Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
39 changes: 39 additions & 0 deletions docs/LAYA_TBS_LEARNED_SNARC_BRIDGE_2026-10-10.md
Original file line number Diff line number Diff line change
@@ -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/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.

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). 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.

## 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.
14 changes: 14 additions & 0 deletions docs/LEARNED_SNARC_EXPLORATION_2026-10-10.md
Original file line number Diff line number Diff line change
@@ -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.
Loading
Loading