Skip to content
voxlang.orgv0.6.0SSOT
Internals — roadmap: describes planned work that may not match the current code.

JARS, FST/FTT Token Reduction, and Local Ollama Harness Architecture (2026)

This architecture specification documents the unified Joint Adaptive Routing and Scoring (JARS), Just-About-Right (JAR) penalty calibration, Compiled Finite-State Transducer (FST) intent classification, and Fuzzy-Trace Theory (FTT) Gist-Capsule token-reduction mechanics shared across Vox, gigme, greaterFool, and platform-ssot, alongside zero-spend local Ollama evaluation protocols and deterministic grammar-constrained sampling gates.

1. Purpose & Cross-Repo Context

Section titled “1. Purpose & Cross-Repo Context”

This specification documents the unified JARS / JARVS + FST/FTT Classifier Token-Reduction & Local Ollama Architecture shared across Vox (vox-orchestrator, vox-eval, vox-mcp-registry, vox-sampler), gigme (Scout v2.0), greaterFool (FTT Dual-Trace Orchestrator v1.2), and the polyglot single source of truth in ngi.video/packages/platform-ssot.

It extends two foundational Vox architecture documents:

2. The Four-Pillar JARS + FST/FTT Architecture

Section titled “2. The Four-Pillar JARS + FST/FTT Architecture”

2.1 Pillar 1: Compiled FST + Grammar-Constrained Classifier (<0.1 ms to <40 µs/token)

Section titled “2.1 Pillar 1: Compiled FST + Grammar-Constrained Classifier (<0.1 ms to <40 µs/token)”

Across our chat and agent harnesses, replacing ad-hoc surface regexes with a two-stage deterministic + grammar-constrained classifier eliminates both regex brittleness and cloud router latency:

  1. Tier 0 — Compiled FST & Clause-Scoped Concept-Frame Scorer (<0.1 ms, $0):
    • Uses tantivy-fst in Rust (crates/vox-mcp-registry, crates/vox-orchestrator) and semanticIntentScorer.ts in TypeScript (gigme, ngi.video/packages/platform-ssot) to map morphological stems to weighted semantic frames (constraint_correction, search_recalibration, memory_audit, deep_synthesis, tool_discovery).
    • Enforces the Clause-Scoped Negation Parity Invariant (N_snip === N_win (mod 2)): negation operators (not, never, without, refuse, cannot) invert concept polarity only within their clause boundary (but, however, ;, .), preventing false-positive intent triggers on negated statements.
  2. Tier 1 — Pushdown / Earley Grammar-Constrained Local Classification (crates/vox-sampler/src/earley.rs):
    • When Tier 0 confidence falls in the borderline ambiguity band $[0.35, 0.72)$, a local model (huihui_ai/qwen3-abliterated:8b via Ollama or vox-populi local inference) is invoked with EarleySampler / JSON-schema logit masking (<40 µs/token overhead), guaranteeing valid structured output in a single pass.

2.2 Pillar 2: Fuzzy-Trace Theory (FTT) Gist-Capsule Multi-Turn Token Reduction

Section titled “2.2 Pillar 2: Fuzzy-Trace Theory (FTT) Gist-Capsule Multi-Turn Token Reduction”

Instead of running a heavy secondary neural token pruner (LLMLingua-2, which adds 50–180 ms and risks stripping negation words or identifiers), multi-turn history is compressed using Fuzzy-Trace Theory (Reyna & Brainerd, 1995) dual-trace memory (compressTurnToGistCapsule in ngi.video/packages/platform-ssot/src/harness.ts):

  • Verbatim Trace Window (Last $W = 4$ messages): Retained verbatim to preserve immediate local co-reference, code blocks, and exact numeric values.
  • Gist Capsule Ledger (Older turns $> W$): Distilled deterministically in <0.05 ms into structured single-line capsules (T{i}[frame=...,tools=...,choice=...]: anchor), preserving 100% of quantitative anchors, entity IDs, and negation polarity while cutting 8-turn conversational history tokens by 54.1–60.9%.
  • Synergy with Vox ContextLifecycleManager: Complements clear_tool_uses_20250919 (which strips stale tool-result payloads) by compressing conversational prose turns before compact_20260112 is ever needed.

2.3 Pillar 3: Just-About-Right (JAR) $[-2..+2]$ Calibration for Tool & Context Sizing

Section titled “2.3 Pillar 3: Just-About-Right (JAR) $[-2..+2]$ Calibration for Tool & Context Sizing”

Borrowing the psychometric Just-About-Right (JAR) $[-2..+2]$ bipolar scale (Rothman & Parker, 2009), every turn is scored on three sizing axes before dispatch:

  • toolCountJar $\in [-2, -1, 0, +1, +2]$:
    • -2 / -1 (too_little): Complex multi-step or cross-domain query given too few tools $\rightarrow$ widens dynamic top-$K$ selection or triggers vox_tool_search / discover_tools.
    • 0 (just_about_right): Pinned core prefix (8 tools, preserving L1 prompt cache) + 4–6 semantically matched domain tools (12–14 total, well below the 30–50 tool degradation wall).
    • +1 / +2 (too_much): Simple turn given $> 25$ tools $\rightarrow$ prunes unreferenced domain schemas, saving 45–85% of tool-schema tokens.
  • contextBudgetJar $\in [-2..+2]$: Calibrates the verbatim window $W$ and RAG chunk count against query complexity.
  • verbosityJar $\in [-2..+2]$: Calibrates target response length so simple factual lookups do not emit 600-token essays.

2.4 Pillar 4: APA JARS-Quant/Mixed & Local Ollama Trial Flywheel (crates/vox-eval/src/trial_flywheel.rs)

Section titled “2.4 Pillar 4: APA JARS-Quant/Mixed & Local Ollama Trial Flywheel (crates/vox-eval/src/trial_flywheel.rs)”
  1. Local Ollama Integration (http://127.0.0.1:11434):
    • Uses huihui_ai/qwen3-abliterated:8b for fast $0 routing/tool-selection trials (pass@k and $\text{pass}^k = (c/n)^k$) and huihui_ai/Qwen3.8-abliterated:27b as the local LLM-as-Judge for rubric scoring in vox-eval.
  2. JARS-Quant/Mixed Provenance Ledger:
    • Every benchmark artifact records git SHA, corpus fingerprint, model tier, $\text{pass}^k$ reliability across $k=3$ trials, JAR penalty drop, token-reduction percentage, and a ranked weakness table generated by surfaceHarnessWeaknesses.

3. Concrete Vox Integration Map

Section titled “3. Concrete Vox Integration Map”
Vox Crate / ModuleCurrent MechanismJARS + FST/FTT + Ollama Enhancement
crates/vox-mcp-registry & vox-capability-registry330 MCP tools (~42k tokens if dumped flat); select_tools_for_turnEnforce prefix-stable pinned_core tool ordering (tools -> system -> messages L1 cache invariant) + JAR $[-2..+2]$ dynamic top-$K$ sizing + vox_tool_search active discovery
crates/vox-orchestrator/src/models/mod.rsHeuristic model cascadeAdd Tier 0 Compiled FST + Clause Negation Parity scorer (<0.1 ms) and Tier 1 Local Ollama (qwen3-abliterated:8b via EarleySampler JSON schema mask) before cloud escalation
crates/vox-orchestrator/src/context_lifecycle.rsclear_tool_uses_20250919 + compact_20260112Add deterministic FTT Gist-Capsule compression (compress_turn_to_gist_capsule) for turns older than $W = 4$ before triggering paid LLM compaction
crates/vox-eval/src/trial_flywheel.rspass_at_k, pass_pow_k, EvalTrialTaskAdd JarScaleEvaluation ($[-2..+2]$ penalty analysis), local Ollama trial runner (http://127.0.0.1:11434), and automated surface_harness_weaknesses failure-mode miner