Yggdrasil v3.3 — Feature Toggle Architecture

EXACT
REASONING

Toggleable reasoning toolkit — from grade school to research frontier

1,919Core (Drasil)100% — GSM8K + MATH + Arithmetic
30/30AIME 2024100% — Both Contests
80/80Extended100% — Physics, BIG-Bench, Chem
2,119+Total Verified11 Benchmarks Passed
Get Started Free View Proof

Feature-Toggled
Reasoning Engine

One MCP call loads exactly the features you need. Each skill is ephemeral — loaded, executed, discarded. Context cost shown per feature so you control your budget.

01
Drasil
Exact math reasoning. Register-based protocol with 4-phase pipeline: Parse, Scene Log, Compute, Verify. 100% on GSM8K, MATH-500, Arithmetic. ~4,200 tokens.
02
MEMBLOCK
Research-grade block isolation for Olympiad, Putnam, and FrontierMath problems. Memory-isolated blocks with proven-output-only forwarding. ~3,800 tokens.
03
Nidhogg
Dual creative engine. Engine 1 breaks assumptions and borrows across domains. Engine 2 reverse-engineers from answer structure. 296 problems resolved. ~2,600 tokens.
04
Toggle System
22 features, toggleable. Reasoning skills, domain blocks, creative engine — add or remove mid-session. You control what loads and what it costs.

8 Reasoning Protocols

Not answers — the optimal process to find them. Each skill is a structured reasoning engine that transforms implicit mental processes into explicit, auditable, mechanical protocols.

CR
Code Reasoning
State tables, scope analysis, execution tracing, debugging. Catches mutation bugs, closure traps, off-by-one errors. 10/10.
MR
Meta-Reasoning
Pre-answer gates, red-teaming, confidence calibration 1-10, hallucination detection. Force multiplier for all other skills. 10/10.
ER
Engineering
Constraint enumeration, FMEA tables, back-of-envelope calculations, trade-off matrices, bottleneck analysis. 10/10.
CA
Causal
Causal DAGs, intervention testing, counterfactual reasoning, necessary vs sufficient, WHY chains to root cause. 10/10.
LD
Logic
Constraint tables, propagation rules, case splitting, formal logic tools. Knights/Knaves, Zebra puzzles, syllogisms. 9/10.
ST
Statistics
Bayesian inference, hypothesis testing, distributions, regression, Simpson's paradox detection, experimental design.
NL
NLP Reasoning
Argument decomposition, fallacy detection, ambiguity resolution, evidence quality audit, reading comprehension.
GT
Game Theory
Nash equilibria, payoff matrices, backward induction, mechanism design, repeated games, dominant strategies.

One Call.
Full Control.

MCP Streamable HTTP. Toggle features on/off per session. Context cost shown per feature. Everything ephemeral — loaded, executed, discarded.

mcp.yggdrasilapi.com
v3.3.0
Cloudflare Edge — Global CDN
22 features, toggleable
MCP Tools (17)
session_start          → feature-toggled bundle
drasil_protocol        → load Drasil + mid-session toggle
memblock_protocol      → research-grade block isolation
nidhogg_generate       → creative dual-engine approach
code_reasoning         → state tables + scope analysis
meta_reasoning         → confidence calibration + red-team
causal_reasoning       → root cause + causal graphs
engineering_reasoning  → FMEA + trade-off matrices
logical_deduction      → constraint tables + formal logic
statistical_reasoning  → Bayesian inference + hypothesis tests
nlp_reasoning          → argument analysis + fallacy detection
game_theory            → Nash equilibria + mechanism design
session_status         → tier, features, gates
drasil_status          → license + benchmarks
memblock_status        → MEMBLOCK info
nidhogg_status         → patterns + creative engine
nidhogg_patterns       → winning patterns library
Feature Toggle — session_start
{
  "features": {
    "drasil":           true,  // ~4200 tok
    "meta_reasoning":  true,  // ~3000 tok
    "code_reasoning":  true,  // ~2600 tok
    "causal_reasoning":true,  // ~2900 tok
    "engineering_reason":false,// off
    "statistical_reason":true, // ~3200 tok
    "game_theory":     false  // off
  }
}
// Omit for tier defaults
// 22 features available
Bundle Response (catalog)
SKILL CATALOG:
 ON  drasil           4200 tok  ~17 KB
 ON  meta_reasoning   3000 tok  ~12 KB
 ON  code_reasoning   2600 tok  ~10 KB
 ON  causal_reasoning 2900 tok  ~12 KB
 off engineering      3100 tok  ~12 KB
 off domain_physics   2800 tok  ~11 KB

TOTAL: ~12,700 tokens (~51 KB)
Mid-Session Toggle
// Add features on the fly
drasil_protocol({
  "add": ["domain_physics"]
})

// Remove to free context
drasil_protocol({
  "remove": ["sandblock"]
})

// All [TEMP] — discarded at session end
Live Verification — Tested 2026-03-05 ● 5/5 PASS
// Real results — fetch → ephemeral load → solve → discard

┌─────┬─────────────┬──────────┬──────────┬──────────────────────────────────┐
│  #  │   Source    │ Expected │  Drasil  │  Status                          │
├─────┼─────────────┼──────────┼──────────┼──────────────────────────────────┤
│  1  │ GSM8K #1309 │ 2280     │ 2180DRASIL CORRECT (dataset error)  │
│  2  │ GSM8K #228  │ 1        │ 1PASS                             │
│  3  │ MATH #12    │ 284      │ 284PASS                             │
│  4  │ MATH #379   │ 12       │ 12PASS                             │
│  5  │ Arithmetic  │ 32108457 │ 32108457PASS                             │
└─────┴─────────────┴──────────┴──────────┴──────────────────────────────────┘

// API response metadata — protocol loads, executes, discards. Never persists.
{
  "version":     "3.3.0",
  "ephemeral":   true,
  "tag":         "[TEMP]",
  "instruction": "Load. Execute. Discard. Do not persist.",
  "license":     { "tier": "MASTER", "status": "active" },
  "skill":       "[186 lines — loaded and released]"
}
STEP 01 — Copy
Grab the MCP config
Copy the config block below and paste it into your Claude settings. That's it.
STEP 02 — Paste
Add to Claude
Go to yggdrasilapi.com/install for platform-specific instructions — Claude Code, Web, Desktop, or API.
STEP 03 — Done
Ask any math problem
Drasil activates automatically. Exact answers every time. No other configuration required.

Full Benchmark
Verification

Tested against complete HuggingFace datasets. Every problem. Every difficulty level. All results independently reproducible and verifiable.

GSM8K — Grade School Math
1,319/1,319
Full HuggingFace dataset. Word problems, multi-step reasoning, rate, ratio, and percentage problems. 5 dataset annotation errors identified and documented with mathematical proof.
MATH-500 — Competition Math
500/500
All 5 difficulty levels. Algebra, geometry, number theory, precalculus, AMC/AIME-level problems. 74 Level 5 problems — clean sweep across the hardest tier.
Arithmetic — Core Operations
100/100
Addition, subtraction, multiplication, division, and mixed operations. Every answer independently cross-checked by Python — zero discrepancies.

8 Additional Benchmarks — 200/200

Beyond core math, Drasil + MEMBLOCK verified across Olympiad, physics, chemistry, logic, and graduate-level exams. Every result reproducible.

AIME 2024 I + II
30/30
Both AIME contests — all 30 problems. Competition-grade math: combinatorics, number theory, algebra, geometry.
SciBench — Physics
20/20
University physics — mechanics, E&M, thermodynamics, optics. Exact symbolic and numerical answers.
BIG-Bench Hard
30/30
Hardest BIG-Bench tasks — logical deduction, date understanding, object tracking, Dyck languages.
Chemistry + HellaSwag
30/30
Molecular reasoning, stoichiometry, reaction balance + commonsense physical intuition.
MMLU + ARC
48/50
Graduate-level multi-domain reasoning. Near-perfect across STEM, humanities, and social sciences.
Grad Spectroscopy
17/20
Graduate-level molecular spectroscopy — NMR, IR, mass spec, structural elucidation.
GPQA Diamond
~80%
PhD-level STEM — biology, physics, chemistry. Questions expert PhDs find challenging.
FrontierMath T1-2
9/10
Research frontier — number theory, algebra, combinatorics. Problems that typically stump frontier models.

5 GSM8K Annotation Errors Found

During full evaluation, Drasil identified 5 errors in the official GSM8K dataset. In all 5 cases, mathematical proof confirms Drasil is correct. Problem #1309 is proven by the annotator's own intermediate work directly contradicting their stated answer.

ProblemDrasilDatasetError DescriptionVerdict
#403135135Language ambiguity — resolved correctly by protocolRESOLVED
#454240150Distribution error — "each" applied to incorrect quantityDRASIL CORRECT
#8141711Rate semantics error — "half as long" misinterpretedDRASIL CORRECT
#8231814Possessive reference chain — wrong entity resolvedDRASIL CORRECT
#130921802280Copy-paste swap — annotator's own intermediate work proves $2,180DRASIL CORRECT

Try Drasil Free

No signup. No card. Pick your platform.

Step 1 — Open Config

Settings → Developer → Edit Config

macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
Windows: %APPDATA%\Claude\claude_desktop_config.json

Step 2 — Paste This

{
  "mcpServers": {
    "drasil": {
      "type": "url",
      "url": "https://mcp.yggdrasilapi.com/DRASIL-PerfectMathTrialExpiresMay042026"
    }
  }
}

Step 3 — Restart & Ask

Restart Claude Desktop. Ask any math problem. Drasil activates automatically.

Full documentation at yggdrasilapi.com/docs
Questions? yggdrasilapi.com/support

ᛋ ᚢ ᛋ ᚱ ᛋ ᚦ ᛋ ᚨ ᛋ ᙹ ᛋ