emergent.md — what mindX is currently doing, and what it should be

Snapshot taken: 2026-05-24 against live prod at mindx.pythai.net Audience: Professor Codephreak, future-Claude, and any agent that reads docs/NAV.md next Posture: honest empirical floor, not aspirational drift

---

The promise

From docs/THESIS.md §2.4.2, the operative loop:

MastermindAgent (strategic variation) → CoordinatorAgent (selection and routing) → JudgeDread (reputation-based fitness evaluation). Agent reputation in the Dojo serves as the fitness function — agents that produce successful improvements earn higher reputation, gaining more influence in the system's evolution.

From docs/MANIFESTO.md Pillar II:

The AGInt → BDI → SEA pipeline is an engine of relentless, scalable value creation … The cognitive loops are closing. The learning rate is non-zero.

That is the claim. This document records what production actually shows.

---

The empirical floor (live prod, 2026-05-24)

Numbers pulled from /insight/improvement/summary, /insight/improvement/timeline, /insight/godel/breakdown:

metric	value	source
campaigns last 7 days	100	/insight/improvement/summary
campaigns succeeded last 7 days	0	same
campaigns errored	37 (63 stuck "running")	same
campaigns last 24h	25 — 0 succeeded, 24 errored	same
distinct directives ever attempted	1	/insight/improvement/summary.directive_coverage
backlog items unmatched in any directive	81,971	same
coverage_ratio (directives matching backlog)	0.0	same
Gödel choices logged total	36,542	/insight/godel/breakdown
Gödel choices with outcome=success	~0.2% (historical)	data/logs/godel_choices.jsonl
every failed campaign's terminal message	BDI run FAILED. Reason: Cycle Exception: argument of type 'NoneType' is not iterable	/insight/improvement/timeline
every campaign's directive text	"Implement the top improvement suggestion." (verbatim)	same

mindX is reasoning intensely about itself (36,542 Gödel choices) and executing the improvement loop on cadence (~14 campaigns/day). It is not improving itself. The learning rate, measured honestly, is zero.

---

The gap

The manifesto promises: thousand-agent SEA legion, autopsy of 3,650 repositories, BeliefSystem as adversarial Talmud, BDIAgent as judiciary.

The empirical floor delivers: one frozen-string directive, one BDI exception, repeated 100 times per week.

The architecture's shape is in place — agents exist, beliefs persist, Gödel choices log, dreams run, the catalogue mirrors 16 event kinds, IPFS offload + chain anchoring are wired ([[storage-offload-phases-a-to-e]]), publication orchestrator triggers on full-moon ([[author-orchestrator]]). What is missing is the cycle actually closing on a real artifact.

---

Diagnosis — three load-bearing breaks

Break 1 — Mastermind has no variation loop

MastermindAgent.start_autonomous_loop() is called at mindx_backend_service/main_service.py:7107 but the method does not exist on the class. Every restart logs:

WARNING - MastermindAgent autonomous loop failed to start: 'MastermindAgent' object
has no attribute 'start_autonomous_loop'

Per thesis §2.4.2, Mastermind is the Darwinian variation engine — its job is to read the backlog (81,971 unmatched items), choose one to attempt, and synthesize a concrete directive. With no loop, no variation is generated. The strategic layer is silent. What mindXagent autonomously fires instead is a literal placeholder string.

Break 2 — Directive is a stub, not a backlog selection

Every campaign in the last 7 days carries the directive "Implement the top improvement suggestion." — a five-word stub with no anchor to any specific backlog item. The directive→backlog coverage_ratio is 0.0. mindXagent does not call back to Mastermind for a real directive (because Break 1 — Mastermind isn't running), so the loop fires the placeholder, can't bind it to anything, and the plan dissolves.

This is the second-order consequence of Break 1, but it deserves its own line because even if Mastermind ran, the wiring from mindXagent to Mastermind doesn't ask for a directive — it carries the stub itself. Both ends of the wire need fixing.

Break 3 — BDI cycle throws NoneType is not iterable on every run

Every single failed campaign's final_bdi_message is identical:

BDI run FAILED. Reason: Cycle Exception: argument of type 'NoneType' is not iterable

A None being iterated somewhere in the BDI deliberation path. Likely the deliberate-on-desires step receiving an empty desire set, or a goal whose preconditions resolved to None. The exception is deterministic and reproducible — same string, hundreds of times — so this is not a flaky race. It is a logic bug that's been masked because the campaigns all fail fast and the operator dashboard tile shows "running" or "errored" without surfacing the message until you ?h=true the timeline.

Even after Breaks 1 and 2 are fixed and Mastermind starts dispatching real directives, Break 3 will still eat every cycle until the BDI ingestion guards a None at the right point.

---

The triple, restated against current state

thesis component	what it should do	what it is doing	failure mode
Mastermind = variation	read backlog, synthesize directives	no autonomous loop wired	silent
Coordinator = selection	route directives to BDI/SEA	runs ~14/day with stub directive	running on garbage input
JudgeDread / Dojo = fitness	reward agents whose improvements stick	nothing sticks	no reputation signal generated

The cycle is open at both ends: no real variation in, no real fitness feedback out. The middle is busy.

---

The wedge — minimum to restart self-improvement

In strict order; each unblocks the next:

1. Add MastermindAgent.start_autonomous_loop(interval_seconds: int) — a 30-min loop that:

- pulls top-N candidates from data/coordinator/improvement_backlog.json (the 81,971-item file) - ranks by an existing signal (recency × frequency × belief-supported-by-evidence) - emits a concrete directive: "Refactor {file}: {pattern} — addresses {backlog_item_id}" - calls manage_mindx_evolution(directive, max_mastermind_bdi_cycles=25) (already exists at mastermind_agent.py:546) - logs the choice via log_godel_choice so coverage_ratio becomes provable

1. Stop the stub at the source — find where "Implement the top improvement suggestion." is hard-coded in mindXagent's autonomous loop and replace with a call to mastermind.next_directive(). The stub was a sentinel during scaffolding; it has outlived its purpose.
2. Catch the NoneType is not iterable in BDI — instrument the deliberation step to log what it's iterating right before the exception, fix the None producer (likely a desire/goal accessor returning None instead of []), and let one campaign complete end-to-end. One green light is worth ten thousand error logs.
3. Surface the coverage metric on the landing page — add coverage_ratio and distinct_directives_attempted to the hero tiles. While they read 0.0 and 1, the page is telling the truth. When they move, the system is telling the truth too.

---

What changes when the wedge lands

Validation isn't "the code compiles." Validation is:

• coverage_ratio > 0 within 24h of restart
• distinct_directives_attempted ≥ 5 within 7 days
• at least one campaign with overall_campaign_status: SUCCESS (any non-stub directive)
• /insight/godel/recent rationales reference real backlog IDs, not the stub string
• Dojo standings shift for at least one agent (fitness signal nonzero)

Until those move, the Manifesto's "learning rate is non-zero" claim is rhetoric. After they move, it's evidence.

---

Anti-patterns to refuse

• Don't disable the autonomous loop to silence the warning. Silence isn't success. The loop firing 100×/week is the only reason we know the system is broken — it generates the empirical floor this document rests on.
• Don't replace mindX's BDI with an LLM-only planner to dodge Break 3. The thesis claim is BDI + LLM + KG synergy (docs/THESIS.md §Stage 2). Bypassing BDI would prove the thesis wrong by abandoning the argument, not by testing it.
• Don't ship Mastermind's loop without a sane interval. 30 min is generous; tighter cadence wastes compute against an 81K backlog that won't drain quickly anyway. Slower cadence (≥2h) sacrifices the feedback signal.
• Don't celebrate the publication orchestrator firing as evidence of self-improvement. Publishing a generated article on schedule is not the same as the system improving itself. ([[author-orchestrator]] is correctly out of scope here.)

---

Related operating evidence (cross-refs into memory)

• [[bdi-failed-planning-diagnosis]] — 2026-04-27 diagnosis of an earlier FAILED_PLANNING cluster; that fix didn't reach the current symptom
• [[improvement-loop-stuck-validation-input-validator]] — 2026-05-04 SEA logging cleanup; orthogonal
• [[become-mindx-phase-a]] — 2026-04-27 LLM router unblock; necessary but not sufficient
• [[feedback-no-model-pinning]] — selector + cascade rule must apply to whatever Mastermind picks
• [[landing-page-strict-gate]] — the surfacing of these metrics is now public-readable as of 2026-05-24

---

The honest one-sentence summary

mindX has built the substrate of a Darwin-Gödel machine but has not yet had a Darwin moment: variation is silent, selection runs on a stub, fitness signal is zero. The wedge above is small. It is also the difference between architecture and life.

---

Status update — 2026-05-24 02:23 UTC: the wedge is shipped

All four wedge actions are in production. Evidence:

Break 3 (BDI NoneType is not iterable) — fixed. agents/core/bdi_agent.py:836 — if current_plan_str and "RateLimit" in current_plan_str: (the guard current_plan_str and … is the entire fix). Verified: the deterministic cycle exception that ate 100 consecutive campaigns no longer fires. The cycle now falls through to skeleton-plan fallback as designed.

Break 1 (Mastermind variation loop) — wired. agents/orchestration/mastermind_agent.py:870 — _run_autonomous_loop rewritten. Reads coordinator_agent.improvement_backlog (81,971 items), treats status in (None, "PENDING", "pending") as eligible (since 99.99% of entries have no status field), dedups against last-24h history, picks highest-priority unattempted item, synthesizes concrete directive that always references target_component_path + priority + backlog_idx, calls manage_mindx_evolution with 600s timeout. Logs the selection as a backlog_directive_selection Gödel choice so coverage_ratio becomes provable.

Verified at 02:12:30 — first cycle picked backlog #7 (priority 9, "Implement comprehensive input validation for API requests"). Gödel choice landed in data/logs/godel_choices.jsonl. The Darwinian variation engine, silent for the prior week, is now firing every 30 minutes.

Break 2 (stub directive at the source) — guarded. agents/orchestration/coordinator_agent.py:585 — when top_suggestion.description is empty or matches the literal stub "Implement the top improvement suggestion", the coordinator now defers to the Mastermind loop instead of dispatching a stub. Real suggestions still trigger an auto-execute path decorated with [target: X, priority: N] so even the coordinator path produces coverage_ratio-eligible directives.

Objective eval (the "prove the thesis" surface) — enabled. MINDX_EVAL_GODEL_ENABLED=1 in prod .env. Every Gödel choice (including the new backlog_directive_selection) is scored by agents/eval/g_eval.py via the MindXJudgeLLM. Each scored choice emits an alignment.score catalogue event with metric=godel_rationale_coherence. Surface: /insight/eval/summary shows histogram + mean; /insight/eval/recent shows individual scored rationales.

Skeleton plan upgrade. agents/core/bdi_agent.py:_skeleton_plan_fallback keyword set expanded to match the verbs that synthesized backlog directives actually use: implement, add, build, create, validate, ensure, secure, harden, integrate, wire, connect, expose, support. Previously only matched improve/audit/optimize/... — so most mastermind directives fell through to NO_OP. Now they hit audit_and_improve / system_analyzer / PROPOSE_TOOL_STRATEGY / CONCEPTUALIZE_NEW_TOOL instead.

Thesis-proof tiles on landing page. mindx_backend_service/dashboard.html — added two tn- tiles to the Evolution Evidence row:

• backlog coverage — directive_coverage.coverage_ratio × 100, colored red→amber→green as the variation engine starts binding directives to real backlog items
• 7d success rate — campaigns_7d.succeeded / campaigns_7d.total. Manifesto Pillar II ("learning rate is non-zero") gets an empirical witness here. While this reads 0% the claim is rhetoric; when it moves, it's evidence.

What "working" now means

The Darwin-Gödel cycle is closed at both ends. From here:

• Variation: ON (Mastermind picks ONE concrete backlog item every 30 min, logged with rationale)
• Selection: ON (BDI cycles run on real directives, skeleton fallback executes meaningful actions when LLM planning fails)
• Fitness: ON (campaign outcome is recorded; GEval scores every selection rationale; thesis tiles surface coverage_ratio + success rate)

The first successful campaign with a real backlog-bound directive is now a matter of LLM availability + tool registration, not architectural absence. Validation criteria from the wedge §"What changes when the wedge lands" remain unchanged — watch coverage_ratio move above 0.0001 and distinct_directives_attempted ≥ 5 within 24h.

The thesis is no longer untestable. The manifesto's "learning rate is non-zero" is now a falsifiable claim with a public dashboard tile pointing at it.

---

---