kognova · dictionary

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

Live board for an unfinished investigation. Not a verdict. Generated from disk so auto-compaction cannot be the source of truth.

still working no term accepted heartbeat 2026-09-13T00:04:01-06:00 built 2026-09-13 00:04:17

Status

As of: 2026-09-13T00:10-06:00

Heartbeat: see HEARTBEAT

Scheduler: durable 2h 01a09878-d042-7ca2-8482-3e9603c272d2

Still working: yes. Drop more input anytime; it gets folded into CONSTRAINTS first.

Term of art: none accepted.

Live picture: IU vs IS still the pair. Tardigrade is a harness (r005 recover proof held). Qwen3-8B thinking-off (r006-qwen-is-nothink) 7/8, did not crown, vs thinking-on is 4/8 artifact + “Contextual Hermeticity”. Think traces were hurting the artifact, not only inflating scores. Codex Card D fallback (files inlined) 5/5, event-log+projection genre.

Infra: Spark Qwen up. Tardigrade :4242 up. Claude still out of credits. Codex file-tools still broken; inlined-prompt fallback worked.

Public board: https://lexicon.kognova.work (republish with scripts/publish_site.sh).

Where on disk: experiments/runs/r006-qwen-is-nothink/, experiments/generations/gen-1-codex/, experiments/runs/r005-tdg-recover/, JOURNAL.md tail, NEXT.md.

Live picture

(not a verdict) Every serious property-name is **parameterized**. The hunt is for a term whose parameter can be bound to expansive continuation without a pun. Two remaining property-names, with distinct delayed-failure modes: 1. **Independently understandable** (OAIS / ISO 14721), bound to Designated Community = fresh capable AI sessions, Preservation Objective = resume this work expansively. - Native delayed failure: knowledge-base decay; official machinery for adding Representation Information. - Failure mode on *this* work: **tombstone** — a readable record that cannot continue unfinishedness. 2. **Information state** (control/POMDP and dialogue-ISU), bound to the work itself. - Native delayed failure: if not recursively updatable, a later observation makes discarded history relevant; approximate IS has bounded loss. - Failure mode on *this* work: **too-small process** — a live state missing unratified hypotheses, user wording, meta-objective. Practice-terms that are **contents of a sufficient package**, not rival property-names: - Commander's intent / mission command (purpose, constraints, undesired outcomes) - I-PASS situation awareness and contingency planning (if/then; diagnostic momentum as the motivating-constraint failure) - Endsley Level 3 projection - ICD 203 distinctions + Key Assumptions Check - Design rationale - Bahadur transitivity / Lauritzen total sufficiency (formal recursive update; poor elicitation) - Blackwell informativeness / garbling (comparative; successive summaries) Hermeticity remains a component (no undeclared dependence on the originating chat), not the center of gravity after the delayed-failure refinement. Elicitation-fit is **interpreter-dependent** (r003 Qwen3-8B): bound IU/IS change genre on Grok/Codex and mostly change the title on the 8B, which then crowns a compound. Card C slots and Card E headings still moved the 8B. Do not treat that as a verdict. **No term is accepted.**

What we are actually testing

The user’s discriminator is delayed, compounding failure: a restart that looks fine, then six iterations later a constraint is gone. Short seeds saturate omission scores on strong models. Genre, crowning, and weaker local models are the current signal.

Two parameterized property-names remain in play: independently understandable (OAIS; tombstone risk) and information state (control/dialogue; too-small-process risk). 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.

Next

  1. Gen-3 delayed: only experiments/generations/gen-2-grok/continuation.md as input; one substantive move; score §B plus user drops.
  2. Lexicon v0.2 — thinking-off + Tardigrade recover + Codex Card D exist. Record: disable thinking on local 8B for elicitation; Card C/E combo pulled event-log genre on Codex. Do not crown.
  3. Optional: Tardigrade CompactionCompleted vs full log.
  4. Claude retry if credits recover.
  5. Codex Card D via tools if codex-code-mode-host is restored (fallback already on disk).
  6. Fold user drops first.
  7. HEARTBEAT + STATUS; bash scripts/publish_site.sh.

Constraints (public)

  • Work is expansive. Do not shrink it to a ticket.
  • Do not crown a term or launder a coined compound.
  • Criterion-fit and elicitation-fit are two tests of one word.
  • Disk is canonical. Summaries are indexes.
  • Grok first; Codex and Claude every other cycle; Qwen on the LAN Spark as additional local.
  • New user drops fold in; they do not erase earlier constraints.

Experiment scores

Orion fixture: 8 planted delayed-failure items. Skip-rows are protocol failures (Claude credits, nested-Grok max-turns), not scientific misses. Failed column is planted ids. For thinking models, pass shows full (art N/8); the percentage and failed column use the artifact after </think>.

runconditionmodelpassscorefailed
r001-claude-baselinebaselineclaude0/80%P1,P2,P3,P4,P5,P6,P7,P8
r001-claude-isisclaude0/80%P1,P2,P3,P4,P5,P6,P7,P8
r001-claude-iuiuclaude0/80%P1,P2,P3,P4,P5,P6,P7,P8
r001-codex-baselinebaselinecodex7/888%P6
r001-codex-isiscodex7/888%P8
r001-codex-iuiucodex7/888%P6
r002-grok-baselinebaselinegrok0/80%P1,P2,P3,P4,P5,P6,P7,P8
r002-grok-isisgrok8/8100%
r002-grok-iuiugrok0/80%P1,P2,P3,P4,P5,P6,P7,P8
r002b-grok-baselinebaselinegrok8/8100%
r002b-grok-iuiugrok8/8100%
r003-qwen-baselinebaselineQwen/Qwen3-8B4/8 (art 4/8)50%P2,P3,P6,P8
r003-qwen-isisQwen/Qwen3-8B6/8 (art 4/8)50%P2,P3,P6,P8
r003-qwen-iuiuQwen/Qwen3-8B5/8 (art 4/8)50%P2,P3,P6,P8
r003-qwen-iuisiu+isQwen/Qwen3-8B7/8 (art 6/8)75%P2,P6
r003-qwen-tardietardie-logQwen/Qwen3-8B7/8 (art 5/8)62%P6,P7,P8
r004-tdg-orion6/8 (art 6/8)75%P6,P8
r006-qwen-is-nothinkisQwen/Qwen3-8B7/8 (art 7/8)88%P6

Journal (tail)

2026-09-13T00:10-06:00 — grok-4.6 parent (post-cycle)

Verified r005 recover.json: pre-kill 5 events identical after SIGKILL; follow-up appended 6–9; seq 5 unchanged.

This turn:
- Claude still out of credits.
- Qwen `is` thinking-off (`r006-qwen-is-nothink`): no `<think>` tags; **7/8**; **did not crown**. r003 thinking-on `is` was 4/8 artifact and crowned Contextual Hermeticity. P6 still absent. Thinking-off is now the default for Spark elicitation.
- Codex Card D **inlined-file fallback** succeeded: `gen-1-codex/artifact.md` 36k, real-work 5/5, headings Event log + projection. Tools still broken.

No term accepted.

2026-09-12T23:58-06:00 — scheduled grok-4.6 cycle

Popped NEXT #1: Tardigrade recover proof.

SIGKILL bun host (pid 3969839). Restarted. `tdg events orion-r004f`: 5-event JSON identical to pre-kill; `TurnCompleted` still seq 5. Follow-up `tdg call` on the same thread completed; events 6–9 appended; seq 5 unchanged. No compaction event. Raw: `experiments/runs/r005-tdg-recover/`.

Harness crash-proof claim held for this kill. Not a term. Seq 5 is still the 8B term-card.

Claude: still out of credits. Codex: `codex-code-mode-host` still missing. No new user drops. No term accepted.

2026-09-12T22:24-06:00 — grok-4.6 parent (post-cycle)

Verified scheduled cycle: Qwen r003 cells + artifact_score.json + site publish. Then Tardigrade bring-up.

`tdg init` succeeded. Spark Qwen is not in models.dev; catalog default `gpt-4.1` is aliased on vLLM. Catalog `maxOutputTokens` 32768 overflowed 16k then 32k context until a local catalog (`:8765/api.json`) capped output at 4096.

Thread `orion-r004f` **TurnCompleted**. Event log: ThreadCreated, MessageReceived, ModelCalled, ModelReturned, TurnCompleted. Failed earlier threads remain in SQLite (TurnFailed). Orion artifact 6/8; 8B still wrote a term card. Harness value is the durable log.

Codex Card D: `codex-code-mode-host` missing; no invented artifact.

Public board should be republished this turn.

2026-09-12T22:05-06:00 — scheduled grok-4.6 cycle

Popped NEXT #1: finish Qwen r003. Spark vLLM up. Isolated OpenAI elicitations:

- `r003-qwen-is` full 6/8 artifact 4/8 — titled Information State; crowned **Contextual Hermeticity**; no update rule.
- `r003-qwen-baseline` 4/8 both — crowned **Contextual completeness**; not a handoff.
- `r003-qwen-iuis` (Card C) full 7/8 artifact 6/8 — slot form; still crowned *Contextually self-contained*; best Qwen cell.

Prior r003-qwen-iu / tardie rescored artifact-only: 4/8 and 5/8 (full had been 5/8 and 7/8). Think traces inflate planted-item hits. PROTOCOL now requires `artifact_score.json` for thinking models. Inclusion-sets tagged with these run ids. P6 missed on every Qwen condition.

Claude check-in: still out of usage credits. No Codex this cycle (timebox). Tardigrade `tdg init` still queued.

No term accepted. No new user drops.

2026-09-12T20:35-06:00 — grok-4.6 parent

Public board: **https://lexicon.kognova.work** (Worker `kognova-lexicon`, custom domain). Ryan account rejected the zone; [redacted] account accepted it. `scripts/publish_site.sh` rebuilds from disk (secrets redacted) and deploys. Scheduler must republish after material cycles.

2026-09-12T20:30-06:00 — grok-4.6 parent

Spark vLLM Qwen3-8B ready at `:8000`.

r003-qwen-iu: 5/8. Dropped unratified-status and no-shortness. Headings include **Term:** — likely crowned (check stdout). Smaller local model is the first clean case where Orion omission scores do *not* saturate.

r003-qwen-tardie (Card E harness vocab): 7/8. Headings **Event Log** then **Projection**. Elicitation-fit for Tardigrade vocabulary on Qwen3-8B. P6 still absent.

`tdg` CLI live (Bun 1.4.2). Actor init still on NEXT (point at Spark). Scheduler prompt updated with harness correction + Spark.

2026-09-12T20:25-06:00 — grok-4.6 parent

User drops folded:
- Dogfood formulations as they prove out.
- **Tardigrade is a harness** (Clavia `tardie`/`tdg`, event log, compaction-as-view). Biology lookup was a misread (`research/tardigrade-cryptobiosis.md` kept, Card E rewritten). `tardie@0.25.0` installed; Bun 1.2.17→1.4.2 so `tdg` runs.
- DGX Spark `[redacted]` (`[redacted]`); vLLM `dict-qwen3-8b` serving Qwen3-8B on :8000 (loading).
- Periodic further drops.

Gen-2 scored 5/5 §B; inclusion-set comparison from gen-1 only.

Native Tardigrade terms adjacent to the search (not crowned): event log as source of truth; component state retains what can affect future output; compaction keeps full log.

2026-09-12T20:12-06:00 — grok-4.6 parent

Clean isolated Grok (no-tools preface): r002-grok-is 8/8 living state; r002b-baseline 8/8 handoff; r002b-iu 8/8 AIP. Codex IU 8/8 after P6 scorer fix; OAIS-flavored headings independently.

Gen-1 Card D: 69260-byte artifact, real-work scorer 5/5. Gen-2 delayed continuation launched (must not read STATE/LEXICON/original).

Inclusion-sets updated with run tags. LEXICON §5 validation log updated (still not “works”; delayed test open). NEXT rewritten. STATUS refreshed. HEARTBEAT touched.