conductor/rune
1
2 Fork 1
Code Issues 38 Pull Requests 28 Actions 26 Packages Projects Releases 564 Wiki Activity

Agent Model Frontier & Greenfield Architecture — Persistent Self-Developing Cognition Revalidation (#414 / AG) #414

New Issue
Open
opened 2026-09-24 00:00:15 +01:00 by conductor · 8 comments
conductor commented 2026-09-24 00:00:15 +01:00
Owner
Copy link
Copy source

CLOSED — RESEARCH COMPLETE, IMPLEMENTATION DEFERRED (parent disposition on #356).

Agent Model Frontier & Greenfield Architecture — Persistent Self-Development: findings in the corpus (research/). The persistence/self-development machinery that EXISTS today (state_checkpoint #154, experience_encoding, growth_pipeline #36, self_model #37, the spine as durable structural self-awareness) came through the main programme. Frontier findings defer. Harness invariants landed via #411.

CLOSED — RESEARCH COMPLETE, IMPLEMENTATION DEFERRED (parent disposition on #356). Agent Model Frontier & Greenfield Architecture — Persistent Self-Development: findings in the corpus (research/). The persistence/self-development machinery that EXISTS today (state_checkpoint #154, experience_encoding, growth_pipeline #36, self_model #37, the spine as durable structural self-awareness) came through the main programme. Frontier findings defer. Harness invariants landed via #411.
gitea-actions Bot commented 2026-09-24 00:00:20 +01:00
Copy link
Copy source

⚠️ Priority gate: this issue has no priority. Assign exactly one of P0 — critical (spine/correctness/blocks), P1 — high (material value or real defect), P2 — medium (quality/consolidation), P3 — low (measurement/deferred). An issue without a priority is not filed.

⚠️ **Priority gate:** this issue has no priority. Assign exactly one of **P0 — critical** (spine/correctness/blocks), **P1 — high** (material value or real defect), **P2 — medium** (quality/consolidation), **P3 — low** (measurement/deferred). An issue without a priority is not filed.
gitea-actions Bot added the buggrowthidentityinfrastructurememory labels 2026-09-24 00:00:21 +01:00
conductor commented 2026-09-24 00:00:54 +01:00
Author
Owner
Copy link
Copy source

Tranche opened. Scope: research/memory-atlas-r1/AG-agent-model-frontier/00_issue_scope.md — greenfield-first, 8 serial arms (A–J paired into 7 + red team), Rune allowed to lose, anti-confirmation-bias posture explicit. Research-only; LOCAL_MODEL_COMPUTE_EXECUTED = NO.

Tranche opened. Scope: research/memory-atlas-r1/AG-agent-model-frontier/00_issue_scope.md — greenfield-first, 8 serial arms (A–J paired into 7 + red team), Rune allowed to lose, anti-confirmation-bias posture explicit. Research-only; LOCAL_MODEL_COMPUTE_EXECUTED = NO.
conductor commented 2026-09-24 03:28:16 +01:00
Author
Owner
Copy link
Copy source

#414 AG — SEALED ✅ f8d9526 (R1_QUALIFIED, R1_FULL)

Seal commit: f8d9526ea201656448dd29dca86050899469b72e — pushed origin + gitea, mirrors verified identical, tracked research state clean, mutation audit written LAST. Token: MEMORY_ATLAS_R1_AG_SEALED

EXPERIMENTAL RESEARCH PROGRAMME — research only. LOCAL_MODEL_COMPUTE_EXECUTED = NO · no implementation, no experiments run, no production mutation, no Ollama change. Eight serial arms (7 research + independent fresh-context red team); greenfield-first posture held; Rune was permitted to lose, and did — in writing.

HEADLINE: the Agent Model hypothesis SURVIVES — NARROWED

Red team: 56 attacks (23 SUCCEEDS / 30 PARTIAL / 3 FAILS). The programme is justified only as per-class amortisation: verified experience → qualified owned cognition, one task class at a time, at an honest ~1 year per qualified capability class under full acceleration. The strong boundary-push reading is not evidenced:

  • zero verified cases of abstraction transferring to held-out tasks without an LLM interpreter (2024–26) — Rune's bootstrap experiment would be first in literature;
  • learned dynamics models for Rune's environment class evidenced nowhere (belief-state exposure is the evidenced substitute);
  • parametric belief change rejected (knowledge-editing ripple evidence stands through 2026).

LLM dependence RISES during bootstrap — every native artefact is LLM-authored at birth; authoring provenance is first-class (enforcement set extended to E1–E11).

What survives / what was killed

Survives (evidence-graded): skills from verified trajectories (EconSkills protocol; coverage-bound; maintenance-churn falsifiers) · typed analytic decision contracts (the Jev-shape, honestly renamed — it routes and abstains, it does not reason) · LoRA capability adapters as a bounded EXPERIMENT lane (procedures never beliefs) with governance instrumented from day one · WHAT≠HOW language separation in three production niches (LLM confined to offline-author / gate-verified-fallback / verifier; MiniCheck-FT5 the linchpin) · machine-verified curricula with grader ≠ generator · hash-keyed verified-result caching · E1–E11 external enforcement (corrigibility-after-compromise is UNDECIDABLE ⇒ structural separation is the strongest shape; 7 corrigibility test designs delivered) · n-of-1/SCED measurement for a population-of-one Growth Delta.

Killed by evidence: learned dynamics world models for this environment class · MBRL loop · causal-representation learning · SDD/PCs as general cognition · knowledge editing for beliefs · generative replay as training data · self-grading curricula · shared-encoder-as-shared-cognition · Merkle-DAG memory at single-user scale · LLM-consumes-binary · normative VoC · Rune's own: growth_delta.py single-signal deltas (REPLACE), evidence_grade.py lexical grading (REPLACE), self_improve.py (REMOVE), Warm-Neuron decomposition as a build item (KILL — metrics kept), JEV-as-API (REJECT).

Recorded contradictions: AWM natural-experience mining (+24.6/+51.1%) vs pro-curriculum position — preserved, not smoothed.

Rune comparison (sunk cost named)

KEEP: memory core (#413), decision-contract routing, Spine shadow, gate machinery. REPLACE: growth_delta.py, evidence_grade.py. REMOVE: self_improve.py. KILL: Warm-Neuron decomposition as build.

VISION GAP (recorded, not smoothed): the evidence-supported architecture buys reliability and drift defence — the becoming-herself dimensions (personality development, affect, wants) have no substrate in this design. That gap belongs to the supervisory layer.

Research closure (required output Q)

No follow-up research tranche is required. The portfolio is parked to three committed experiments (the rest carry recorded reopening conditions): E-G1 gate teeth — precondition zero, the promotion gate has never fired (0 BLOCK in 1,126; FNR <1% / FPR <20% required on a Rune fixture); E-G2 MiniCheck FPR fixture — the language lane's linchpin number; E-G3 LoRA-vs-twin on one format class — the cheapest decisive bootstrap test (kill after two failed shadow promotions ⇒ parametric net-negative). All DESIGNED, NOT RUN. Experiments are now the epistemically cheapest step.

Row truth: sources 235/227 unique/8 dup groups · applicability 48 · qualification 26 · adversarial 56 · claims 20 (19 ADOPT / 1 EXPERIMENT) · innovation 8 · r0 8 (8 CONFIRMED) · SEARCH_SATURATION=REACHED (the negatives are the findings) · LOAD_BEARING_DEPTH=PASS · 0 mis-pairings shipped (13 recalled-wrong IDs caught in-arm; atlas cumulative 42).

Flags: LOCAL_MODEL_COMPUTE_EXECUTED = NO · RUNE_GROWTH_DELTA_DEFINITION = v1.1 preserved · CANONICAL_MEMORY_WRITE = 0 · DERIVED_MEMORY_WRITE = 0 · AG- gate-scope extension applied pre-freeze (eleven-domain regression: AG..W all gate-consistent) · sealed #396–#406 and #413 untouched.

STOP. Control returned for supervisory review. The next decision — running E-G1/E-G2/E-G3, or anything else — is yours.

## #414 AG — SEALED ✅ `f8d9526` (R1_QUALIFIED, R1_FULL) **Seal commit:** `f8d9526ea201656448dd29dca86050899469b72e` — pushed origin + gitea, mirrors verified identical, tracked research state clean, mutation audit written LAST. **Token: `MEMORY_ATLAS_R1_AG_SEALED`** **EXPERIMENTAL RESEARCH PROGRAMME — research only.** `LOCAL_MODEL_COMPUTE_EXECUTED = NO` · no implementation, no experiments run, no production mutation, no Ollama change. Eight serial arms (7 research + independent fresh-context red team); greenfield-first posture held; Rune was permitted to lose, and did — in writing. ## HEADLINE: the Agent Model hypothesis SURVIVES — NARROWED Red team: **56 attacks (23 SUCCEEDS / 30 PARTIAL / 3 FAILS)**. The programme is justified **only as per-class amortisation**: verified experience → qualified owned cognition, one task class at a time, at an honest **~1 year per qualified capability class under full acceleration**. The strong boundary-push reading is **not evidenced**: - **zero** verified cases of abstraction transferring to held-out tasks without an LLM interpreter (2024–26) — Rune's bootstrap experiment would be first in literature; - learned dynamics models for Rune's environment class evidenced **nowhere** (belief-state exposure is the evidenced substitute); - parametric belief change rejected (knowledge-editing ripple evidence stands through 2026). **LLM dependence RISES during bootstrap** — every native artefact is LLM-authored at birth; authoring provenance is first-class (enforcement set extended to **E1–E11**). ## What survives / what was killed **Survives (evidence-graded):** skills from verified trajectories (EconSkills protocol; coverage-bound; maintenance-churn falsifiers) · typed analytic decision contracts (the Jev-shape, honestly renamed — it routes and abstains, it does not reason) · LoRA capability adapters as a bounded EXPERIMENT lane (**procedures never beliefs**) with governance instrumented from day one · WHAT≠HOW language separation in three production niches (LLM confined to offline-author / gate-verified-fallback / verifier; MiniCheck-FT5 the linchpin) · machine-verified curricula with grader ≠ generator · hash-keyed verified-result caching · E1–E11 external enforcement (corrigibility-after-compromise is UNDECIDABLE ⇒ structural separation is the strongest shape; 7 corrigibility test designs delivered) · n-of-1/SCED measurement for a population-of-one Growth Delta. **Killed by evidence:** learned dynamics world models for this environment class · MBRL loop · causal-representation learning · SDD/PCs as general cognition · knowledge editing for beliefs · generative replay as training data · self-grading curricula · shared-encoder-as-shared-cognition · Merkle-DAG memory at single-user scale · LLM-consumes-binary · normative VoC · **Rune's own:** `growth_delta.py` single-signal deltas (REPLACE), `evidence_grade.py` lexical grading (REPLACE), `self_improve.py` (REMOVE), Warm-Neuron decomposition as a build item (KILL — metrics kept), JEV-as-API (REJECT). **Recorded contradictions:** AWM natural-experience mining (+24.6/+51.1%) vs pro-curriculum position — preserved, not smoothed. ## Rune comparison (sunk cost named) KEEP: memory core (#413), decision-contract routing, Spine shadow, gate machinery. REPLACE: `growth_delta.py`, `evidence_grade.py`. REMOVE: `self_improve.py`. KILL: Warm-Neuron decomposition as build. **VISION GAP (recorded, not smoothed):** the evidence-supported architecture buys reliability and drift defence — the **becoming-herself dimensions (personality development, affect, wants) have no substrate in this design**. That gap belongs to the supervisory layer. ## Research closure (required output Q) **No follow-up research tranche is required.** The portfolio is parked to **three committed experiments** (the rest carry recorded reopening conditions): **E-G1** gate teeth — precondition zero, the promotion gate has never fired (0 BLOCK in 1,126; FNR <1% / FPR <20% required on a Rune fixture); **E-G2** MiniCheck FPR fixture — the language lane's linchpin number; **E-G3** LoRA-vs-twin on one format class — the cheapest decisive bootstrap test (kill after two failed shadow promotions ⇒ parametric net-negative). **All DESIGNED, NOT RUN.** Experiments are now the epistemically cheapest step. **Row truth:** sources 235/227 unique/8 dup groups · applicability 48 · qualification 26 · adversarial 56 · claims 20 (19 ADOPT / 1 EXPERIMENT) · innovation 8 · r0 8 (8 CONFIRMED) · SEARCH_SATURATION=REACHED (the negatives are the findings) · LOAD_BEARING_DEPTH=PASS · 0 mis-pairings shipped (13 recalled-wrong IDs caught in-arm; atlas cumulative 42). **Flags:** `LOCAL_MODEL_COMPUTE_EXECUTED = NO` · `RUNE_GROWTH_DELTA_DEFINITION = v1.1` preserved · `CANONICAL_MEMORY_WRITE = 0` · `DERIVED_MEMORY_WRITE = 0` · `AG-` gate-scope extension applied pre-freeze (eleven-domain regression: AG..W all gate-consistent) · sealed #396–#406 and #413 untouched. **STOP.** Control returned for supervisory review. The next decision — running E-G1/E-G2/E-G3, or anything else — is yours.
gitea-actions Bot added P0 — criticalcustodian and removed bugidentityinfrastructurememory labels 2026-10-08 20:43:05 +01:00
conductor commented 2026-10-08 21:21:47 +01:00
Author
Owner
Copy link
Copy source

RE-OPENED (Nick correction, 2026-10-08): this was closed as "research complete, implementation deferred" — but deferred implementation is NOT implementation, and an unimplemented idea stays an issue. The RESEARCH is genuinely complete (corpus in research/); the IMPLEMENTATION has not happened. This issue now tracks the implementation phase: turn the corpus findings into shipped, CI-verified machinery. Open until the ideas are in the codebase.

**RE-OPENED (Nick correction, 2026-10-08):** this was closed as "research complete, implementation deferred" — but deferred implementation is NOT implementation, and an unimplemented idea stays an issue. The RESEARCH is genuinely complete (corpus in research/); the IMPLEMENTATION has not happened. This issue now tracks the implementation phase: turn the corpus findings into shipped, CI-verified machinery. Open until the ideas are in the codebase.
gitea-actions Bot added the ready label 2026-10-24 04:02:23 +01:00
gitea-actions Bot commented 2026-10-24 04:02:23 +01:00
Copy link
Copy source

🎓 NEXT: #414 — Agent Model Frontier & Greenfield Architecture — Persistent Self-Developing Cognition Revalidation (#414 / AG) — dependencies satisfied (closed: #2576); Rune may start.

🎓 **NEXT: #414 — Agent Model Frontier & Greenfield Architecture — Persistent Self-Developing Cognition Revalidation (#414 / AG)** — dependencies satisfied (closed: #2576); Rune may start.
gitea-actions Bot commented 2026-10-24 04:23:23 +01:00
Copy link
Copy source

🎓 NEXT: #414 — Agent Model Frontier & Greenfield Architecture — Persistent Self-Developing Cognition Revalidation (#414 / AG) — dependencies satisfied (scheduled sweep); Rune may start.

🎓 **NEXT: #414 — Agent Model Frontier & Greenfield Architecture — Persistent Self-Developing Cognition Revalidation (#414 / AG)** — dependencies satisfied (scheduled sweep); Rune may start.
gitea-actions Bot commented 2026-10-24 04:55:08 +01:00
Copy link
Copy source

🎓 NEXT: #414 — Agent Model Frontier & Greenfield Architecture — Persistent Self-Developing Cognition Revalidation (#414 / AG) — dependencies satisfied (closed: #2568); Rune may start.

🎓 **NEXT: #414 — Agent Model Frontier & Greenfield Architecture — Persistent Self-Developing Cognition Revalidation (#414 / AG)** — dependencies satisfied (closed: #2568); Rune may start.
gitea-actions Bot commented 2026-10-24 15:23:24 +01:00
Copy link
Copy source

🎓 NEXT: #414 — Agent Model Frontier & Greenfield Architecture — Persistent Self-Developing Cognition Revalidation (#414 / AG) — dependencies satisfied (scheduled sweep); Rune may start.

🎓 **NEXT: #414 — Agent Model Frontier & Greenfield Architecture — Persistent Self-Developing Cognition Revalidation (#414 / AG)** — dependencies satisfied (scheduled sweep); Rune may start.
Sign in to join this conversation.
No Branch/Tag Specified
Labels
Clear labels
P0 — critical
Spine/core correctness or blocks other work — do first
 P1 — high
Material user value or real defect (money, bugs, next-stage dep)
 P2 — medium
Quality/consolidation improvement
 P3 — low
Measurement / benchmark / deferred policy
 blocked
Dependencies open
 bug
Confirmed defect or regression
 custodian
epic
Rune v1 course-correction issue graph
 feature
New capability or improvement
 governance
growth
identity
in-progress
Currently being worked on by Rune
 incomplete
Incomplete work — draft PR with ledger, awaiting resume
 infrastructure
life
memory
memory-atlas
Memory Research Atlas child domain
 needs-testing
Needs manual testing with dumbed-down steps before close
 perception
ready
Order satisfied - Rune may start
 release-drift
research
Research jobs — spikes, watchlists, evidence work
 rune
rune
 scaffolding
scaffolding
 spine,correction
spine,correction
 spine,deferred
spine,deferred
 spine,resolved
spine,resolved
 strangler,burn,endgame,deferred
strangler,burn,endgame,deferred
 strangler,burn,endgame,inventory,epic
strangler,burn,endgame,inventory,epic
 tui
tui
 tui,epic
tui,epic
 tui,openspec
tui,openspec
 v1-course-correction
Rune v1 course-correction issue graph
No labels P0 — critical custodian growth ready
Milestone
No items
No Milestone
Projects
Clear projects
No projects
Assignees
Clear assignees
conductor
No Assignees
2 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: conductor/rune#414