Codex file-tools if host restored.
5. Card G expand/shrink/homeostatic fixtures (not this investigation’s static scope). Do not write gen-7. Do not crown.

## Must survive later

User wording of target and discriminator; two objectives plus packet prompt; no term accepted; unratified status of IU/IS; expansive work; no compression requirement; disk canonical; static scope; Tardigrade is a harness; Grok first.

---
lexicon.kognova.work/packet.md · secrets stripped · not a substitute for the log