Autonomous Skills & Workflows Hub

Active Custom Skills

4 Dedicated

Audit, Runner, Flywheel, Health Guard

Consulting Consortium

7 Firms

MBB (3) + Tactical Systems & Risk (4)

Execution Channels

Dual Surface

Natural Chat UI + 1-Word CLI Wrappers

Local Compute Cost

$0.00

Zero-token Ollama autopsies & invariant checks

🧠 Global Skill Architecture & Cognitive Routing

How incoming developer prompts route through specialized Antigravity skills, cognitive tiers, and runtime execution daemons.

🏛️ strategic-tactical-audit Location: .agents/skills/

Adversarial Consulting Consortium

Challenges capital allocation, detects proxy traps, eliminates academic toil, and enforces invariant checks using a 7-firm consortium: McKinsey, BCG, Bain, Deloitte, Accenture, Ernst & Young (EY), and PwC Forensics.

Trigger: "audit <track>" or ./audit

⚡ run-project Location: .agents/skills/

Autonomous Project & Solver Driver

Forces immediate execution of headless solver loops, unblocks hibernating/paused tracks, clears stale PID locks, and dispatches ticks with polite OS priority (os.nice(5)).

Trigger: "run <track>" or ./run

🔄 competition-flywheel Location: .agents/skills/

Universal Autonomous Evolutionary Flywheel

Multi-tier validation screening (Tiers 0–3), zero-cost loss autopsies using local DeepSeek-R1, targeted code mutations, and quota-governed daily submissions with SHA-256 deduplication.

Engine: tools.flywheel

🛡️ model-health-guard Location: .agents/skills/

Zero-Null Interceptor & VRAM Manager

Detects failed inference (eval_count == 0), intercepts empty responses, flushes VRAM, auto-offloads idle models (keep_alive: 0), and cascades down fallback tiers.

Daemon: PID 2491 Active

🏛️ Strategic & Tactical Audit Consortium

Simulated executive consulting review challenging competition methodology, expected value ($EV$), and systems integrity.

7 Independent Personas

STRATEGIC TIER (MBB) Capital, Growth & Turnaround

1. McKinsey Strategy Practice Capital & $EV$

Calculates Expected Value: $EV = P(win) \times Purse - Burn$. Enforces pre-registered kill criteria when algorithmic returns flatline.

2. Boston Consulting Group (BCG) Experience Curve

Audits flywheel velocity (Δscore / Δsub) and leaderboard crowd dynamics. Exploits orthogonal architectural vectors where 90% of competitors have blindspots.

3. Bain & Company PE Turnaround

Ruthless zero-toil audit. Freezes academic paper drafting (e.g. 14.7k words on dead ends) and institutes 14-day rapid restructuring sprints focused 100% on what pays.

TACTICAL TIER (BIG 4 + TECH) Architecture, Rigor & Risk

4. Deloitte Technical Advisory Systems Architecture

Forensics on hidden runtime invariants (e.g. gateway 1-run locks, rolling 168h submission locks) and multi-account GPU failover routing (kaggle_pool.py).

5. Accenture / Thoughtworks Software Rigor

Bit-exact determinism, eliminating non-deterministic hash() usage, enforcing single-variable ablation isolation, and building CI/CD rollback harnesses.

6. Ernst & Young (EY) Model Assurance

Noise floor modeling (MDE), extreme boundary perturbation stress-testing, and mandatory FAIR / open-source reproducibility auditing for prize payouts.

7. PwC Forensics Data Integrity

Train/test leakage detection, public string ID trap audits (e.g. if "re86" in gid failing on hashed test sets), and offline wheel licensing compliance.

Consortium Activation Triggers

Trigger Mode	Condition	Mandate & Outcome	Executable Command
Stagnation Plateau	Score delta < MDE for >48h	Falsifies current baseline; forces algorithmic pivot	python3 tools/adversarial_audit.py --mode stagnation
Pre-Milestone Freeze	≤ 14 days to prize freeze	Freezes paper drafts; commits 100% compute to high-ROI solver	python3 tools/adversarial_audit.py --mode pre-freeze
Resource / Quota Alert	GPU hours < 20% or slot locked	Activates secondary account pool (kaggle_pool.py) failover	python3 tools/adversarial_audit.py --mode emergency
On-Demand Direct	Chat UI prompt or CLI shortcut	Full 360-degree forensic report & JSON matrix deliverable	./audit <track-id>

⚡ Autonomous Run Project Engine (`project_driver.py`)

Dispatches background headless solver ticks, validates compute headroom (Rule 2), and manages process lifecycles.

os.nice(5) Priority

Execution Safety & Concurrency Lifecycle

1. Headroom Audit

Samples os.getloadavg(). Requires ≥4 cores reserve for WindowServer unless forced.

2. Lock / PID Guard

Inspects driver_locks/*.lock. Prevents duplicate runs; --force terminates stale PIDs.

3. Venv Resolution

Detects project-isolated .venv/bin/python. Isolates dependencies without host pollution.

4. Polite Priority

Spawns subprocess with os.nice(5). Logs output into project logs with fallback safety.

Registered Track Execution Matrix

Track ID	Category	Status	Configured Tick Command	One-Line Trigger
arc-rig	Competition	SURGING (#61)	python3 core/continuous_engine.py --tick	./run arc-rig
arc-agi-3	Paper + Code	PIVOTING	python3 runner.py	./run arc-agi-3
biohub-cell-tracking	Competition	ADVANCING (#902)	python3 loop/continuous_loop.py	./run biohub-cell-tracking
umud-muscle	Competition	HIBERNATING (#63)	python3 run_eval.py	./run umud-muscle --force
kaggriculture	Simulation	ACTIVE_EXP	python3 simulation/run.py	./run kaggriculture
facetec-spoof-bounty	Bounty ($200k)	STANDBY	python3 audit/probe_liveness.py	./run facetec-spoof-bounty

🔄 Universal Competition Flywheel (`competition-flywheel`)

Autonomous continuous evolutionary loop enforcing multi-tier validation screening and local zero-cost loss autopsies.

Ground Truth Feasibility

The 4-Tier Screening Gate Hierarchy

TIER 0 Static AST

Package & Path Check

AST syntax parsing. Bans host path leaks (/Users/, /tmp/) and ensures complete offline package isolation.

TIER 1 Runtime Smoke

Container Smoke Test

Executes 180s container smoke run under strict Kaggle resource constraints. Verifies valid submission.parquet generation.

TIER 2 CV & MDE

Statistical Benchmark

Disjoint validation folds. Validates score improvement exceeds the Minimum Detectable Effect (MDE) beyond the measured noise floor (σ).

TIER 3 Regression

Champion Pool Check

Tests against historical champion edge-case tasks. Prohibits any candidate that introduces net negative regressions into master.

🔬 Zero-Cost Cognitive Loss Autopsy Protocol

When a candidate fails Tier 1, 2, or 3, the flywheel never guesses or burns cloud tokens. It invokes the local AutopsyEngine:

1. Failure Extraction Extracts top 5 worst-performing episodes or tasks.

2. Local DeepSeek-R1 Derives root cause & broken heuristic invariants (0 cloud spend).

3. Targeted Mutation Local Qwen2.5-Coder synthesizes repaired candidate.

🛡️ Model Health Guard (`model-health-guard`)

Automated diagnostic watchdog intercepting null inference, eval_count == 0, and managing unified memory.

14 Models Available

Self-Healing Fallback Cascade

STAGE 1: INTERCEPTION

Detects eval_count == 0, empty strings, or timeout exceptions during local reasoning steps.

STAGE 2: VRAM FLUSH

Sends keep_alive: 0 to Ollama server to unload VRAM and clear memory deadlocks on Unified Memory.

STAGE 3: TIER FALLBACK

Cascades model down: deepseek-r1:32b → 14b → latest.

Local Probed Model Fleet Telemetry

Model Identifier	Role	Health Status	Latency	Offload Policy
deepseek-r1:32b	Math Invariants & Loss Autopsy	HEALTHY (eval_count > 0)	9.6s	Dynamic (keep_alive: 0)
deepseek-r1:14b	Fast Invariant Derivation	HEALTHY (eval_count > 0)	4.5s	Dynamic (keep_alive: 0)
qwen2.5-coder:14b	Unit Tests & Code Mutations	HEALTHY	3.2s	Dynamic (keep_alive: 0)
qwen3-vl:30b	Visual Grid Mask Verification	HEALTHY	8.1s	Dynamic (keep_alive: 0)

Autonomous Skills & Workflows Hub

🧠 Global Skill Architecture & Cognitive Routing

Adversarial Consulting Consortium

Autonomous Project & Solver Driver

Universal Autonomous Evolutionary Flywheel

Zero-Null Interceptor & VRAM Manager

🏛️ Strategic & Tactical Audit Consortium

Consortium Activation Triggers

⚡ Autonomous Run Project Engine (`project_driver.py`)

Execution Safety & Concurrency Lifecycle

Registered Track Execution Matrix

🔄 Universal Competition Flywheel (`competition-flywheel`)

The 4-Tier Screening Gate Hierarchy

Package & Path Check

Container Smoke Test

Statistical Benchmark

Champion Pool Check

🔬 Zero-Cost Cognitive Loss Autopsy Protocol

🛡️ Model Health Guard (`model-health-guard`)

Self-Healing Fallback Cascade

Local Probed Model Fleet Telemetry

🎯 Bug Bounty Governance & Specialized Skill Suite

bounty-scope-gatekeeper

bounty-dedup-sentinel

bounty-poc-verifier

📊 The 3-Level Task-Based Stagnation Engine

📈 Codebase Maturity & Yield Curve

⏱️ Cadence: Sprint vs. Marathon

⌨️ Quick Execution Command Center

💬 Interactive Chat UI Triggers

💻 1-Word CLI Terminal Shortcuts