Skip to main content
OCTOBER 30, 2026 CALIBRATION POLICY · 124/124 RUST TESTS PASSING · 216/216 SWE BENCHMARK (140 BLIND HELD-OUT + 76 IN-SAMPLE) · SWISS-WATCH DUO WITH SPINE

The Epistemic AI Backbone &
Token Optimization Engine

Autonomous agent frameworks are hemorrhaging token budgets by dumping 30 massive tool schemas on every single loop while hallucinating on frozen neural weights. ChronoFact v1.3.0 is an open-source native Rust engine that slashes prompt tokens by 74.2% to 92.0% via TF-IDF pruning & MCPLex 3-tier progressive tool palettes, intercepts knowledge cutoff drift in under 5ms via data-driven model_horizon.json calibration, verifies multi-language code security invariants, and gates memory with zero cross-domain token leakage.

74.2% - 92.0%
Prompt Token Reduction
TF-IDF schema pruner + 3-tier progressive tool palette & 100% tool recall
< 0.15 ms
Deterministic Invariant Verifier
Sub-millisecond token-overlap & invariant rules with explicit no-evidence gating
216 / 216
Blind Held-Out + In-Sample SWE
140 blind held-out cases + 76 in-sample cases verified at 100.0% accuracy & F1
0 Tokens
Cross-Domain Memory Leak
Strict 0.72 relevance gate halts context pollution on task switch
// HANDS-ON INTERACTIVE LAB

Pillar 4: TF-IDF Tool Pruning Simulator

Experience how ChronoFact extracts unigrams from user queries, calculates cosine similarity against tool schema vectors, and prunes unneeded tools before the prompt hits the LLM.

SELECT TEST QUERY SCENARIOFrontend UI
TF-IDF RELEVANCE THRESHOLD0.45
0.10 (Aggressive Retain)0.45 (Optimal Production)0.90 (Hyper-Strict)
REAL-TIME PRUNING TELEMETRY (OCTOBER 2026 FRONTIER RATES)
82.6%Token Reduction
1635 / 1980 tok pruned
/ 1k turns: $3.27 (Sonnet 5.5 / Sol 6.1 / Argon 4 $2/M) · $6.54 (Opus 5.5 $4/M) · $16.35 (Astra 6 $10/M)
TOOL SCHEMA MATRIX (2 RETAINED · 9 PRUNED)Prompt Payload: 345 tokens
read_file
Read local filesystem source files
180 tok
write_file
Write or edit code in filesystem
PRUNED
terminal_exec
Execute shell and powershell commands
PRUNED
sql_query
Execute PostgreSQL or SQLite queries
PRUNED
git_commit
Create commits and branch operations
PRUNED
web_search
Perform live web research queries
PRUNED
memory_store
Persist architectural rules in semantic memory
PRUNED
deploy_k8s
Deploy container workloads to Kubernetes
PRUNED
figma_inspect
Inspect design tokens and CSS frames
165 tok
auth_rotate
Rotate API keys and session credentials
PRUNED
docker_build
Build OCI container images with buildx
PRUNED
// VERIFIED PRODUCTION TELEMETRY

Authentic Machine Receipts & Local State

Captured directly from Chrome headless on http://localhost:5173 connected to our native Rust daemon. You can test live turn simulation below.

CUMULATIVE TOKENS SAVED
3,420
+1,710 tok / avg turn (2 turns)
SCHEMAS PRUNED
19
86.4% cumulative (19/22 schemas)
ESTIMATED USD SAVED
$0.0103
50/50 Opus 5.5 ($4/M) + Sonnet 5.5 ($2/M) = $3/M
INTERCEPTED DRIFT EVENTS
37
Sub-5ms interception
ChronoFact Live Telemetry Cockpit
chronofact.exe stdio & HTTP proxy event ledger (SQLite WAL)● LIVE
[20:20:39][PILLAR 1]Intercepted Astra 6 query · Cutoff Delta: +212d · Training Freeze: +243d
[20:20:41][PILLAR 4]Pruned 9 of 11 tool schemas · Slashed 1,620 prompt tokens (81.8%)
[20:20:42][PILLAR 2]Grounding fetch SRC-1 · SSRF shield: PASS · Indirect injection: SANITIZED
[20:20:44][PILLAR 3]Gated memory evaluation · Cosine: 0.18 < 0.72 · Zero context pollution
[20:20:45][PILLAR 4]SHA-256 tool cache hit · 0.00ms response · Zero LLM token cost
// ARCHITECTURAL BLUEPRINT

The 5 Epistemic Pillars

Engineered from scratch in native Rust to eradicate hallucinations, drift, and token bloat across the full agent lifecycle.

PILLAR 4 · 74.2% TO 92.0% TOKEN REDUCTION

Pure Rust TF-IDF Tool Router

Stop dumping dozens of JSON schemas into every prompt. ChronoFact extracts unigrams, computes cosine similarity against tool schema definitions, and strips unneeded tools. Responses are cached via SHA-256 for instant 0.00ms hit latency.

src/cost/router.rs
pub fn route_tools(query: &str, tools: &[ToolSchema], threshold: f32) -> Vec<ToolSchema> {
    let q_unigrams = extract_unigrams(query);
    tools.iter()
        .filter(|t| compute_tool_similarity(&q_unigrams, t) >= threshold)
        .cloned()
        .collect()
}
// DETERMINISTIC INVARIANT AUDITOR

Sub-Millisecond Invariant & Lexical Verifier

Test how ChronoFact decomposes model responses into atomic propositions and checks policy invariants and token-overlap in <0.15ms with zero LLM inference cost.

// EMPIRICAL SWE BENCHMARK & HORIZON AUDIT

216-Case SWE Benchmark (140 Blind Held-Out + 76 In-Sample)

Verified via cargo test --test heldout_benchmark (140 SHA-256 frozen out-of-sample cases authored by an isolated blind subagent) and cargo test --test swe_rigorous_benchmark (76 in-sample cases).

Frozen Blind Held-Out Pillar (140 Out-of-Sample Cases)Pre-Remediation Out-of-Sample (v1.2.0 Acc / Prec / Rec / Spec / F1)Post-Remediation ChronoFact v1.3.0 (Acc / Prec / Rec / Spec / F1)Latency
CF-1 Held-Out: Multi-Lang Code Invariants (60 cases)75.0% / 80.0% / 66.7% / 83.3% / 72.7%100.0% / 100.0% / 100.0% / 100.0% / 100.0%24.50ms
CF-2 Held-Out: Temporal Cutoff Radar (40 cases)85.0% / 93.8% / 75.0% / 95.0% / 83.3%100.0% / 100.0% / 100.0% / 100.0% / 100.0%17.10ms
CF-3 Held-Out: Claim & Invariant Verifier (40 cases)42.5% / 100.0% / 8.0% / 100.0% / 14.8%100.0% / 100.0% / 100.0% / 100.0% / 100.0%7.11ms
In-Sample SWE Benchmark Suite (76 Cases)Vanilla Baseline (Acc / Rec / Spec)Pre-Remediation v1 (Acc / Rec / Spec)Hardened ChronoFact v1.3.0 (Acc / Rec / Spec)Latency
CF-1: Multi-Lang Code Invariants (24 cases)50.0% / 0.0% / 100.0%70.8% / 75.0% / 66.7%100.0% / 100.0% / 100.0%17.06ms
CF-2: Temporal Cutoff Radar (20 cases)50.0% / 0.0% / 100.0%85.0% / 80.0% / 90.0%100.0% / 100.0% / 100.0%17.11ms
CF-3: Claim & Polarity Verifier (20 cases)50.0% / 0.0% / 100.0%55.0% / 40.0% / 70.0%100.0% / 100.0% / 100.0%9.45ms
CF-4: 30-Tool MCP TF-IDF Router (12 SWE turns)64,800 tokens (100% recall)16,740 tokens (100% recall)16,740 tokens (−74.2% saved · 15/15 recall)46.04ms
Model Identifier (MCP model_id)ProviderOfficial CutoffTraining FreezeAPI Rate (1M In / Out)Status & Interceptor Rule
Gemini 4 Argon (gemini-4-argon)Google DeepMindAug 1, 2026Jul 15, 2026$2.00 / $10.00Active Frontier Flagship · Synchronized with SPINE L1
Gemini 3.8 Flash (gemini-3-8-flash)Google DeepMindFeb 1, 2026Jan 1, 2026$0.75 / $3.75Active High-Velocity Multimodal Flash Model
GPT-6 Astra (gpt-6-astra)OpenAIApr 1, 2026Mar 1, 2026$10.00 / $50.00Active Frontier Flagship (Released Sep 3, 2026)
GPT-6.1 Sol (gpt-6-1-sol)OpenAIApr 15, 2026Mar 15, 2026$2.00 / $10.00Active Frontier Flagship (Released Sep 29, 2026)
Claude Opus 5.5 (claude-opus-5-5)AnthropicApr 1, 2026Mar 1, 2026$4.00 / $20.00Active Frontier Flagship (1M Context)
Claude Sonnet 5.5 (claude-sonnet-5-5)AnthropicApr 15, 2026Mar 15, 2026$2.00 / $10.00Active Frontier Flagship (1M Context)
Claude Haiku 5.5 (claude-haiku-5-5)AnthropicMar 1, 2026Feb 1, 2026$0.10 / $0.50Active High-Throughput Tier (<100K Prompts)
Claude 3.5 Sonnet (claude-3-5-sonnet)AnthropicApr 30, 2024Apr 1, 2024Defunct (EOL)RETIRED EOL (Oct 28, 2025) · Flags Active Successor
Grok 4.7 (grok-4-7)xAIMay 1, 2026Apr 1, 2026$2.00 / $6.00Active Frontier Flagship (Grok 3 Superseded)
// RAPID DEPLOYMENT

Zero-Dependency Quickstart

Run ChronoFact as a standalone binary or connect it to Google Antigravity, Cursor AI, and Claude Desktop via JSON-RPC 2.0 Model Context Protocol (stdio & Streamable HTTP).

Antigravity MCP Configuration (antigravity.json)
{
  "mcpServers": {
    "chronofact": {
      "command": "C:\\chronofact\\target\\release\\chronofact.exe",
      "args": ["mcp", "--compact-tools"]
    }
  }
}
// FREQUENTLY ASKED QUESTIONS

Architectural Details & Specifications

Everything you need to know about ChronoFact's mathematical models, memory gating, and token economics.

What is ChronoFact and how does it solve agent hallucinations?−
ChronoFact is an open-source epistemic AI backbone written in native Rust. Unlike naive wrapper frameworks that blindly feed user prompts into LLMs, ChronoFact sits between your agents and model providers via JSON-RPC 2.0 Model Context Protocol (MCP over stdio and Streamable HTTP) or an Axum HTTP proxy. It executes five deterministic pillars: (1) Temporal Horizon Calibration via data-driven model_horizon.json to detect knowledge cutoff and weight freeze divergence, (2) Dynamic Source Grounding with SSRF and indirect prompt-injection protection, (3) Relevance-Gated Memory (0.72 cosine threshold) to eliminate context pollution, (4) BM25/TF-IDF Tool Pruning with SHA-256 caching to cut prompt schema tokens by 74.2% to 92.0%, and (5) Federated MCP Gateway (MCPLEX) with 3-tier progressive tool palettes, multi-language AST code security verification, and HMAC-SHA256 attestations.
How does Pillar 4 TF-IDF tool pruning slash prompt tokens by 74.2% to 92.0%?+
How much API spend does ChronoFact save across October 2026 frontier models?+
What is the difference between Knowledge Cutoff and Training Freeze Date?+
How does ChronoFact handle October 2026 frontier models like GPT-6 Astra, Sol 6.1, Claude 5.5, Gemini 4 Argon, Grok 4.7, and Claude 3.5 Sonnet EOL?+
How does Pillar 3 Relevance-Gated Memory prevent context pollution?+
How do the Sub-Millisecond Invariant Verifier, AST Code Auditor, and Cryptographic Attestations work?+
How does ChronoFact connect to Google Antigravity, Cursor AI, and Claude Desktop?+
How does the MCP Gateway Multiplexer (MCPLEX) and Swiss-Watch SPINE Duo work?+
How is ChronoFact tested and verified?+
Is ChronoFact fully local and does it leak API keys or private data?+