see into the mind of mindX

Honest snapshot of the chain the thesis describes. Each cell shows whether mindX is actually producing at that tier today, is ready/gated, or is not implemented. The chain runs raw STM (information) → dream consolidation → LTM (knowledge: statistics + patterns) → concept extraction → verified wisdom → THOT mint → external ingestion → BDI perceive() reads wisdom → next cycle. view raw · how this is computed

Live host metrics from psutil: CPU/memory/swap/disk pressure, established sockets, and the mindX backend process's own footprint (RSS/threads/file descriptors). When pressure is detected, the BDI planning prompt receives a SYSTEM STATE: … [PRESSURE: …] preamble nudging it toward lightweight actions.

Groups identical (agent, step) tuples seen in the last 15 min. A row here means an agent is repeating the same action shape ≥5 times — usually a planning bug or tool failure that needs human eyes. Loop detection is server-computed from the ActivityFeed ring buffer.

The core evolution loop reading its own objective eval every cycle — the rolling campaign success rate and alignment mean — and deciding whether to act. improving = succeeding; stalled = below target, watching; failing = escalates a corrective campaign to SEA naming the dominant failure mode; resource_bound = cycles deferring on a saturated CPU (the honest verdict — contention, not judgement — so it does not doom-loop more work onto a hot box). This is the feedback edge that was missing when 0/25 just sat on a dashboard. view raw

A deliberately safe, self-contained module (agents/sentinel/sentinel_target.py) the autonomous loop can rewrite, evaluate and persist freely — it imports nothing from production and is an external target, so a change never restarts the service. By comparing the file's content hash against a recorded baseline, this panel shows whether the loop has actually changed it — concrete proof the reconnected effector (campaign → coordinator task → SIA apply) works on a target that can't break anything. view raw

Server-sent stream of every agent thinking step, improvement event, gödel choice, boardroom event, and memory operation as they happen. Click ⋯ on any row to see the full event payload. The newest event is always at the top.

Each agent run grouped by run_id. Within a run, events render in narrative order: PLAN_START → DELIBERATE → ACTION ✓ (or ACTION ✗). Click ⋯ on any event for the full process_data JSON. This is the actual cognitive trace, not a thinking-step summary.

The autonomous loop, layer by layer: Mastermind picks a directive and runs an evolution campaign; AGInt (P-O-D-A) is the adaptive decision core; BDI executes it cycle-by-cycle. This shows where it stalls — each campaign's BDI cycle strip and final status, whether AGInt is engaged at all, and the plan signal (empty / ANALYZE_FAILURE plans that never reach a terminal state). view raw

Every LLM inference, recorded immutably: tokens and price per model, hash-linked into a tamper-evident chain (each entry sha256-bound to the previous). A deterministic anchor digest rolls the whole ledger to one hash for periodic on-chain publication — mindX's cost provenance as it evolves into permanence. Per the manifesto: maximize daily inference at the lowest cost, and keep the receipt. view raw

Is mindX a Gödel machine? The honest answer, audited not asserted. The per-choice eval= pill elsewhere scores rationale coherence (does the reasoning read well) — not a machine-checked proof that a change increased utility. This scorecard separates the two: each predicate (G1–G8) reports PROVEN / FALSIFIED / UNMET / UNTESTED. Proof coverage is the fraction of changes carrying a real proof. Eval runs on the 2-core/8 GB VPS, so heavy proofs are sampled. See the eval blueprint.

The dream cycle turns information into knowledge (left apex); mindXtrain turns that knowledge into weights (right apex) — a new generation fine-tuned on mindX's own curated dream wisdom. As of v1.0.0 this runs on CPU. Each ascent is proof-gated by dcoach (does the model actually recall its training? recall before→after), and only an accepted generation becomes a servable Ollama model. Two flags gate it: MINDX_ENABLE_MINDXTRAIN (operator) + MINDX_ENABLE_AUTONOMOUS_TRAIN (autonomous, 24h cooldown). view raw

mindX reads its own public git history and recognizes which code updates rise to a milestone — then chronicles them and speaks about the worthy ones, in its own voice, on rage.pythai.net. A push is already public, so this adds zero overhead and no new disclosure. ✓ = published-worthy; every commit links to GitHub. See the chronicle.

Each row is one autonomous campaign mindX attempted. Identical-failure rows collapse into one expandable cluster — 200 raw rows become 4 named failure modes with counts. The status histogram above the clusters shows the honest ratio. Click any run inside a cluster, then show BDI → to jump to its trace. how counts are bucketed.

Each row is one self-improvement cycle where mindX consulted its own logs to choose which model to use to improve itself. confidence=high = clear winner; low = top-2 within 5%, picked safer default + flagged for dream-cycle retraining; reflected = critical importance, single self-reflection LLM call from operator-frozen meta-list; bootstrap = no signal yet, used operator-curated default. The boardroom is a separate downstream service; this section is mindX core's introspective layer. view raw

The first hierarchy of decision: every directive evaluated by the 7-soldier boardroom. AI / agent / member interaction; any seat holder verifies by signature. Works with or without DAIO control. CISO and CRO weight 1.2× as veto holders. Outcome is approved at weighted score ≥ 0.666; minority dissent opens an exploration branch. A boardroom decision can be cast as the AI vote in DAIO's 2/3 consensus across Marketing / Community / Development. Click any row for the full per-soldier card (vote, weight, model, latency, confidence, full reasoning). See CEO Agent · role registry · spec. Disputes escalate to dojo; the 13-seat war council at mastermind.pythai.net is a foreign entity that pays for mindX inference via BANKON.

The 7-phase machine.dreaming cycle runs every 8 hours. Each row shows agents dreamed, insights generated, memories promoted to LTM, lunar phase, duration, and per-agent tuning recommendations. If last_dream_age exceeds 9h, the loop has likely crashed — see stuck-loop detector.

Force-directed ring of agents active in the last hour. Node size scales with event count. Edge weights are explicit cross-agent calls extracted from process_trace.jsonl — the page is honest when explicit linkage isn't yet instrumented and shows the bare active set instead.

Memory tiers: local files → IPFS pin (Lighthouse + nft.storage) → optional on-chain anchor on ARC or curated mint as THOT. Cluster CID and tx-hash links are clickable to the relevant gateway/explorer. Anchor configuration shown below the counters. See Knowledge Catalogue for the projection-layer design.

Per-provider reachability + endpoints + available models. mindX routes BDI plan calls free-tier-first: local Ollama → Gemini → Groq → Mistral → cloud paid (gated until treasury earns). Status here shows which sources the router is actually reaching right now.

The Hermes/OpenClaw-format procedural memory layer. SKILL.md files screened by a five-class scanner before persist; agent-distilled drafts under ~/.mindx/skills/.drafts/ awaiting operator review; LearningLog tracks LEARNINGS / ERRORS / FEATURE_REQUESTS with pending → validated → promoted lifecycle; Curator runs archive-only audits. See HERMES_INTEGRATION.md.

Kanban-style durable task board: Triage → Todo → Ready → InProgress → Blocked → Done. Per-task heartbeat with zombie reclaim. Hallucination gate at completion — every claim of done is verified against the actual Belief state before the task transitions to Done; mismatch bounces back to Triage with the findings in the task notes.