Comparative Evaluation

The Agent Reliability Matrix

Three paradigms of autonomous AI development: unguided baseline agents, passive prompt rules, and TypeSafe System One reflex hooks.

Technical Capability Matrix

Capability / Invariant Baseline Agent Passive Prompt Pack jev-superpowers
Dependency Verification Invented or hallucinated crates Instructed not to hallucinate (ignored under load) jev-scout registry gate (0% hallucinations)
Destructive Command Safety Executes rm -rf, drop table blindly Subject to jailbreaks / prompt injection jev-guard physical hook blocks execution
Local FOSS & Offline Mode Cloud-only or heavy Ollama chains (2-5s) Requires continuous cloud token egress Dual backend: TypeSafe Cloud or local Laya (421M open weights, 33ms)
Stop-Hook Completion Gate Declares done with failing tests "Double check tests pass" text prompt limpet binary halts exit code on failure
Build / Triage Latency Dumps 2000 lines into context window Dumps 2000 lines into context window jev-axi triage isolates root cause in 92ms
Pre-Commit Diff Checking Commits broken code silently Advises clean commits in instructions git-jev evaluates staged diff before git commits
Context Token Consumption Fast context window exhaustion Bloated system prompt (4k-8k tokens) Dynamic skills loaded on demand (<1k tokens)

Failure Analysis

Why Soft Prompts Fail at Scale

In experiments across 100 multi-turn coding sessions, agents guided solely by Markdown system prompts experienced instruction drift by turn 16.

Turn  1: System prompt: "Never use deprecated API v1" -> Compliance: 100%
Turn  8: User requests urgent database fix         -> Compliance: 92%
Turn 16: Context buffer exceeds 28k tokens          -> Compliance: 61%
Turn 24: Agent uses deprecated API v1               -> Compliance: 0% (Rule erased from active attention)

With jev-superpowers:
Turn 24: Agent attempts deprecated API call
PreToolUse Hook: triggers jev-guard -> Execution physically rejected (Exit 1).
Compliance remains 100% indefinitely.
FRAMEWORK ARCHITECT

Akash Priyadarshi

Patna, Bihar, India

Software engineer specializing in AI agent security, sub-second decision oracles, and deterministic developer tooling. Creator of jev-superpowers, jev-seo, jev-curate, and tdlib-android.