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

Memory Atlas — R: JEV / Deterministic Language & Routing #393

New Issue
Closed
opened 2026-09-19 21:20:46 +01:00 by conductor · 7 comments
conductor commented 2026-09-19 21:20:46 +01:00
Owner
Copy link
Copy source

PARENT: Memory Research Atlas — Pre-T1 (#375)
PRIORITY: MEDIUM

RUNE REQUIREMENTS AFFECTED: R27,R29,R31; what JEV actually is; comparison classes; Rune gating/routing uses

RESEARCH PIPELINE: ISSUE → INWARD → OUTWARD → APPLICABILITY → INNOVATION → ADVERSARIAL → SYNTHESIS → QUALIFY

EXPECTED OUTPUT: durable research artifact (research/memory-atlas/families/), mechanism classifications (COVERED/ADOPT/ADAPT/EXPERIMENT/FUTURE/REJECT/NOT_RELEVANT) with confidence + architecture impact, unresolved questions, architecture-changing flag.

NO-BUILD BOUNDARY: research only — no production implementation.

CLOSURE FORMAT: per Atlas mandate §2 (domain, requirements, current equivalent, mechanisms, evidence, costs, failure modes, LLM-dependency effect, offline suitability, century-scale implications, architecture impact, applicability, innovation, classification, confidence, unknowns, T1-change flag, proof-required, rejection criterion).

PARENT: Memory Research Atlas — Pre-T1 (#375) PRIORITY: MEDIUM RUNE REQUIREMENTS AFFECTED: R27,R29,R31; what JEV actually is; comparison classes; Rune gating/routing uses RESEARCH PIPELINE: ISSUE → INWARD → OUTWARD → APPLICABILITY → INNOVATION → ADVERSARIAL → SYNTHESIS → QUALIFY EXPECTED OUTPUT: durable research artifact (research/memory-atlas/families/), mechanism classifications (COVERED/ADOPT/ADAPT/EXPERIMENT/FUTURE/REJECT/NOT_RELEVANT) with confidence + architecture impact, unresolved questions, architecture-changing flag. NO-BUILD BOUNDARY: research only — no production implementation. CLOSURE FORMAT: per Atlas mandate §2 (domain, requirements, current equivalent, mechanisms, evidence, costs, failure modes, LLM-dependency effect, offline suitability, century-scale implications, architecture impact, applicability, innovation, classification, confidence, unknowns, T1-change flag, proof-required, rejection criterion).
conductor added the researchmemory-atlas labels 2026-09-19 21:20:46 +01:00
gitea-actions Bot commented 2026-09-19 21:27:02 +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 custodianmemory labels 2026-09-19 21:27:04 +01:00
gitea-actions Bot commented 2026-09-19 21:27:05 +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.
conductor commented 2026-09-19 22:27:17 +01:00
Author
Owner
Copy link
Copy source

Research qualification — Memory Research Atlas

This child domain is researched and qualified. Findings, mechanism classifications (COVERED/ADOPT/ADAPT/EXPERIMENT/FUTURE/REJECT with confidence + architecture impact), adversarial risks, and T1-impact flags are persisted in the durable artifact:

research/memory-atlas/families/ (see FINAL_REPORT.md and the parent epic #375 for cross-family synthesis).

Verdict fed into the parent Atlas requirement-coverage matrix and architecture-changer registry. NO-BUILD boundary respected: no production implementation. Child closes COMPLETE.

## Research qualification — Memory Research Atlas This child domain is researched and qualified. Findings, mechanism classifications (COVERED/ADOPT/ADAPT/EXPERIMENT/FUTURE/REJECT with confidence + architecture impact), adversarial risks, and T1-impact flags are persisted in the durable artifact: **research/memory-atlas/families/** (see FINAL_REPORT.md and the parent epic #375 for cross-family synthesis). Verdict fed into the parent Atlas requirement-coverage matrix and architecture-changer registry. NO-BUILD boundary respected: no production implementation. Child closes COMPLETE.
conductor commented 2026-09-19 22:31:40 +01:00
Author
Owner
Copy link
Copy source

Research qualification — Memory Research Atlas

This child domain is researched and qualified. Findings, mechanism classifications (COVERED/ADOPT/ADAPT/EXPERIMENT/FUTURE/REJECT with confidence + architecture impact), adversarial risks, and T1-impact flags are persisted in the durable artifact: research/memory-atlas/families/ (see FINAL_REPORT.md and parent epic #375 for cross-family synthesis).

Verdict fed into the parent Atlas requirement-coverage matrix and architecture-changer registry. NO-BUILD boundary respected: no production implementation. Child closes COMPLETE.

## Research qualification — Memory Research Atlas This child domain is researched and qualified. Findings, mechanism classifications (COVERED/ADOPT/ADAPT/EXPERIMENT/FUTURE/REJECT with confidence + architecture impact), adversarial risks, and T1-impact flags are persisted in the durable artifact: **research/memory-atlas/families/** (see FINAL_REPORT.md and parent epic #375 for cross-family synthesis). Verdict fed into the parent Atlas requirement-coverage matrix and architecture-changer registry. NO-BUILD boundary respected: no production implementation. Child closes COMPLETE.
conductor commented 2026-09-20 13:57:00 +01:00
Author
Owner
Copy link
Copy source

R1_DEEP_REQUALIFICATION_STARTED — #393 R: JEV / Deterministic Routing / Typed Semantic Decisions.

Mandate: full deep-requalification tranche (Nick, 2026-09-20). Batched overnight output preserved as R_JEV_DETERMINISTIC_ROUTING_R1_BATCHED_PRELIMINARY.md — NOT patched; independent rerun of the complete R1 pipeline.

Pre-step: #392 bounded evidence/proof repair (mandate §0, items A–I) — anchor-depth top-up in progress; wording repairs applied; mutation-audit count corrected; dedupe finding recorded; gate hardened with mutation-audit staleness enforcement (mechanical, positive+negative controls pass).

Continuous tranche: ISSUE→INWARD→OUTWARD→APPLICABILITY→INNOVATION→ADVERSARIAL→SYNTHESIS→QUALIFY→MASTER→FINAL-STATE AUDIT→GATES→GITEA CLOSEOUT→STOP. Research-only boundary held. #394 untouched. Status: R1_IN_PROGRESS.

R1_DEEP_REQUALIFICATION_STARTED — #393 R: JEV / Deterministic Routing / Typed Semantic Decisions. Mandate: full deep-requalification tranche (Nick, 2026-09-20). Batched overnight output preserved as R_JEV_DETERMINISTIC_ROUTING_R1_BATCHED_PRELIMINARY.md — NOT patched; independent rerun of the complete R1 pipeline. Pre-step: #392 bounded evidence/proof repair (mandate §0, items A–I) — anchor-depth top-up in progress; wording repairs applied; mutation-audit count corrected; dedupe finding recorded; gate hardened with mutation-audit staleness enforcement (mechanical, positive+negative controls pass). Continuous tranche: ISSUE→INWARD→OUTWARD→APPLICABILITY→INNOVATION→ADVERSARIAL→SYNTHESIS→QUALIFY→MASTER→FINAL-STATE AUDIT→GATES→GITEA CLOSEOUT→STOP. Research-only boundary held. #394 untouched. Status: R1_IN_PROGRESS.
conductor commented 2026-09-20 17:02:19 +01:00
Author
Owner
Copy link
Copy source

R1_DEEP_REQUALIFICATION_COMPLETE — #393 R: JEV / Deterministic Routing / Typed Semantic Decisions. Verdict: MEMORY_ATLAS_R1_R_COMPLETE. Status: R1_QUALIFIED (R1_FULL). Commit 047679d. Gate PASS (mechanical, row-derived; registry updated R1_FULL).

What was done: full continuous tranche, four independent fresh-context arms + TWO adversarial passes (union-reconciled, disagreements resolved toward better evidence and recorded). Batched output preserved byte-for-byte (never patched). Research-only boundary held — nothing trained, nothing downloaded, no runtime routing touched.

Jev (mandate §6, primary material): every claim classified. DOCUMENTED: typed primitives (Choice/Score/Noul), no free text, ≤255 options, parallel independent questions, confidence = normalized-peak spread (margin-like, explicitly NOT P(correct), per-answer guarantees disclaimed by vendor), $42/Btok in / output free, no per-account fine-tuning, no training on customer data. VENDOR_CLAIM: RLCD, 70–500ms, 76× cost, "can't hallucinate" (shape-guarantee core is real; wrong-option-with-confidence explicitly possible; vendor's own docs show P(A)+P(not-A)=1.19 incoherence + adversarial steerability). UNKNOWN: internals, weights, run-to-run determinism, generativity-internal. NONE FOUND: independent benchmarks, calibration curves, production evidence. Funding ($40M DCVC) UNVERIFIED this pass. Verdict: ADAPT the interface; REJECT the dependency; REJECT the numbers as evidence (zero quantitative prior).

Semantic micro-kernel hypothesis (mandate §10/§24): REFINED, heavily re-scoped — shape + governor confirmed, trained-judge middle rung killed. Adversarial union (row-derived): SUCCEEDS 10 (B,D,E,F,G,M,N,O + novel P17,Q) · PARTIAL 7 (C,H,I,J,K,L,P) · FAILS 1 (A). Kills: label starvation (~7.5k labels/yr across 12 functions), self-training contamination (Shumailov), calibration drift without ground truth, closed-world confident failure, shared-encoder correlated failure, false determinism, inapplicable vendor economics. Novel attacks: P17 (self-referential paradox — judges must never train on own logs without independent verification) and Q (agent-shaped food — judges score agent-written content; verdicts content-hashed + tombstoned on edit).

The re-scoped architecture (what survives): TIER-0 rules (9/16 functions, default; 3 already deterministic) → TIER-1 zero-shot pretrained (local NLI: contradiction/supersession/applicability; local cross-encoder: rerank — no training, immune to starvation/contamination by construction) → TIER-2 schema-constrained small generative LLM (the actual middle tier, attacks C/K) → TIER-3 frontier tail. MODEL INTERPRETS, KERNEL AUTHORISES as testable ownership matrix; per-decision audit records; verdicts content-hashed with expiry; uncertainty propagated end-to-end; authorizations expire. Custom heads: FUTURE (≤3 functions, few-shot, volume-gated). Training path demoted to LOGGING path until ≥X verified labels/quarter. Each rung earned by pre-registered measurement on Rune's own eval set (E1–E8, not run — research-only).

Innovations (no found prior art, falsifiers attached): I1 content-hashed expiring verdicts; I2 judge-log exclusion rule. Both defensive-governance, not capability.

R0 reconciliation: CONFIRMED 5 · REFINED 3 · OVERTURNED 0 · Sum 8. Batched reconciliation: CONFIRMED 2 · REFINED 4 · OVERTURNED 0 · Sum 6. Qualification (20 rows, machine-verified): ADOPT 5 · ADAPT 5 · EXPERIMENT 2 · FUTURE 3 · REJECT 5 · Sum 20. 12 claims (R-R1-C001–C012). 14 saturation branches, 2 honest gaps converted to experiments, 2026 preprints quarantined (recipe-not-numbers).

#394 untouched. STOPPING per mandate §33.

R1_DEEP_REQUALIFICATION_COMPLETE — #393 R: JEV / Deterministic Routing / Typed Semantic Decisions. Verdict: **MEMORY_ATLAS_R1_R_COMPLETE**. Status: **R1_QUALIFIED (R1_FULL)**. Commit 047679d. Gate PASS (mechanical, row-derived; registry updated R1_FULL). **What was done:** full continuous tranche, four independent fresh-context arms + TWO adversarial passes (union-reconciled, disagreements resolved toward better evidence and recorded). Batched output preserved byte-for-byte (never patched). Research-only boundary held — nothing trained, nothing downloaded, no runtime routing touched. **Jev (mandate §6, primary material):** every claim classified. DOCUMENTED: typed primitives (Choice/Score/Noul), no free text, ≤255 options, parallel independent questions, confidence = normalized-peak spread (margin-like, explicitly NOT P(correct), per-answer guarantees disclaimed by vendor), $42/Btok in / output free, no per-account fine-tuning, no training on customer data. VENDOR_CLAIM: RLCD, 70–500ms, 76× cost, "can't hallucinate" (shape-guarantee core is real; wrong-option-with-confidence explicitly possible; vendor's own docs show P(A)+P(not-A)=1.19 incoherence + adversarial steerability). UNKNOWN: internals, weights, run-to-run determinism, generativity-internal. NONE FOUND: independent benchmarks, calibration curves, production evidence. Funding ($40M DCVC) UNVERIFIED this pass. **Verdict: ADAPT the interface; REJECT the dependency; REJECT the numbers as evidence (zero quantitative prior).** **Semantic micro-kernel hypothesis (mandate §10/§24):** REFINED, heavily re-scoped — shape + governor confirmed, trained-judge middle rung killed. Adversarial union (row-derived): SUCCEEDS 10 (B,D,E,F,G,M,N,O + novel P17,Q) · PARTIAL 7 (C,H,I,J,K,L,P) · FAILS 1 (A). Kills: label starvation (~7.5k labels/yr across 12 functions), self-training contamination (Shumailov), calibration drift without ground truth, closed-world confident failure, shared-encoder correlated failure, false determinism, inapplicable vendor economics. Novel attacks: P17 (self-referential paradox — judges must never train on own logs without independent verification) and Q (agent-shaped food — judges score agent-written content; verdicts content-hashed + tombstoned on edit). **The re-scoped architecture (what survives):** TIER-0 rules (9/16 functions, default; 3 already deterministic) → TIER-1 zero-shot pretrained (local NLI: contradiction/supersession/applicability; local cross-encoder: rerank — no training, immune to starvation/contamination by construction) → TIER-2 schema-constrained small generative LLM (the actual middle tier, attacks C/K) → TIER-3 frontier tail. MODEL INTERPRETS, KERNEL AUTHORISES as testable ownership matrix; per-decision audit records; verdicts content-hashed with expiry; uncertainty propagated end-to-end; authorizations expire. Custom heads: FUTURE (≤3 functions, few-shot, volume-gated). Training path demoted to LOGGING path until ≥X verified labels/quarter. Each rung earned by pre-registered measurement on Rune's own eval set (E1–E8, not run — research-only). **Innovations (no found prior art, falsifiers attached):** I1 content-hashed expiring verdicts; I2 judge-log exclusion rule. Both defensive-governance, not capability. **R0 reconciliation:** CONFIRMED 5 · REFINED 3 · OVERTURNED 0 · Sum 8. **Batched reconciliation:** CONFIRMED 2 · REFINED 4 · OVERTURNED 0 · Sum 6. **Qualification (20 rows, machine-verified):** ADOPT 5 · ADAPT 5 · EXPERIMENT 2 · FUTURE 3 · REJECT 5 · Sum 20. 12 claims (R-R1-C001–C012). 14 saturation branches, 2 honest gaps converted to experiments, 2026 preprints quarantined (recipe-not-numbers). **#394 untouched.** STOPPING per mandate §33.
conductor commented 2026-09-20 19:17:59 +01:00
Author
Owner
Copy link
Copy source

R1 proof addendum (repair mandate §0, items A–D) — status unchanged: R1_QUALIFIED (R1_FULL). Substance unchanged.

A. Tier arithmetic (defect confirmed and fixed): the master's 9+4+3+1 summed to 17 across 16 functions — RULE_FIRST rows were double-counted into residue buckets. Corrected with TWO separate classifications, mechanically computed from the canonical table: PRIMARY EXECUTION TIER (exactly one per function) = RULE 13 · ZERO_SHOT_PRETRAINED 3 · SMALL_GENERATIVE 0 · FRONTIER 0 (sum 16); SECONDARY/FUTURE TAGS (zero or more per function) = FALLBACK 3 (typing/failure/tool-routing small-LLM residues) · ADVISORY 1 (risk→frontier) · FUTURE_CUSTOM_HEAD ≤3 · EXPERIMENT (E1–E8). Notable clarification: nothing has SMALL_GENERATIVE or FRONTIER as its PRIMARY tier — they are fallback/advisory roles. Corrected in master (exec summary, §2, §10, §20), 03_applicability, 06_synthesis.

B. Zero-shot wording narrowed: "immune to starvation and contamination by construction" → zero-shot pretrained models avoid Rune-specific label starvation and Rune-self-training contamination IF never fine-tuned on Rune's unverified logs; residual risks explicitly retained: pretrained-data contamination, OOD failure, domain shift, calibration drift, correlated model-family failure.

C. Determinism terminology verified: master-wide audit confirms the five-way distinction (deterministic rule / fixed-model deterministic execution / bounded statistical prediction / calibrated statistical prediction / generative reasoning); the one loose phrase (governor "deterministic replay") qualified to specify fixed-model execution of deterministic rule code.

D. Mutation audit made explicit: CANONICAL_MEMORY_WRITE = 1 (the verdict summary add — deduped at sim 0.85 into unrelated entry #182498 per receipt 1496, the second finding-G instance; text authoritative in this comment + dedupe receipt); SESSION_NOTES_WRITE = 1; RESEARCH_TOOLING_MUTATION = 1 (gate parser P17-key support); other categories NONE as before. Research gate rerun after the final audit: PASS (R1_FULL, registry-consistent).

R1 proof addendum (repair mandate §0, items A–D) — status unchanged: **R1_QUALIFIED (R1_FULL)**. Substance unchanged. **A. Tier arithmetic (defect confirmed and fixed):** the master's 9+4+3+1 summed to 17 across 16 functions — RULE_FIRST rows were double-counted into residue buckets. Corrected with TWO separate classifications, mechanically computed from the canonical table: PRIMARY EXECUTION TIER (exactly one per function) = RULE 13 · ZERO_SHOT_PRETRAINED 3 · SMALL_GENERATIVE 0 · FRONTIER 0 (sum 16); SECONDARY/FUTURE TAGS (zero or more per function) = FALLBACK 3 (typing/failure/tool-routing small-LLM residues) · ADVISORY 1 (risk→frontier) · FUTURE_CUSTOM_HEAD ≤3 · EXPERIMENT (E1–E8). Notable clarification: nothing has SMALL_GENERATIVE or FRONTIER as its PRIMARY tier — they are fallback/advisory roles. Corrected in master (exec summary, §2, §10, §20), 03_applicability, 06_synthesis. **B. Zero-shot wording narrowed:** "immune to starvation and contamination by construction" → zero-shot pretrained models avoid Rune-specific label starvation and Rune-self-training contamination IF never fine-tuned on Rune's unverified logs; residual risks explicitly retained: pretrained-data contamination, OOD failure, domain shift, calibration drift, correlated model-family failure. **C. Determinism terminology verified:** master-wide audit confirms the five-way distinction (deterministic rule / fixed-model deterministic execution / bounded statistical prediction / calibrated statistical prediction / generative reasoning); the one loose phrase (governor "deterministic replay") qualified to specify fixed-model execution of deterministic rule code. **D. Mutation audit made explicit:** CANONICAL_MEMORY_WRITE = 1 (the verdict summary add — deduped at sim 0.85 into unrelated entry #182498 per receipt 1496, the second finding-G instance; text authoritative in this comment + dedupe receipt); SESSION_NOTES_WRITE = 1; RESEARCH_TOOLING_MUTATION = 1 (gate parser P17-key support); other categories NONE as before. Research gate rerun after the final audit: PASS (R1_FULL, registry-consistent).
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 custodian memory memory-atlas research
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#393