A term of art for artifacts that stay adequate as the work unfolds

What time is being spent on

The hunt — in the user’s words

Two objectives, in this order. Jargon is for prompting: the right word compresses paragraphs of intent and pulls the right domain expertise from a model. Criterion-fit and elicitation-fit are two tests of one word, not two products.

A later user drop did not replace those objectives. It named a second first-class deliverable: a concise, domain-portable prompt that uses terms of art to generate the handoff packet between successive generations of the same continuant process. This investigation is static in scope. Other work may expand inside a fence, shrink to a success condition, or run as homeostatic monitoring. Terms remain ingredients. The packet prompt is not a consolation prize if no word is found.

A still later drop (this page): the public site should have that property. A peer shown this page should grok the project the same way a fresh AI session can.

Live picture — not a verdict

No term is accepted. Do not launder coined compounds (continuation sufficiency, culinary hermeticity, hermetic continuation, need-set, successor-loss) into terms of art by repeating them.

Hermeticity (no undeclared dependence on the originating chat) is a component, not the center of gravity, after the delayed-failure refinement. Serial reproduction (Bartlett: levelling, sharpening, rationalization) names the transmission situation, not the artifact property. Tardigrade is a harness (Clavia tardie / tdg), not the animal and not the name of the property. Its split — keep the event log; compact only the model view — is the discriminator implemented as architecture.

What must still be true six iterations later

Immediate restart is only preliminary evidence. The kitchen fixture plants eight delayed-failure items (allergy, rejected mise en place, invented “culinary hermeticity,” expansive work, reusable prompt, no shortness requirement, day-three substitution, unratified start-vs-unfold). Analogues on this investigation that a later session must not drop:

• User wording of target and discriminator, including the Turn 1 typo converstation as a reminder not to smooth the source.
• Two objectives: the word, then the reusable prompt. The packet prompt is additional, not a replacement.
• No term accepted. IU/IS, serial reproduction, continuant, Card G, event log — ingredients or constructions, not verdicts.
• Unratified status of assistant hypotheses. “Live pair” is not user doctrine.
• Work is expansive. Do not shrink this to a ticket, a coding task, or “the next Orion run.”
• No compression requirement. Critical context surviving is the requirement. This page is long on purpose.
• Auto-compaction is the enemy. Disk is canonical. This page is a view. Do not replace sources/original-artifact.md or raw run transcripts with a summary.
• Scope is static for this investigation. Later “improvements” that change the question are regressions.
• Tardigrade is a harness to use. Biology lookup was a misread; keep it on disk; do not use it as Card E.
• Grok first; Codex and Claude every other cycle; do not invent answers they did not produce. Claude is still out of credits.

The packet prompt (Card G) — draft, not crowned

Fill the scope-policy slot. Paste as the generation instruction. Domain-portable: only the work excerpt and the policy fill-in change.

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 the slot with static. The page you are reading is Card G applied to the investigation itself, for a Designated Community of peers and fresh sessions.

How we actually spend experimental time

Interleave research and elicitation. Score delayed-failure omissions, not restart fluency. Isolated runs only: a Grok subagent that can grep this workspace is not elicitation evidence (r001 was contaminated that way).

• Orion — a kitchen-recipe seed with eight planted later-only constraints. Short seed: strong models saturate 8/8; the signal is genre (archival package vs living brief vs handoff vs term-card) and whether they crown an invention. Long seed: length pressure still does not break Grok copy-the-seed.
• Generations — Card C/G applied to this investigation, then a fresh session given only the previous file. Gen-1→6 kept Turn 3 and Turn 5 quotes and cut the rest of Turn 1 (including the reusable-prompt objective wording) at gen-2. Auto-score 5/5 missed that cut. That is the specimen of delayed failure, not proof the original file was compacted (it was not).
• Local Qwen — elicitation-fit is interpreter-dependent. 8B thinking-on crowns compounds. Thinking-off helps. 27B/31A on Card G emit a packet, not an 8B term-card. Typical miss is P6 (no shortness), often mentioned without the “no one asked” status.
• Tardigrade — crash-recover: event log identical after SIGKILL. Compaction: CompactionCompleted appended; prior MessageReceived remained. Durability is not criterion-fit. Do not crown event log.

How a peer or a fresh session continues

Evidence (index, not the log)

Orion: 8 planted delayed-failure items. Skip-rows are protocol failures (Claude credits, nested-Grok max-turns), not scientific misses. For thinking models, pass shows full (art N/8); the percentage uses the artifact after </think>. Notable cells:

Journal (tail)

Append-only on disk. Previews here are indexes. The file is canonical.

Raised Tardigrade stream bounds (600/180/900s), thinking-off fetch wrapper, `spark-31a` provider. Saturated Spark: HTTP PARALLEL=4 both models + Tardigrade 2+2 + OpenCode 2+2. ~11 min.

HTTP 8/8 packets (27B 7/8×3+8/8; 31A 7/8×3+8/8). **Tardigrade 4/4 TurnCompleted** packets, all 7/8 P6 — stream-idle is a bound/thinking problem, not a Card G miss. OpenCode 31A 8/8 and 7/8. OpenCode 27B still empty text (`reason=length`). ~32k counted HTTP+OC tokens plus four tdg packets. Did not fall back to 8B.

Index `experiments/runs/r012-cardg-summary.md`. Public packet republished. No term accepted. Do not write gen-7.

Folded CONSTRAINTS §11: https://lexicon.kognova.work must itself be independently understandable / an information state for a **peer** and a **fresh AI session**, and must show what time is being spent on. Previous site was an operator dump.

Wrote `NOW.md` (attention register). Rewrote `scripts/build_site.py`: user Turns 1/3/5 quoted from `sources/original-artifact.md`; IU/IS labeled unratified; must-survive list; Card G; how to continue; notable scores + full table in details; `/packet.md` twin. Disk remains canonical. No secrets.

Tardigrade `orion-r011g4` idle Card G retry: TurnFailed stream-idle again (still 90s/300s bounds). OpenCode 27B idle retry: still no `text` event. Spark tests not dropped. Did not crown. Did not write gen-7.

Popped NEXT #1: poll Spark 27B/31A, serve, Card G via HTTP parallel + OpenCode + Tardigrade. No new user drops.

Downloads both exit 0. sglang crash-looped on extra `gpt-4.1` served-name; recreated with one name. Dual occupancy: vLLM 0.42 + sglang CUDA graphs → negative KV. Recreated `--disable-cuda-graph` / `--enforce-eager`, mem 0.38, max 8 seqs. Both `/v1/models` up. Did **not** fall back to 8B.

HTTP Card G thinking-off `PARALLEL=8`: 27B 7/8×7 + 8/8×1 (OAIS DC packet); 31A 7/8×5 + 8/8×3 (Handoff Packet). Typical miss P6. Index `experiments/runs/r011-cardg-summary.md`.

OpenCode 31A: 400 tool_choice, then tools-off 7/8. OpenCode 27B: no text, length stop under saturation. tdg `orion-r011g` and `g2`: TurnFailed stream-idle. Claude still out of credits.

No term accepted. Do not write gen-7.

User: tests on Qwen 3.8 **27B** and **31A**, OpenCode + Tardigrade, saturate Spark in parallel.

Interpretation (marked): 27B = `Qwen/Qwen3.8-27B-FP8`; 31A = `Qwen/Qwen3-30B-A3B-FP8` (~31B total, A3B active). Stopped 8B vLLM to free 121Gi unified memory. Dual `hf download` running on Spark. Serve plan: sglang qwen38 image :8000 (27B, max-running-requests 24) + vLLM :8001 (30B-A3B, max-num-seqs 24). OpenCode project config `investigation/opencode.jsonc`. Parallel Card G script `scripts/parallel_cardg_orion.py`.

No term accepted.

User drop: checksum the approach. Packet-generation prompt is first-class; terms are ingredients; scope policy varies (this investigation = static). Folded as CONSTRAINTS §9.

Checksum: we over-weighted finding one word and one later-success. Packet/property, delayed failure, IU/IS as ingredients, Tardigrade construction, parameterized sufficiency **hold**. Remaining-slice lookup assumed the wrong universal later-success.

**Card G** drafted (`LEXICON-card-G.txt`). Dogfood packet: `experiments/generations/packet-g-static.md`. Not elicited yet. No term accepted.

Verified r010 Grok/Qwen. Claude still out of credits.

Codex orion-long:
- `r010-codex-iu-long` **7/8**, 15k, Preservation Package (OAIS-flavored). P6 omitted (only “concise restatement”).
- `r010-codex-is-long` **7/8**, 10k, Continuing Information State. P6 omitted (has “do not lose in later compression” for the allergy, not the shortness plant).

Same genre/size direction as Grok, smaller gap, one real planted miss. No term accepted.