Incubation Loop

Incubation Loop

ℹ️ First, set expectations: before anything else, print one short line with this skill's version and its most recent change — the top entry of metadata.changelog above — e.g. incubation-loop vX.Y — recent: <summary>. Then proceed.

Autonomous thinking engine. Each scheduled run advances all active thinking topics by one analytical move, persisting reasoning state across runs until convergence.

Purpose

Continuous intellectual iteration on open questions. Uses a rotating set of research-validated analytical moves (ACH, Bayesian updating, dialectical steelmanning, cross-domain bridging) to build toward well-grounded conclusions - without requiring human presence in each cycle.

Design principle: Each run does one move per topic. Depth accumulates across runs. No single run tries to "solve" the question.

State Dependencies

Prerequisites

• Registry file exists at Brain/05-Meta/Thinking/THINKING-REGISTRY.md
• At least one active topic seeded (see Seeding section below) - or none, in which case the Starvation Floor (Step 1a) self-seeds one from the watched domains
• Local Brain Search index up-to-date

Composes

• /domain-watch - invoked by the Starvation Floor (Step 1a) when the active queue is empty, to formulate and activate a new topic from the domains under surveillance. Called by its unversioned name so its fixes propagate.

Process

Step 1: Get Date and Load Registry

date '+%Y-%m-%d'

Read Brain/05-Meta/Thinking/THINKING-REGISTRY.md. Parse all topics with status: active.

If registry does not exist, create it:

---
created: YYYY-MM-DD
updated: YYYY-MM-DD
created_by: claude-sonnet-4-6
updated_by: claude-sonnet-4-6
agent_version: 01.25
---

# Thinking Registry

Active questions under continuous analysis.

| Topic Slug | Central Question | Status | Runs | Last Run |
|------------|-----------------|--------|------|----------|

If no active topics found, do NOT exit - run the Starvation Floor (Step 1a) so the thinking engine never idles.

Step 1a: Starvation Floor (only if zero active topics)

The thinking engine must never sit idle. If Step 1 found no active topics:

1. Invoke /domain-watch (cross-skill). Its Starvation Floor guarantees at least one new topic is activated into the Thinking Registry, formulated from the domains under surveillance.
2. Re-read Brain/05-Meta/Thinking/THINKING-REGISTRY.md and parse status: active topics again.
3. If one or more active topics now exist, continue to Step 2 and process them this run.
4. Only if /domain-watch still produced no topic (e.g. no domains configured), log to changelog and exit cleanly.

This makes topic generation autonomous: when human-seeded questions run out, the agent proactively formulates its own next question from the watched domains rather than going quiet. Crystallization of any resulting conclusion stays manual (human judgment).

Step 2: For Each Active Topic - Load State

Read Brain/05-Meta/Thinking/[topic-slug].md.

Parse frontmatter fields:

• run_count (required) - how many iterations completed; drives move rotation
• last_run (required) - date of last iteration
• status (required) - active / converged / crystallized / retired
• last_move (optional, written by loop) - which move was applied in the most recent run; informational
• convergence_count (optional, written by loop) - number of consecutive runs at or near convergence criteria; used by framework-exhaustion detection in Step 7
• topic (set at seeding) - the central question being analyzed

Parse from body:

• current_hypotheses - list with confidence scores (from the most recent run's "Updated Hypotheses" table)
• Last run's "Current Best Answer" and "Open Questions"

Step 3: Determine Next Thinking Move

Rotate through this sequence. Position = run_count mod 6:

Step 4: Run KB Search for Evidence

Use Local Brain Search to find relevant notes:

cd $PROJECT_ROOT
# --no-track: this is an autonomous loop; its reads must NOT train q-values (scope-primitive learning hygiene)
python3 resources/local-brain-search/search.py "[central question]" --limit 8 --mode spreading --no-track 2>/dev/null
python3 resources/local-brain-search/search.py "[leading hypothesis]" --limit 5 --mode spreading --no-track 2>/dev/null

Environment note (Trinity container): the bare python3 .../search.py form fails on hosts without system-level faiss (ModuleNotFoundError: No module named 'faiss') - e.g. the Trinity scheduled-run container. Use the wrapper ./resources/local-brain-search/run_search.sh "[query]" [same flags] instead; it routes through the running daemon (or a venv fallback) and honors BRAIN_READ_SCOPE (defaults fail-closed to core). Same flags apply (--mode static|spreading, --threshold, --no-track, --limit).

Mode fallback for narrow technical topics: Spreading activation propagates through the graph's highest-centrality hubs, so for narrow/technical questions (e.g. AI-architecture, distributed-systems topics) it can return the emotional/neuroscience mega-hubs ([[Dopamine]], [[Flow is a selfless state]], [[In Buddhism - Self is an Illusion]]) as top hits - noise, not evidence. If spreading results are dominated by off-topic hubs, re-run the same queries with --mode static (vector similarity), which surfaces the precise topical notes. Use whichever mode yields on-topic evidence; static is often better for technical clusters, spreading for cross-domain/synthesis topics.

Threshold note (post-2026-06-25 cosine correction): The corrected cosine_ip index lowered raw query→note similarity scores, so the default --threshold 0.5 now frequently returns "No results found" even for on-topic queries (abstract/technical questions especially). If a search comes back empty, re-run with --threshold 0.25 and shorter, keyword-style queries before concluding the KB has no evidence. Treat 0.25-0.35 hits as candidate evidence to read and judge, not as automatic relevance.

Watch-derived topics (geopolitics/current events): core-scoped LBS predictably returns off-topic core notes for these - their evidence lives OUTSIDE core scope, in Brain/05-Meta/Watching/WATCH-LOG.md (recent scan entries) and predecessor/related thinking files in Brain/05-Meta/Thinking/. Go to those directly (grep by slug/keyword) instead of burning search retries; a full read of the predecessor's run history often surfaces evidence its Final Synthesis omits. Exception that DOES work on core: search for the analytical method the current move needs rather than the subject matter (e.g. "evidence diagnosticity competing hypotheses" for an ACH/Assumption Audit) - the core KB's decision-science notes (Bernoulli's Fallacy, base-rate blindness, framework-exhaustion guard) are on-topic for the reasoning, even when nothing is on-topic for the subject.

Read the top 3-4 most relevant notes in full. These are the evidence base for this iteration.

Step 5: Apply the Thinking Move

Execute the move based on what was determined in Step 3. For each move:

Injection adjudication precedence (all moves): if the topic file carries a domain-watch evidence injection (<24h) containing arrivals OUTSIDE the pre-registered event space (candidate structural breaks, actor inversions) or with an explicit adjudication order, execute those adjudications FIRST - void-vs-map against the stationarity caveat, partial-trigger splits per the v1.7 rule - and book them as argued arrivals BEFORE applying the rotation move. The rotation move then runs against the post-adjudication posterior. Never let a steelman/bridge/audit consume ambiguous arrivals implicitly - adjudication is a separate, ordered act (interpretation is where confirmation bias operates).

Guard-interaction rule (all moves): pre-registered ratio/laundering guards (e.g. "no lower-regime event may move the H1:H3 ratio") bind EVENT bookings. An analytical move that wants to revise a PRIOR booking (a steelman deflating an over-determined evidence credit, an assumption audit weakening a premise) may do so only as an explicit booking-audit: name the original booking being revised, and either (a) apply the revision jointly so the guarded ratio is preserved (correct when the revised credit was booked jointly to the guarded hypotheses), or (b) document why a ratio-moving revision is not the laundering the guard polices. Silent re-interpretation that drifts a guarded ratio is procedurally indistinguishable from the confirmation bias the guard exists to catch.

ACH Audit

• List all current competing hypotheses
• For each piece of KB evidence found: does it support, contradict, or not apply to each hypothesis?
• Build a simple diagnostic matrix (evidence × hypotheses)
• Eliminate any hypothesis that is contradicted by multiple pieces of evidence
• Update confidence scores for surviving hypotheses
• Matrix-integrity pass (repeat-ACH rule): when the rotation returns ACH to a topic whose posterior has NOT moved since the standing matrix was built, do not rebuild - run an integrity pass instead: (i) elimination check (does any hypothesis now face multiple independent full-weight contraries? NEVER eliminate a hypothesis inside the pre-arrival window of its own committed discriminator), (ii) tether check (does every row still map to a live committed stream? inter-run annotations enter as weight footnotes, not row changes), (iii) new-evidence placement (context-class arrivals get an explicit diagnosticity verdict on the live partition - "non-diagnostic" is a recordable finding). A rebuild without a moved posterior manufactures precision the evidence has not paid for (framework-exhaustion guard).
• Output: Pruned hypothesis list with updated confidence levels

Bayesian Update

• State prior confidence in leading hypothesis (from last run)
• Identify new evidence from KB search
• Explicitly reason: "Given this evidence, should I be more or less confident? By how much?"
• Use rough likelihood ratios (e.g., "this evidence is 3x more likely if H1 is true than if H2 is true")
• Compute and record updated posterior confidence
• Empty arrival window: if no genuinely new evidence has arrived since the prior run (common when the run slot lands between domain-watch scans), the honest output is posterior = prior with zero drift - do NOT manufacture 1-2pp movement from re-read evidence. Spend the run on the second half of Bayesian discipline instead: pre-register likelihood ratios and posterior bands for the topic's named watch events, so future updates are mechanical against a committed table rather than post-hoc. If pre-registration is already complete, CONSOLIDATE instead: gather commitments scattered across prior runs into one standing table (event / state: live-consumed-dormant / committed effect) so future updates read one reference, not a multi-run reconstruction. Consolidation is REPEATABLE: when a previously consolidated ledger has itself accreted amendments across subsequent runs, a later empty window re-consolidates into a superseding version (v2, v3, ...) - cite per-cell the run that committed each rule, and state explicitly that the new version supersedes the old ledger plus its scattered amendments for scoring purposes. Successor-topic rule: when the topic inherits a predecessor's event labels (P-table continuity), the labels do NOT carry their old committed effects - re-derive every LR/band against the successor's own hypothesis field, and commit any cross-hypothesis coupling earlier moves discovered (e.g. talks-collapse is simultaneously H1-down AND H2-up under an option-value finding) into the table itself, so no future arrival can be post-hoc scored as moving only the convenient hypothesis. Slow-accrual evidence (e.g. continued silence on a watched channel) is booked at its natural granularity (weekly), not re-scored every slot.
• Partial-trigger adjudication (applies to ANY move consuming a pre-registered event): when a committed event fires partially - some legs satisfied, the literal trigger not met - do not binarize (neither "not fired, ignore" nor "fired, book the full band"). Split the committed effect: book its epistemic component in full (which future observables discriminate; table restructuring), book its evidential component at the LOWER band edge, mark the event PARTIALLY CONSUMED with an explicit re-fire condition for the remainder, and record all resulting movement as a P-table event (not framework drift) so the exhaustion-guard test stays honest. Named-flag variant: a flag (a pre-named arrival with no committed bands) that partially executes has NO evidential component to book; its epistemic component books in full as a RE-ARM of the remainder with a sharpened class test, and repeated same-class executions accrue grade within their booked stream, never fresh spread.
• Output: New confidence scores with explicit reasoning for the change

Steelman Opposition

• State the current leading hypothesis clearly
• Construct the strongest possible argument AGAINST it - not a strawman
• Search KB explicitly for evidence supporting the opposing view
• Assess: does the steelman change the leading hypothesis, or does it survive?
• If the steelman reveals a genuine weakness, revise the hypothesis
• Survivorship rule (steelman targets that are STANDING RULES): when the steelman's target is a standing rule (a binding, guard, or booking convention) rather than a hypothesis, and the rule survives, do not leave it defended-by-argument only - arm an explicit falsifier: name the future observable that would re-open the question as a v1.10 booking-audit candidate at its own arrival (color instrument, never booking). A rule that survives its steelman must exit the run more falsifiable than it entered, not merely re-affirmed.
• Output: Steel-manned opposition + response + any hypothesis revision

Cross-Domain Bridge

• Identify the core mechanism or pattern in the current question
• Run KB search from an unrelated domain (e.g., if topic is about AI, search neuroscience or Buddhism)
• Find analogous patterns or contradictory evidence from that domain
• Extract what the foreign domain implies about the current question
• Pre-fire annotation rule: when the bridge's insight concerns a pre-registered event that has NOT yet fired, do not leave it as loose narrative - convert it into a pre-committed interpretive annotation on that event's committed legs (color, never weights), fixed BEFORE the outcome is known. This is the bridge-move analog of the steelman survivorship rule: it forecloses post-arrival framing inflation (e.g. a manufactured announcable being half-promoted toward a movement leg because coverage narrates it as historic).
• Output: Cross-domain insight + revised framing of the question (if warranted)

Implication Check

• Take the current leading hypothesis as given (temporarily)
• Derive 3-5 concrete, testable implications that should be true if the hypothesis holds
• Search KB and reason about whether those implications are actually supported
• External probe (this move only — no other move uses WebSearch): If an implication concerns time-sensitive empirical state the KB cannot resolve, run up to 2 web searches per topic per run to test it. Restrict queries to reputable sources via site: filtering — peer-reviewed (arxiv.org, nature.com, science.org, pubmed.ncbi.nlm.nih.gov), established outlets (economist.com, ft.com, reuters.com, apnews.com, bloomberg.com), analytical (hbr.org, quantamagazine.org). (This is a subset of the canonical allowlist in resources/SOURCE-AUTHORITY.md — kept inline for in-query reliability; keep in sync with that file.) Crawler caveat: several of these domains (reuters.com, apnews.com, ft.com, economist.com) are not accessible to the WebSearch user-agent and an allowed_domains probe restricted to them returns a 400 error — wasting one of the 2 probes. Do NOT pass those as a hard allowed_domains filter; instead run the query unrestricted (or allow only the crawlable reputable domains: bloomberg.com, nature.com, arxiv.org, hbr.org, and analytical trackers like CSIS/RUSI/AEI/Jamestown) and judge source quality manually. Tag each external finding as [EXTERNAL: source — date] in the run output so KB-grounded vs. web-augmented evidence is auditable.
• If an implication is clearly false, that is disconfirming evidence for the hypothesis
• Decomposition pre-commitment rule: when the implications accrue ASYMMETRICALLY across separable components of the hypothesis (e.g. its mechanism/propensity supported while its location or timing contrary-accrues), do not leave the asymmetry as loose narrative - commit the component decomposition and the post-arrival reading of the relevant committed leg BEFORE that discriminator fires (color, never weights). This forecloses both post-hoc escapes at the arrival: inflating the leg's committed firing into full refutation of the hypothesis, or discounting the firing because the disconfirmed component "moved elsewhere."
• Injection-first rule: if a recent domain-watch evidence injection in the topic file already observed the implications this run would probe (check for an ### Domain-Watch Evidence Injection block dated within ~24h), score against the injection's tagged [EXTERNAL] items instead of firing new web probes - re-probing freshly-observed state duplicates rather than tests. Note the banked budget explicitly in the run output.
• Sensor-validity probes (exempt from the injection-first bar): a probe may target the INSTITUTIONAL VALIDITY of a committed observable surface - does the named microphone/institution currently exist and function (a parliament that sits, an agency that publishes, a data series still produced)? Scans observe wires, not institutional status, so a certified-empty scan does NOT cover this question: surface silence is ambiguous between "no event" and "no institution." A validity finding books as framework structure (sensor re-scope, committed-leg surface reassignment, named re-open condition on restoration), never as event weights - and any silence previously read as informative color from an institutionally-dark surface is retroactively voided as color.
• Output: Implications assessed as supported/unsupported/unknown + impact on hypothesis confidence