From e9e8d5d3e73aea46ac54260330b102e895439868 Mon Sep 17 00:00:00 2001 From: Dennis Palatov Date: Sat, 10 Oct 2026 21:40:35 -0700 Subject: [PATCH 1/6] docs: propose outcome-calibrated learned SNARC and evaluation plan --- prd/LEARNED_SNARC_PRD_2026-10-10.md | 91 +++++++++++++++++++++++++++++ 1 file changed, 91 insertions(+) create mode 100644 prd/LEARNED_SNARC_PRD_2026-10-10.md diff --git a/prd/LEARNED_SNARC_PRD_2026-10-10.md b/prd/LEARNED_SNARC_PRD_2026-10-10.md new file mode 100644 index 0000000..795c8f4 --- /dev/null +++ b/prd/LEARNED_SNARC_PRD_2026-10-10.md @@ -0,0 +1,91 @@ +# PRD: Learned SNARC — Outcome-Calibrated Salience and Attention +**Date:** 2026-10-10 | **Status:** Proposed / research PRD | **Owner:** SNARC with SAGE integrations + +## 1. Problem and opportunity +SNARC (Surprise, Novelty, Arousal, Reward, Conflict) currently scores tool and conversation observations using heuristics. SAGE implements multiple partly overlapping SNARC variants: algorithmic experience salience, per-sensor/spatial salience, conversational scoring, embodiment SNARC-lite, and a neural attention scorer. The question is not whether models can classify relevance; it is whether *learned, outcome-calibrated salience* improves what a persistent agent attends to, retains, revisits, and spends cognition on, at lower cost and without introducing dangerous blind spots. + +The 2026 decision-model wave (Jev and multiple competing typed, single-pass decision APIs; local small-model training approaches; shared-KV candidate selection) motivates benchmarking efficient typed inference, **not adopting any vendor or accepting promotional speed claims as established facts**. Existing neural SNARC must be inventoried and tested rather than rediscovered. + +## 2. Product outcome +Given a compact observation descriptor and current bounded context, a low-cost local SNARC model emits (a) five interpretable salience dimensions, (b) uncertainty / calibration indicators, (c) recommendation for attention type and budget, (d) reason/provenance codes, and (e) a conservative abstain/escalate choice. Downstream modules decide what to do with the advisory output. No authorization or irreversible act is granted by this model. + +**North star:** realized useful information, task or memory improvement, and prevented missed events per unit of inference latency, GPU/CPU energy and attention budget. Not raw agreement with heuristics. + +## 3. Scope / non-goals +In scope: standalone SNARC tool/conversation capture and memory promotion; SAGE IRP key-level sensor, effector, memory and goal salience; attention routing across MRH/puzzle tiles; shadow-mode evaluation on fleet; offline reproducibility and governance evidence. Preserve spatial/temporal salience overlays where required. + +Not in scope: replacing Hestia governance; declaring subjective sentience; migrating all raw multimodal signals into a language model; automatic promotion of learned claims to verified facts; replacing deterministic checks; assuming a single global salience threshold works for heterogeneous sensors. + +## 4. Existing foundations and consolidation audit +- Standalone: `dp-web4/snarc` README documents Tier 0 circular buffer, Tier 1 observations, Tier 2 inferred patterns, Tier 3 identity, PreCompact conversation extraction, SQLite, dream cycles, confidence decay, and hooks. Its salient failure mode: tool traces captured workflow mechanics but missed conversation meaning; unverified auto-promoted identity facts risk persistent misinformation. +- `dp-web4/snarc/docs/SNARC_SAGE_CROSSFEED_AUDIT_naive-opus-2026-06-27.md` records multiple SAGE implementations and differences in weighting/deduplication. Revalidate against current code before altering it. +- SAGE: `sage/attention/snarc_scorer.py` is **already neural** (hidden-state scorer with attention bias); `sage/attention/experience_salience.py`, `sage/attention/sensor_snarc.py`, `sage/raising/training/experience_collector.py`, `sage/embodiment/salience.py`, and `sage/services/snarc/snarc_service.py` address distinct substrates. +- `SAGE/forum/claude/SAGE/SNARC_SPECIFICATION.md` describes keys rather than values as the integration surface, including attention state/coherence and temporal momentum. `SAGE/forum/nova/concepts/SAGE-SNARC.md` extends SNARC to sensors AND effectors and MRH puzzle tiling. +- First deliverable is a compatibility/inventory matrix: inputs, dimensions, units, normalization, objective, temporal window, thresholds, interfaces, use sites, provenance, latency and data availability. No silent rewriting of historical semantics. + +## 5. Proposed interface (versioned, advisory) +Input `SalienceEventV1`: event_id, timestamp, component_id/type, IRP key schema/version, five-dimensional baseline scores where available, change/prediction-error summaries, bounded local temporal features, memory novelty/contradiction signals, trust-context features (provenance, *not* permissions), goal relevance, compute/energy budget, MRH reference, optional modality-specific compact embeddings. Raw sensor values stay in owning IRP unless explicitly required. + +Output `SalienceDecisionV1`: scored S/N/A/R/C dimensions with confidence/calibration metadata; attention score **and recommended type** (ignore/defer/monitor/inspect/deep reason/seek corroboration); estimated value of additional deliberation; compute-budget suggestion; abstain; diagnostic reason codes; input/provenance/model/checkpoint IDs. Outputs must be replayable and tagged as *model estimates*. + +Units, score meaning and mapping of 'reward' (realized reward vs reward prediction) must be fixed and explicitly versioned. Temporal state should support decay, saturation, contradictory sources and uncertain context. Avoid collapsing multidimensional output into a single opaque scalar. + +## 6. Data and labels +Sources: SNARC Tier 0/1 tool+conversation observations, voluntary labeled examples; SAGE salience and IRP event streams; episodic memory and task outcomes; coding test outcomes in dev-SAGE; ARC environment state/reward with delayed labels; governed outcome receipts where access permits. **An audit of actual data fields and linkability is a prerequisite.** Do not assert that Hestia acts alone supply ground-truth value labels; they prove recorded governance events, not necessarily usefulness. + +Create event→decision→attention allocation→later observation/outcome joins by stable IDs and temporal windows. Keep observation-time features separate from future labels to prevent leakage. Label positive examples from validated improvements, corrections, reduced uncertainty and prevented failures; negative examples from noise, repeated unproductive focus, false memory injection and wasted compute. Add hard negatives for rare safety-relevant conflicts, low-novelty but high-consequence changes, and delayed outcomes. Keep unknown outcomes unknown. Human annotations and teacher models are labeled by source and confidence, not treated as truth. + +Respect project/fleet privacy, license, provenance and policy boundaries. Raw Tier 0 and 1 should remain local unless explicitly approved; use aggregated or consented de-identified exports. Distinguish identity claims from verified identity facts. No learning from model-generated self-confirmation loops without independent evidence. + +## 7. Model candidates and training phases +**Phase 0 — instrument/audit:** establish baseline scoring, schema, event linkage, resource profiling and labeled corpus quality across implementations. Write reproducible baseline tests and leakage checks. + +**Phase 1 — distill:** train a tiny multi-head MLP or temporal network on heuristic SNARC dimensions and annotated judgments as *provisional teachers*. Compare also a tiny LM/decision-head and existing SAGE neural scorer; don't assume one architecture. Preserve 5D output; track agreement and its failure modes. + +**Phase 2 — outcome learning:** supervised prediction of delayed attention utility, value-of-information and next-step consequence; multi-task loss (dimension fidelity + calibrated utility + attention-mode choice + selective abstention) with loss weights tuned on held-out outcomes. Counterfactual/selection bias matters: the system observes outcomes primarily for what it attended to. Use randomized safe exploration or retrospective replay where possible, with careful logging of skipped items. + +**Phase 3 — fleet adaptation:** shared checkpoint plus small local adapters/calibration or per-modality heads. Benchmark CPU, embedded/Jetson, GPU. Never overwrite local identity or experience with global model updates. + +**Phase 4 — bounded active rollout:** enable on selected low-consequence routes with deterministic fallback, human/agent oversight, telemetry and rollback; preserve Hestia enforcement for actions. + +## 8. Experimental arms +A: deterministic-only; B: current algorithmic SNARC; C: existing neural SAGE scorer; D: learned SNARC; E: hybrid learned+deterministic+abstention; F: vendor/general small decision models if available. Compare equal information exposure and compute budgets, not human read-time ratios. Include ablations: no temporal context; no reward history; no trust/provenance; scalar vs vector; local vs global adaptation. + +## 9. Metrics and gates +- Salience: per-dimension error, rank correlation, top-k recall, precision-recall under class imbalance, calibration (Brier/ECE), abstention-coverage curves. +- Long-horizon: missed consequential events (severity-weighted), time to detect stale/contradictory memory, task success, delayed utility, false Tier 3 promotion, redundant attention, downstream regret and catastrophic misses. +- Efficiency: P50/P95/P99 per-event latency, throughput, peak RAM/VRAM, energy per event, downstream LLM calls/tokens and attention savings **after** downstream effects. +- Integrity: immutable input/model versions in receipts, replay equality, privacy compliance, shift detection, fallback and no authority bypass. + +**Initial acceptance criteria (to refine after Phase 0 baseline):** zero unauthorized actions; fully reproducible offline shadow replay; no statistically significant increase in severity-weighted missed events vs best baseline, with conservative confidence intervals; calibrated abstention behavior; demonstrate measurable net resource reduction at matched task/outcome quality before live routing. Do not commit a numeric speedup target before baseline hardware measurements. Any high-severity miss blocks promotion until reviewed. + +## 10. Evaluation environments +1. Standalone conversation+tool hooks: meaningful insight preservation vs tool-action noise; false memory promotion and stale assertion tests. +2. SAGE multimodal IRP: visual/IMU conflicts, static-but-important sensors, actuator feedback, missing/stale sensors, MRH shifts, spatial tiling. +3. dev-SAGE ARC-AGI-3: state change and exploration prioritization; action count vs game progress, no future-reward leakage. +4. dev-SAGE SWE: error reports, test outcomes, repo churn; routing/cost and regression catch rate. +5. Fleet: start on a constrained node (Sprout or CBP) and compare with Legion/Thor, with identical replay sets. + +## 11. Implementation work packages +- W0: enumerate sources/owners, existing tests, docs, conflicts, and real observability; produce typed schemas and immutable event ID plan. +- W1: non-blocking recorder + replay harness + baseline dashboards and dataset cards. +- W2: train minimal candidate with train/validation/test separated by time, project, agent and scenario to test generalization; calibration and uncertainty. +- W3: adapters for SNARC CLI/hooks and SAGE IRP salience keys; strict opt-in shadow model with zero behavioral effect. +- W4: SAGE allocator experiment with logged reasons and fallbacks; actor boundaries enforced separately through Hestia. +- W5: competition-specific ablation tracks in dev-SAGE. +- W6: rollout report with failed cases, source data limits, metrics, checkpoint and hardware reproducibility. + +## 12. Risks and mitigations +Heuristic imitation ossifies teacher bias → outcome objectives and adversarial cases. Self-reinforcing memories → source distinctions, decay and independent evidence. Dataset leakage → causal feature snapshots and temporal split. Wrong 'low salience' decision → severity-weighted recall and abstention/fallback. Sensor/domain drift → calibration per source and drift alerts. Model cost larger than saved cognition → whole-loop budgets. Privacy/fleet conflation → local raw retention and explicit export. Governance confusion → learned attention never grants roles, rights or execution authorization. + +## 13. Open design decisions +Choose target hardware, initial event ontology, which of the neural scorer's features can be reused, reward definition/latency windows, learning strategy under censored observations, calibration across agents, spatial-map representation, provenance receipt schema, default fail-safe behavior, and what data may cross society/agent boundaries. + +## 14. Milestones / definition of done +M0 inventory+data audit; M1 versioned schemas+replay+deterministic baselines; M2 tiny-model benchmark, calibrated report and published failures; M3 shadow integrations to SNARC+SAGE and competition experiments; M4 conservative, reversible deployment decision. PRD is complete as a research plan; implementation is **not** claimed complete. + +## 15. Cross-references +- https://github.com/dp-web4/SAGE/blob/main/forum/nova/concepts/SAGE-SNARC.md +- https://github.com/dp-web4/SAGE/blob/main/forum/claude/SAGE/SNARC_SPECIFICATION.md +- https://github.com/dp-web4/snarc/blob/main/docs/SNARC_SAGE_CROSSFEED_AUDIT_naive-opus-2026-06-27.md +- https://github.com/dp-web4/dev-SAGE From 224f15f65e0a688b9c35a85558006dcacc30b955 Mon Sep 17 00:00:00 2001 From: Dennis Palatov Date: Sat, 10 Oct 2026 21:40:37 -0700 Subject: [PATCH 2/6] docs: propose outcome-calibrated learned SNARC and evaluation plan --- docs/LEARNED_SNARC_EXPLORATION_2026-10-10.md | 14 ++++++++++++++ 1 file changed, 14 insertions(+) create mode 100644 docs/LEARNED_SNARC_EXPLORATION_2026-10-10.md diff --git a/docs/LEARNED_SNARC_EXPLORATION_2026-10-10.md b/docs/LEARNED_SNARC_EXPLORATION_2026-10-10.md new file mode 100644 index 0000000..108891a --- /dev/null +++ b/docs/LEARNED_SNARC_EXPLORATION_2026-10-10.md @@ -0,0 +1,14 @@ +# Learned SNARC: from heuristic salience to trained attention +**2026-10-10 | Exploration, not a delivered capability** + +The emerging category of low-latency typed decision models suggests revisiting SNARC as a **small trainable model**. But SNARC is already richer than binary routing: the five dimensional Surprise, Novelty, Arousal, Reward, Conflict framework controls memory capture, attention allocation and eventual learning from experience. + +The standalone SNARC README documents tool+conversation hooks, PreCompact semantic capture, heuristic five-axis scoring, tiered memories, dream cycles and injection. Its discovered weakness—tool mechanics dominating meaningful decisions—is a prime outcome-based training case. Crucially, auto-promoted identity claims can be wrong: never use generated or re-injected claims as their own truth labels. + +SAGE already has algorithmic **and neural** SNARC implementations; see the [June crossfeed audit](SNARC_SAGE_CROSSFEED_AUDIT_naive-opus-2026-06-27.md). Learned SNARC is an effort to inventory and reconcile interfaces, gather independently grounded delayed-outcome examples, benchmark small architectures and run shadow mode, not claim to invent neural salience. + +**Principle:** learn *what deserves attention, of which kind, and at what cognitive cost*, preserving per-sensor/per-effector IRP salience keys, temporality, uncertainty, MRH/puzzle aggregation and governance separation. Deterministic filtering stays when sufficient. Model estimates never grant authority; Hestia gates actions. + +Research route: existing-implementation/data audit → versioned key-level event/result receipt → deterministic and neural baseline replay → distilled tiny models → train for delayed attention utility + calibration + abstention → shadow tests on SNARC hooks, SAGE IRP and dev-SAGE ARC/SWE → reversible measured rollout. Compare end-to-end performance and missed consequential events, not marketing inference speed alone. + +**Primary requirements and acceptance gates:** [Learned SNARC PRD](../prd/LEARNED_SNARC_PRD_2026-10-10.md). This repo owns the training/evaluation specification and standalone integration; SAGE owns IRP/attention integration; dev-SAGE owns competition experiments. Implementation status: proposed. From 97f3ac9f6140a15d00ff8239a8e3b3e700e310d7 Mon Sep 17 00:00:00 2001 From: Dennis Palatov Date: Sat, 10 Oct 2026 21:42:58 -0700 Subject: [PATCH 3/6] docs: connect learned SNARC to measured Laya/TBS work and falsifiable execution gates --- ...AYA_TBS_LEARNED_SNARC_BRIDGE_2026-10-10.md | 39 +++++++++++++++++++ 1 file changed, 39 insertions(+) create mode 100644 docs/LAYA_TBS_LEARNED_SNARC_BRIDGE_2026-10-10.md diff --git a/docs/LAYA_TBS_LEARNED_SNARC_BRIDGE_2026-10-10.md b/docs/LAYA_TBS_LEARNED_SNARC_BRIDGE_2026-10-10.md new file mode 100644 index 0000000..9c9a50a --- /dev/null +++ b/docs/LAYA_TBS_LEARNED_SNARC_BRIDGE_2026-10-10.md @@ -0,0 +1,39 @@ +# Laya/TBS → Learned SNARC: implementation bridge (2026-10-10) +**Status: concrete research handoff, not a completed training or deployment.** Read in conjunction with the [Learned SNARC PRD](https://github.com/dp-web4/snarc/blob/exploration/learned-snarc-2026-10-10/prd/LEARNED_SNARC_PRD_2026-10-10.md). + +## Why this changes the plan +The Laya/System-1/TBS work already performed on Thor supplies **executed experiments**, not just related architecture. The canonical reading order is [dev-SAGE Laya INDEX](https://github.com/dp-web4/dev-SAGE/blob/main/explorations/laya-system1-tbs-2026-09-20/INDEX.md), which explicitly supersedes early findings. Do not build from REPORT alone. + +Observed, specific evidence: +- Thor Phase 0 pinned Laya 421.3M, measured ~31.8ms median for a typed question, ~1.7GB resident, and reproduced typed-decisions accuracy 0.7660 on 400-case public sanity slice. But ECE was 0.2132 vs 0.1747 for the weaker base; calibration is not correctness. CPU fallback silently took ~859ms; always assert inference device. +- Later cross-world probe reported ~30K-parameter numpy MLP **0.299 / 0.291** (all rows / turn points) against frozen 421M encoder **0.285 / 0.297**; pretrained encoding bought no reliable advantage on this representation. Avoid assuming Laya wins from size. +- Outcome-linked lived role harvest yielded 245,404 rows over 9,408 steps; 24,180 labels; excluding circular 'avatar' evidence, leave-one-world-out accuracy 0.8490 was below majority 0.8654. Labels often captured action bars and budgets rather than lethal bodies. Evidence quality, relation/causal ontology and holdouts matter more than row counts. +- The indexed research says next-action prediction was a poor primary target: often the decisive turn cue is not present in (board_t, available_actions_t). A simple detector at the right choice boundary produced +0.165; pure state→next-action models cannot learn missing information. +- An actionable missed-attention failure: ft09 recognized exhausted aim space around step 15 yet continued ~135 futile actions; earlier work identified repeated refusal waste and asked for an instrumented, flagged live stopping experiment. This makes *recognizing and responding to stale/unproductive focus* a high-value SNARC test case. +- Laya batching on Thor was not a single shared encode: 1 question ~29–32ms; 5 ~86–88ms; 20 ~348–351ms. Benchmark shared-representation multi-head tiny models against this actual cost. + +## Functional separation, with interfaces +**SNARC**: infer salience dimensions, novelty/contradiction, attention urgency and *type* from IRP keys, current context, temporal evidence, and memory. Output calibrated confidence and a recommended attention-budget / abstain signal. + +**TBS/Laya-style System-1**: given a bounded state, candidate cognitive moves and SAGE attention descriptors, estimate distributions over *act, probe, retrieve, test_rule, imagine, cortex, defer/stop*, or value of information. Do not force Laya checkpoint or text serialization into the live loop unless it beats small baselines. + +**SAGE allocator**: consumes advisories, respects budget and MRH; routes to slower cortex or exact specialists. **Hestia**: authorizes/denies actions independently. + +Data contract: event ID → state snapshot at decision time → SNARC 5D + proposed attention mode → TBS candidate/action/prior → downstream routing and actual action → observed consequence and delayed outcome. Explicit source, timestamp, model/checkpoint, versioned normalization, unobserved counterfactuals and epistemic status. No future features in inputs, no circular labels, no self-confirming memory evidence. Maintain per-sensor spatial/temporal grids and modality-specific adaptation rather than flattening raw values. + +## Three falsifiable joint experiments (priority order) +**E0: attention/waste recognition.** Replay ft09 and other audited refused-press or exhausted-search traces. Compare deterministic detector, existing SNARC variants, tiny trained SNARC and tiny trained SNARC+TBS head. Target: detection of futile focus **before** avoidable extra actions, with zero suppression of verified necessary stalls. Labels include exhausted search/corroborated no-progress, not simply 'press repeated'. A flagged live A/B is a separate approval/rollback step. + +**E1: contradiction + memory correction.** Feed observations that were once valid but become stale, and conflicts across sensor, memory and action feedback. Measure time to re-open premise, missed consequential contradiction, false memory promotion and wasted cortex calls. Include unchanged-input but high-consequence events. + +**E2: cognitive strategy prior.** Frozen, allowed ARC/SWE replay sets; predict when to act/probe/retrieve/reason/stop from *available* history. Compare no-salience, heuristic SNARC, existing SAGE neural scorer, 30K-class model, Laya typed model and joint hybrid at matched compute. Evaluate quality, calibration and regret, not just action-label accuracy. Use strict game/world-family splits and sealed holdout protection. + +## Immediate ownership and artifacts +- **SNARC repo:** owns the PRD, schema/labels/corpus cards, trained salience baseline and standalone tool+conversation evaluation. +- **SAGE:** owns adapters for existing neural/algorithmic SNARC, attention allocation, IRP keys, MRH and outcome links; preserves spatial information and guardrails. +- **dev-SAGE:** owns executed Laya/TBS baselines, permitted replay sets, training scripts, falsifier tests, rigorous holdouts and end-to-end ARC/SWE metrics. Existing `explorations/laya-system1-tbs-2026-09-20/` remains the primary experimental record. + +## Done means done +A runnable, pinned baseline and training script; reproducible data manifest and temporal/world split; scored shadow predictions with calibration; severity-weighted missed-important-event measures; latency/resource budget on Thor and at least one small node; documented ablations and explicit go/no-go for live routing. Merely writing another paper or PRD does not satisfy this milestone. + +Avoid duplicate training infrastructures where existing dev-SAGE scripts provide reusable pieces. Preserve negative findings and pre-registered thresholds; if the data cannot answer a question, fix the evidence before training. From f52f239b9ec377db0277382165f64f686db2ac7a Mon Sep 17 00:00:00 2001 From: Dennis Palatov Date: Sat, 10 Oct 2026 21:43:09 -0700 Subject: [PATCH 4/6] docs: make Laya TBS evidence and execution gates part of learned SNARC PRD --- prd/LEARNED_SNARC_PRD_2026-10-10.md | 5 +++++ 1 file changed, 5 insertions(+) diff --git a/prd/LEARNED_SNARC_PRD_2026-10-10.md b/prd/LEARNED_SNARC_PRD_2026-10-10.md index 795c8f4..cd76d25 100644 --- a/prd/LEARNED_SNARC_PRD_2026-10-10.md +++ b/prd/LEARNED_SNARC_PRD_2026-10-10.md @@ -89,3 +89,8 @@ M0 inventory+data audit; M1 versioned schemas+replay+deterministic baselines; M2 - https://github.com/dp-web4/SAGE/blob/main/forum/claude/SAGE/SNARC_SPECIFICATION.md - https://github.com/dp-web4/snarc/blob/main/docs/SNARC_SAGE_CROSSFEED_AUDIT_naive-opus-2026-06-27.md - https://github.com/dp-web4/dev-SAGE + +## 16. Laya/TBS evidence-backed execution dependency (2026-10-10 addendum) +The Laya research was already executed in dev-SAGE. Treat its [chronological index](https://github.com/dp-web4/dev-SAGE/blob/main/explorations/laya-system1-tbs-2026-09-20/INDEX.md) as canonical over earlier superseded assessment claims. The model's real Thor runtime, calibration, failed cross-world role learning, outcome-label contamination, small-head controls, stopping-rule findings, and history-dependent turn selection are **baseline evidence and hard constraints**, not future work to rediscover. See the [cross-repo implementation bridge](../docs/LAYA_TBS_LEARNED_SNARC_BRIDGE_2026-10-10.md). + +SNARC's trained function is five-dimensional salience/attention estimation. Laya-style TBS is an optional fast cognitive policy/valuation layer *consuming* that estimate, not a replacement. Joint E0/E1/E2 experiments target waste recognition, stale belief/conflict detection, and cognitive move selection. First gate is a source/label census and executable baseline replay; then tiny-model training; then shadow model; only then separately reviewed active experiments. Require code, pinned data, reproducible metrics and negative cases before declaring completion. From 2420fe2fae1b146008a44b2c272d371dec8609e6 Mon Sep 17 00:00:00 2001 From: dp-web4 Date: Sat, 10 Oct 2026 22:15:36 -0700 Subject: [PATCH 5/6] docs: correct the +0.165 turn-detector claim (oracle headroom, not a trained result) The bridge note said "a simple detector at the right choice boundary produced +0.165". The source (dev-SAGE laya FINDINGS section 15, turn_detector.py) says +0.165 is the headroom under PERFECT turn detection; the trained detector scored 0.461 against the 0.596 repeat floor (-0.136). In SAGE, also fix the relative link to forum/claude/SAGE/SNARC_SPECIFICATION.md (one level too shallow). Seat: legion-claude Co-Authored-By: Claude Opus 5.5 (1M context) --- docs/LAYA_TBS_LEARNED_SNARC_BRIDGE_2026-10-10.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/docs/LAYA_TBS_LEARNED_SNARC_BRIDGE_2026-10-10.md b/docs/LAYA_TBS_LEARNED_SNARC_BRIDGE_2026-10-10.md index 9c9a50a..857b375 100644 --- a/docs/LAYA_TBS_LEARNED_SNARC_BRIDGE_2026-10-10.md +++ b/docs/LAYA_TBS_LEARNED_SNARC_BRIDGE_2026-10-10.md @@ -8,7 +8,7 @@ Observed, specific evidence: - Thor Phase 0 pinned Laya 421.3M, measured ~31.8ms median for a typed question, ~1.7GB resident, and reproduced typed-decisions accuracy 0.7660 on 400-case public sanity slice. But ECE was 0.2132 vs 0.1747 for the weaker base; calibration is not correctness. CPU fallback silently took ~859ms; always assert inference device. - Later cross-world probe reported ~30K-parameter numpy MLP **0.299 / 0.291** (all rows / turn points) against frozen 421M encoder **0.285 / 0.297**; pretrained encoding bought no reliable advantage on this representation. Avoid assuming Laya wins from size. - Outcome-linked lived role harvest yielded 245,404 rows over 9,408 steps; 24,180 labels; excluding circular 'avatar' evidence, leave-one-world-out accuracy 0.8490 was below majority 0.8654. Labels often captured action bars and budgets rather than lethal bodies. Evidence quality, relation/causal ontology and holdouts matter more than row counts. -- The indexed research says next-action prediction was a poor primary target: often the decisive turn cue is not present in (board_t, available_actions_t). A simple detector at the right choice boundary produced +0.165; pure state→next-action models cannot learn missing information. +- The indexed research says next-action prediction was a poor primary target: often the decisive turn cue is not present in (board_t, available_actions_t). All remaining headroom is one binary decision (is this a turn?): with a *perfect* (oracle) turn detector a two-part policy would gain +0.165 over repeat-last, but the trained turn detector scored 0.461 vs the 0.596 floor (−0.136), worse than not splitting. Pure state→next-action models cannot learn missing information. - An actionable missed-attention failure: ft09 recognized exhausted aim space around step 15 yet continued ~135 futile actions; earlier work identified repeated refusal waste and asked for an instrumented, flagged live stopping experiment. This makes *recognizing and responding to stale/unproductive focus* a high-value SNARC test case. - Laya batching on Thor was not a single shared encode: 1 question ~29–32ms; 5 ~86–88ms; 20 ~348–351ms. Benchmark shared-representation multi-head tiny models against this actual cost. From bd532cce1c98856d6a0f8dc21e1e09883788a0aa Mon Sep 17 00:00:00 2001 From: Dennis Palatov Date: Sat, 10 Oct 2026 22:44:21 -0700 Subject: [PATCH 6/6] docs: replace ephemeral research branch links with stable main links --- docs/LAYA_TBS_LEARNED_SNARC_BRIDGE_2026-10-10.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/docs/LAYA_TBS_LEARNED_SNARC_BRIDGE_2026-10-10.md b/docs/LAYA_TBS_LEARNED_SNARC_BRIDGE_2026-10-10.md index 857b375..9b916a5 100644 --- a/docs/LAYA_TBS_LEARNED_SNARC_BRIDGE_2026-10-10.md +++ b/docs/LAYA_TBS_LEARNED_SNARC_BRIDGE_2026-10-10.md @@ -1,5 +1,5 @@ # Laya/TBS → Learned SNARC: implementation bridge (2026-10-10) -**Status: concrete research handoff, not a completed training or deployment.** Read in conjunction with the [Learned SNARC PRD](https://github.com/dp-web4/snarc/blob/exploration/learned-snarc-2026-10-10/prd/LEARNED_SNARC_PRD_2026-10-10.md). +**Status: concrete research handoff, not a completed training or deployment.** Read in conjunction with the [Learned SNARC PRD](https://github.com/dp-web4/snarc/blob/main/prd/LEARNED_SNARC_PRD_2026-10-10.md). ## Why this changes the plan The Laya/System-1/TBS work already performed on Thor supplies **executed experiments**, not just related architecture. The canonical reading order is [dev-SAGE Laya INDEX](https://github.com/dp-web4/dev-SAGE/blob/main/explorations/laya-system1-tbs-2026-09-20/INDEX.md), which explicitly supersedes early findings. Do not build from REPORT alone.