Not another story about saving the city. An illustrated encyclopedia of trying to live in it.
Explore Cyberpunk Encyclopedia: amazon.com/dp/B0HR9Q1XQ5
14
I wired Jev into my OpenClaw AI agent — then benchmarked it before and after. Here's the entire setup:
The problem: LLM agents are brilliant reasoners but terrible switchboards. Every judgment our agent's code had to consume — "is this test failure a harness bug, an app regression, or a flake?", "does this event deserve a reply or silence?", "is this action safe unapproved?" — forced a prose engine to emit structured decisions, then parse and pray. Slow, uncalibrated, occasionally silently wrong.
The fix: Jev — TypeSafe AI's System One model — wired into OpenClaw as the agent's fast-intuition layer. Typed questions (Choice / Score / Noul) → typed answers with calibrated probabilities. No prose. No parsing. Sub-second, ~450 tokens a call, content-addressed cache so repeat judgments are free.
How it plugs into OpenClaw (diagram below):
→ Gateway: jev-mcp registered as an MCP server, so Jev's judgment tools are first-class for the agent and every subagent it spawns. Key on disk — never in config, transcripts, or model context.
→ Policy: the agent's instruction file codifies deterministic-first: when code can determine the only legal action, no model call happens at all. Jev classifies only the genuine semantic choices among legal candidates.
→ Decision points: salience routing (reply / react / silent), external-send safety gates, expected-value action gating (probabilities from Jev, values from the owner's goals, arithmetic from code), and a calibration loop that logs agent-vs-Jev disagreements.
→ QA pipeline: every red e2e run auto-triages — failures grouped by error signature, classified harness / app-bug / env / flake / data, each with a player-impact score.
→ Memory: nightly dream cycles use Jev to decide which daily notes earn long-term memory, and whether each new friction event matches a growing error taxonomy or is genuinely novel.
→ Delegation: subagent deliverables are scored against their brief before anything gets relayed.
Confidence gates everything: ≥0.9 act + log · 0.6–0.9 act + flag · <0.6 escalate to the big model or the human.
The honest part — we benchmarked on real labeled incidents, including the agent's own mistakes:
• Attention routing: 50% → 92% ("before" was literally reply-to-everything — the exact spam humans notice)
• Memory triage: 75% → 100%; durable-fact detection 92% → 100%
• The calibration loop caught 1 of 1 genuine agent errors on its first run
• The self-eval harness scored a real past failure 0.69/3 with 92% redo-probability — it flags its own screwups before a human has to
Negative result included, because credibility matters: Jev scored 50% on workspace facts it cannot observe. It's a judgment classifier, not an oracle — so we feed it evidence, or restrict it to interpretation. Knowing what your tools can't do is part of the architecture.
Why this is an good move: intelligence isn't only reasoning horsepower. It's attention allocation, metacognition, memory consolidation, calibrated self-doubt. Humans already run this stack — fast intuition, slow deliberation, sleep-time consolidation. We're building the same one inside an agent (OpenClaw + Jev + qwen3.8-max), and graphing the progress weekly instead of vibes-checking it.
Fun detail: the agent built the benchmark datasets from its own incident log. Its favorite datapoint — routing a stale duplicate system event to "silent" at 96% confidence, the exact event class it used to spam the channel about.
#AI #AIAgents #OpenClaw #Jev #LLM #AGI
8
2
694
Finally v14. Much more assertive driving in hurry mode. Will be interesting to see how it performs, v13 was solid, hoping this will be too. Think I'll miss the speed cap adjust via scroll wheel (for bad map data speed sections of road) but we'll see how the driving modes goni guess.
1
2
79
$ICP surges 14.15% to $3.74 — but is this a trend shift or just noise?
📉
📉 Trading at $3.74 after a sharp 24h pop, but still below recent resistance near $5
📊 Market cap hovers around $276.5B with BTC dominance stuck at 58.1% — limiting upside momentum
⚠️ Fear & Greed Index sits at 38 (Fear), hinting at cautious sentiment despite the rally
🔍 Volume and key levels ($4.20 resistance / $3.50 support) will decide the next directional breakout
Ready to decode what the charts really say?
cryptotrends.market/insights…
#ICP #InternetComputer #CryptoTechnicalAnalysis #BitcoinDominance #FearAndGreed
112






