person
INDEX SPREAD+4.18 bpstrending_up
EVALS QUEUED84 ACTIVE
MMLU-PRO MEDIAN74.2%▲ 0.8%
UPTIME: 99.994% DETERMINISTIC
SPEC REVISION
v2.4.2-STABLE
LEDGER SIGNATURE
verified e8f9...3d01 [SIGNED]
ACTIVE HARNESS NODES 64x Bare-Metal Ubuntu 24.04
DETERMINISM INTEGRITY
99.994% ZERO-DRIFT
shield_with_heart Scientific Evaluation Charter

The Airecmark Evaluation Protocol: Quantitative Methodology & Scientific Charter

Standardized specifications governing empirical scoring for generative code systems, autonomous developer agents, and multimodal reasoning engines. We enforce zero commercial bias, isolated Tree-sitter AST syntax validation, microsecond network packet telemetry, and air-gapped test runners.

functions Inspect Scoring Matrix
Governance Standards

The 4 Ironclad Laws of Airecmark Intelligence

STATUTE REFERENCE: AIC-GOV-2026
LAW 01 find_replace

Zero-Sponsored Placement

Vendor spend, sponsorship funds, cloud partnerships, or corporate affiliation can never adjust an Airecmark Score, recalibrate test weights, or modify ranking index position.

ENFORCEMENT: Dual Auditor Sign-off
LAW 02 data_check

Deterministic Execution

Zero subjective qualitative "LLM-as-a-judge" grading. Every synthetic diff and model patch is built, compiled, and unit-tested in isolated ephemeral microVMs against native test suites.

ENV: Firecracker MicroVMs
LAW 03 account_tree

Production Monorepo Scale

Zero synthetic single-file coding exercises. Tests require traversing inter-package dependency graphs in codebases exceeding 100k lines of active TypeScript, Rust, and C.

CORPUS: 182 Package Graph
LAW 04 lock

Cryptographic Attestation

Every evaluation execution log, token logit variance, and compiler stdout dump is hashed and committed to an open, append-only Merkle ledger for third-party recreation.

LEDGER: Merkle Tree v2 Block
Weight Matrix Formula

Composite Score Mathematical Engine

The overall Airecmark Composite Score (0.0 to 10.0) is a weighted vector computed from four deterministic test clusters.

EQUATION 2.4
Score = 0.40(SAST) + 0.25(SContext) + 0.20(SLatency) + 0.15(SErgo)
NORM: MIN-MAX [0.00, 10.00]
WEIGHT: 40% PRIMARY VECTOR

AST Correctness & Dry-Run Execution (SAST)

0.40

Evaluates syntax tree validity and operational execution using Tree-sitter grammar parsers. Zero tolerance for imaginary package imports, incorrect argument arity, or broken interfaces.

TypeScript 5.8 Strict Check tsc --noEmit --strict
Rust 1.85 Cargo Clippy cargo test --all-targets
Python 3.12 Type Strictness mypy --strict --disallow-any
WEIGHT: 25% RETRIEVAL VECTOR

Context Needle Recall @ Scale (SContext)

0.25

Measures precision of multi-file graph resolution across 128k to 200k active context tokens. We inject buried interface definitions across distant sub-modules to verify dependency propagation.

Token Depth Window 128,000 — 200,000 tokens
Needle Dispersion Spread Randomized Quintiles (Q1-Q5)
Cross-File Symbol Trace > 12 File Hop Depth
WEIGHT: 20% PERFORMANCE VECTOR

TTFT & Streaming Velocity (SLatency)

0.20

Real-time TCP socket telemetry measuring packet arrival delays, Time-To-First-Token (TTFT), and sustained token stream throughput (tok/s) sampled concurrently from 4 global test points.

Edge Monitoring Points IAD, FRA, NRT, SGP
TTFT Target P95 Threshold < 280ms Deterministic
Burst Stability Test 20 Concurrent In-Flight Calls
WEIGHT: 15% USABILITY VECTOR

Ergonomics & Atomic Diff Reversibility (SErgo)

0.15

Calculates the cognitive friction and engineering burden of AI code suggestions. Models lose points for generating multi-hundred-line formatting diffs that clobber unrelated git blame records.

Git Blame Preservation > 94% Line Conservation
Ghost Text Flicker Index < 14ms Render Latency
Rejection Overhead 1-Key Clean Escape

Interactive Composite Score Simulator

Adjust benchmark sub-scores to inspect real-time mathematical aggregation.

COMPUTED AIRECMARK SCORE: 8.78 GRADE A+
AST Correctness (40%) 9.2
Context Recall (25%) 8.6
Latency / TTFT (20%) 8.0
Ergonomics (15%) 8.9
Harness Engineering

Isolated Sandboxing & Hardware Cluster Pipeline

High-throughput test runs execute in zero-network Firecracker microVMs launched on dedicated bare-metal iron.

01
Edge Probe Dispatch

4x Equinix Metal c3.medium instances dispatch parallel WebSocket and REST streams to capture raw network latency.

TCP RTT: ~4.2ms
02
MicroVM Isolation

Generated patches are applied inside ephemeral Firecracker MicroVMs with strictly blocked outbound egress.

Egress: 0 bytes permitted
03
Tree-sitter AST Parse

Parsers traverse concrete syntax tree deltas, isolating changed semantic nodes and verifying contract boundaries.

Grammars: TS, Rust, Python, Go
04
Cryptographic Hash

Raw output logs, AST diff reports, and runtime execution codes are concatenated, signed, and broadcast to the public index.

Algorithm: Ed25519
folder_special Official Monorepo Evaluation Corpora (Active Set v2.4)
Total Test Cases: 4,892 Scenarios
Corpus Identifier Target Ecosystem Total Scale Dependency Topology Pass Criterion
CORPUS-TS-POLYGLOT-182 TypeScript 5.8 / Node 22 / Turborepo 182 Pkgs / 284k LoC Multi-tier circular interfaces
CORPUS-RUST-DISTRIB-DB Rust 1.85 / Tokio / Raft consensus 34 Crates / 112k LoC Async lifetime borrow checking
CORPUS-KERNEL-C-SUBSET C11 / Linux VFS / eBPF filters 16 Subsystems / 410k LoC Pointer safety / Memory boundaries
CORPUS-PY-DATA-DISTRIB Python 3.12 / Polars / PyO3 FFI 12 Modules / 95k LoC Cross-language FFI type signatures
Immunity Controls

Scoring Normalization & Anti-Gaming Defense

Foundation models frequently memorize public benchmark datasets. We utilize runtime mutation and prompt perturbation to make benchmark memorization mathematically impossible.

shuffle Dynamic AST Mutation

Prior to injecting an evaluation task into a model context, our compiler synthetically transforms all private internal variable names, interface labels, and module paths with randomized semantic synonyms. Models that memorized canonical GitHub repository solutions immediately trigger syntax compilation failures.

MUTATION ENGINE: Synonym AST Permutation
tune 15-Fold Prompt Perturbations

Every discrete task is evaluated across 15 semantically identical but syntactically divergent instruction variants. If a model passes prompt version A but fails version B and C, a logit stability penalty is automatically deducted from the final score to isolate prompt brittleness.

VARIANCE PENALTY: σ² > 0.12 Triggers Decay
history_toggle_off Silent Canary Probing

Commercial API endpoints are pinged continuously every hour using rotating private synthetic canaries. When an AI vendor silently updates weights or applies quantization degradation behind a static endpoint name, our telemetry flags the regression within 60 minutes.

CADENCE: 3,600s Automated Trigger
Methodological Transparency

Frequently Addressed Protocol Inquiries

Independent Audit Kit

Verify Test Executions Independently

Download the standalone Docker Compose evaluation harness, sanitized AST test definitions, and signed raw JSON telemetry captures to verify any published score on your own hardware.

verified_user PUBLIC MERKLE ROOT: 0x9f4a...83d2c10b
PROTOCOL LICENSE: Creative Commons BY-ND 4.0 / MIT Test Harness