RAM Local Intelligence Spike v2

Phase A · Harness 2 R3 · sterile-prompt semantic retest in the 1.0B–1.3B Safari feasibility window. This is not RAM, not Build 13 R10, and cannot commit memories.

Starting harness…
Isolation rule: deploy this folder to the separate experimental HTTPS origin. It contains no RAM database, MemoryService, Backup, lifecycle, provenance, or canonical persistence code.
H2R3 sterile-prompt repair: the semantic prompt contains no concrete Capture/date/time/person examples. A nonsemantic protocol sentinel is rejected if it leaks into model output. H2R2 timeout, sanitizer, and worker-reset protections remain intact.

1 · Device capability

Harness identity

Build
WebLLM
Context tokens
Engine

WebLLM remains pinned at 0.2.82. Qwen2.5 1.5B is intentionally omitted from this harness after a Safari page reload during smoke inference.

2 · Pick and load a model

Recommended order: Llama 3.2 1B control → Phi-1.5 ~1.3B → TinyLlama 1.1B.

No model loaded

3 · Basic inference checks

Run these once after each model reaches READY. Plain smoke proves generation. Constrained JSON sanity isolates grammar/JSON machinery from RAM semantics.

4 · Atomic interpretation

The validator reports structure, prompt leakage, and semantic benchmark quality separately. H2R3 removes concrete examples from the semantic prompt before retesting model quality.

Raw model output

No inference yet.

External validation + benchmark

No inference yet.

5 · Evidence

Runtime log