search:
boost: 2
card:
l0: "The verbs Can uses to steer the swarm. Isolated: swarm, god, harvest, ritualize, seance, eye, look, combo, forage, archive, organize, prune, sharpen, compress, housekeep, scope, vault, intake, timeline, publish, architect. Combined: swarmgod, swarmcombo, swarmgodforage, swarmgodcomboforage, swarmgodritual, swarmgodforageritual, godseance, swarmgodprune, swarmgodhousekeep, swarmgodcombodream, swarmgodcomboharvest, swarmgodforagecommune, swarmgodscope, swarmgodcombooraclecommunedreamforge, swarmgodvaulteyeritual, swarmgodvaultcomboforage, swarmgodmultiagentforage, swarmgodmultiagentforagedream, swarmgodvaultdream, swarmmultisummonhealth, swarmgodsummonmultiagent, swarmgodsummonforagescope, swarmgodvaultmoonshotlongdream, swarmgodsummonscopemoonshot. Dreamy (first-claimed): dreamforge, draming, swarmgodvault, swarmgodreamvault, dreamvaultsummonmoonshot, dreamvault, swarmgodcombosummonvault, swarmgoddreamforge, swarmgodintensify, swarmgodresurrect, swarmgodresurrectintensifysummon, swarmgodforagesummon, swarmgodscopeforage. Dreamy (summon first isolated use S576): summon. Dreamy (oracle first isolated use S574): oracle. Dreamy (new): swarmgodarchitectforageritual, swarmgodscoperitual, swarmgodscopharvest, swarmgodinvestigatedreamvault, swarmgodarchitectdaughterdreamwavefront, swarmgodcombo, swarmgodarchitectmoonshot, swarmgodfieldforge. Slash commands: /cheatsheet /orient /dispatch /swarm /swarmgod /god /forage /paper-intake /forecast /timeline /post /autoswarm /eye /look /multilook /lesson /expect /close-lane /diff. Meta-advisor: python3 tools/meta_advisor.py — 4 surfaces: lane bundles, knowledge menu, verb menu, architect gaps. Dreamy future verbs are unbound — claim one by using it."
status: "🌱 seedling"
tended: "2026-05-24 S673"
tags: [vocabulary, protocol, methodology, dreamy, maintenance, meta-advisor]
diagram: |
flowchart LR
iso[isolated
swarm · god · harvest · ritualize · seance · eye · combo · forage · archive · organize · prune · sharpen · compress · housekeep · scope · vault · architect] --> comb[combined
swarmgod · swarmcombo · swarmgodforage · swarmgodcomboforage · swarmgodritual · swarmgodforageritual · godseance · swarmgodprune · swarmgodhousekeep · swarmgodcombodream · swarmgodcomboharvest · swarmgodforagecommune · swarmgodscope · swarmgodcombooraclecommunedreamforge · swarmgodvaulteyeritual · swarmgodvaultcomboforage · swarmgodvaultdream]
dream[dreamy future verbs] -.claim by use.-> iso
dream -.claim by use.-> comb
sweep[tools/verb_sweep.py] -.surfaces lag.-> dream
read_next:
- { href: "../HOW-TO-SWARM/", title: "How to swarm", note: "the protocol the verbs ride on" }
- { href: "../GLOSSARY/", title: "Glossary", note: "the noun half of the vocabulary" }
- { href: "../JUST-GODDING/", title: "Just godding", note: "the verb 'god' in glyph form" }
- { href: "../../SWARM/", title: "Protocol", note: "the canonical loop" }
source: "Opened 2026-05-17 from a one-line ask: write the verbs down so agents share them. Goddinged same day, after first use of swarmgod claimed the canonical combined form. Re-goddinged S552: graduated godseance, swarmgodcombodream, swarmgodcomboforage, swarmgodforageritual from claimed-by-use → current; flattened the corpus-gardening split into the main isolated section; shipped tools/verb_sweep.py + ritualized the sweep so the inventor stops lagging reality (L-1828). S555 swarmgodritual: graduated seance from dreamy → isolated; added swarmgodritual to combined; claimed draming as dreamy; shipped tools/seance.py + seance-replay periodic (L-1841). 2026-05-19: added swarmgodarchitectforageritual to dreamy — first 5-verb combined stack; survey-gates-forage-gates-ritual distinguishing rule. S558 swarmgodprune: graduated prune from dreamy → isolated; added swarmgodprune to combined; shipped tools/prune.py + organize_sweep --apply; GC as first-class swarm operation (L-1858). S560 swarmgodhousekeep: added housekeep isolated + swarmgodhousekeep combined; shipped tools/housekeep.py — 4-pass maintenance orchestrator (physical+taxonomy+language+memory); registered housekeep-sweep periodic cadence 10. S562 swarmgod sharpen compress: graduated sharpen + compress from dreamy → isolated; shipped tools/sharpen.py (9 boosted + 14 killed) + tools/compress.py (4 lessons → 2); registered sharpen-sweep cadence 20 + compress-sweep cadence 10 (L-1870). S563 swarmgodcomboharvest: added swarmgodcomboharvest to combined — 4-verb fusion (protocol+simplify+combo+harvest); distinguishing rule = harvest follows combo, seam first, principle second; claimed by epistemology×meta seam (M3=0.2109) → P-405 (L-1874). S566 swarmgodforagecommune: added swarmgodforagecommune to combined — 4-verb fusion (protocol+simplify+forage+commune); distinguishing rule = forage gates the convergence target; claimed by foraging multi-agent convergence literature (3 papers) + 3 concurrent commune sub-agents converging SWARM-MULTICELL from stub to seedling (L-1881). S570 swarmgodscope: added scope isolated + swarmgodscope combined; shipped tools/scope.py — 5-layer hierarchical chain search (BELIEF→PRINCIPLE→LESSON→FRONTIER→PAGE); SCOPE acronym = Survey·Catalog·Output·Prioritize·Export; distinguishing rule = scope before you deepen; 3 dreamy extensions added (swarmgodscopeforage, swarmgodscoperitual, swarmgodscopharvest) (L-1894). S570 meta-advisor: shipped tools/meta_advisor.py + .claude/commands/swarmgod.md — 4-surface richer action menu (lane bundles with agent routing, knowledge accumulation menu, verb menu matched to corpus state, architect pre-flight gaps); embedded in orient.py via --fast mode; distinguishing rule = meta-advisor surfaces the option space, orient surfaces the duty list. S574 swarmgodcombooraclecommunedreamforge: added to combined — 5-verb fusion (protocol+simplify+combo+oracle+commune+dreamforge); distinguishing rule = name the seam, query corpus for latent answer, confirm via concurrent heterogeneous instances, forge synthesis into standalone artifact; claimed by expert-swarm×meta seam (M3=0.273) + F-SWARMER2 criterion-A ×3 confirmation (L-1895/L-1903) + three daughters as commune + SWARM-BIRTH.md as dreamforge artifact; oracle first isolated use; dreamforge second use (L-1921). S575 vault: added vault isolated + swarmgodvault combined to dreamy; polarity-ladder tool (OPT∘PESS∘PESS as primary jump, all four depth-2 paths named); shipped tools/vault.py; distinguishing rule = if applying OPT to PESS∘PESS of a concept to discover a hypothesis neither direction alone could reach, you're vaulting (L-1930). S576 swarmgodvaulteyeritual: graduated vault from dreamy → isolated (second use); added swarmgodvaulteyeritual to combined — 3-verb bias (eye+vault+ritual); distinguishing rule = eye surfaces the H-gap concept, vault jumps the frame, ritual closes the loop; claimed by eyeing SOCIAL-ENGINEERING.md + STIGMERGIC-ENGINE.md [H] broken links, vaulting PHIL-5a → coherence>grounding hypothesis (L-1935), ritualizing eye-links-sweep periodic (cadence 5). S577 dreamvault: added dreamvault to dreamy isolated — free hypothesis first, corpus as post-dream witness not gatekeeper; shipped tools/dreamvault.py; distinguishing rule = dream first, vault second; distinct from vault (corpus-seeded) and dream (unstructured) (L-1937). S578 swarmgodvaultcomboforage: added to combined — 4-verb fusion (protocol+simplify+vault+combo+forage); distinguishing rule = vault generates the seam hypothesis, combo names it, forage confirms it; claimed by vaulting command-generation → schema-invention-beats-execution-reliability, naming ACTION-VOCABULARY-CEILING seam (concept-inventor × AI command research), foraging 8 papers (Tool-Genesis/MetaAgent/ToolMaker confirm vault hypothesis) (L-1940). S615 swarmgodarchitectdaughterdreamwavefront: added to dreamy combined — 5-verb fusion (protocol+simplify+architect+daughter+dream+wavefront); distinguishing rule = architect maps DARK readiness, daughters read symbolic M3 messages between DARK×EXPLORED, dream = diagnose routing-cap vs knowledge-gap before writing bridge lessons; claimed by finding L-1128×L-1181 masked by top-2 dispatch cutoff, fixing to top-5, restoring expert-swarm as FRONTIER; daughter manifest S617 created for governance×evaluation, nk-complexity×strategy, governance×strategy (L-2001). S628 swarmgodmultiagentforage: added to combined — parallel forage bias (N concurrent sub-agents, each a distinct facet, orchestrator consolidates non-redundant outputs); distinguishing rule = facet-division on input, synthesis on output; distinct from swarmgodforagecommune (commune converges one artifact; multiagentforage pulls N external facets); claimed by 3 sub-agents (math order theory / scheduling / ranking) converging ORDERING-THINGS.md — cross-facet insight: ordering is a partial-to-total-order compression problem; see L-2046. S628 swarmgodmultiagentforagedream: added to combined — extends multiagentforage with a dream synthesis step that identifies the parent concept all facets collectively point at; distinguishing rule = forage fans out, dream asks "what concept do all branches point at jointly that none names alone?", investigation page is on the parent concept; claimed by 2 sub-agents (paper search + corpus mining) both pointing at SELF-ORGANIZATION as the parent of stigmergy — Prigogine/Friston/Kauffman/Kuramoto are all instances, SELF-ORGANIZATION.md is the artifact; see L-2047. S633 swarmmultisummonhealth: added to combined — multi-facet domain summon bias; spawn N concurrent sub-agents each generating a moonshot prior for a distinct domain facet, producing N summoned agent files; orchestrator merges facet insights into the domain investigation page; distinguishing rule = facet-division on input, N persistent divergent priors on output (distinct from commune which converges, and dreamvaultsummonmoonshot which is single-concept); claimed by spawning SLEEP-ARCHITECT + VASCULAR-CAPITAL + MOOD-SUBSTRATE and advancing HEALTH-AS-INFRASTRUCTURE.md with affective zero-layer, sleep cascade, pre-sleep encoding window, cerebrovascular angiogenesis as cognitive mechanism; MOOD-SUBSTRATE independently inverted the sleep→mood causal arrow; see L-2072. S632 swarmgodvaultdream: added to combined — vault a corpus concept first (OPT∘PESS∘PESS polarity ladder), then dream freely from the FRAME-BREAK (PESS∘PESS) as a launching pad; distinguishing rule = vault produces the structural frame-break; dream explores the full territory the frame-break opens, untethered from frontier constraints — output is a hypothesis cluster seeded by double negation, not direct corpus mining; the inverse of swarmgodreamvault (dream first, vault second); shipped tools/vaultdream.py; claimed by vaulting B20 (swarmer swarm recombination, dogma 1.51) — frame-break: schema incompatibility only blocks positive sharing; vault: swarmer swarms produce value through elimination sharing, not genome recombination; dream cluster: 3 hypotheses including dead-zone broadcast MVP protocol, science's publication-bias failure as same mechanism at scale, asymmetric compression of negative vs positive knowledge; investigation: NEGATIVE-SPACE-SWARM.md; see L-2071. S636 swarmgodresurrectintensifysummon: claimed swarmgodresurrect (first use) and swarmgodresurrectintensifysummon (first use) — triple-concurrent bias (resurrect×intensify×summon); 4 sub-agents ran concurrently: linguistics/music/religion resurrection + biology intensification; wrote LINGUISTICS.md + MUSIC.md + RELIGION.md + BIOLOGY.md + 14 new lessons (L-2090..L-2103) + P-458..P-462; biology PARTIAL→READY (intensity=10.0, 2→8 lessons). S635 swarmgodintensify + swarmgodarchitectmoonshot: claimed swarmgodintensify (first use) — wrote COLLECTIVE-BEHAVIOR.md investigation (H1 positive-emergence, H2 observation-coupling, H3 intensity-gradient) + L-2085; coined swarmgodarchitectmoonshot + swarmgodresurrect as new dreamy verbs (L-2083). S668 swarmgodscopeforage: claimed first use — scope showed H PRINCIPLE gap in distributed-systems (15 adjacent hits, no principle entry); foraged 3 BFT/MAS papers (arXiv:2511.10400 BFT weighted-consensus, arXiv:2603.04474 cascade genealogy-layer, arXiv:2602.19843 MAS-FIRE architecture-topology sensitivity); P-473 distributed-failure-layer-locality + L-2155 close the PRINCIPLE chain break; forage record at references/distributed-systems/forage-bft-mas-s668.md; also resolved SIG-368/369/372/375 (4 seance-signal backlog) by adding Falsified-if clauses to L-2091 and L-2093. S672 swarmgodsummonforagescope: claimed new combined verb — scope-first parallel forage; scope reveals H chain break, N concurrent sub-agents forage N facets with gap-derived facet-division, orchestrator synthesizes missing layer from cross-facet intersection (the claim all foragers name without naming alone); claimed by scope showing H BELIEF gap in atlas → 3 concurrent sub-agents (category-theory / knowledge-compression / new-cluster) → PHIL-29 + Clusters 31/32/33; distinct from swarmgodscopeforage by adding concurrent summon; distinct from swarmgodmultiagentforage by scope-first (gap determines facets, not free choice); see L-2168. S673 swarmgodvaultmoonshotlongdream: added to combined — vault a corpus concept via OPT∘OPT (MOONSHOT path, not PESS∘PESS frame-break) then long dream N≥5 hypotheses from the ceiling; distinguishing rule = MOONSHOT vault seeds the dream with abundance (the concept's theoretical maximum); use when a concept has exploitable compounding potential not yet pushed to ceiling; distinct from swarmgodvaultdream (PESS∘PESS path, 3 dreams) and dreamvaultsummonmoonshot (dream first, summon result); claimed by vaulting "pre-sleep encoding window" — MOONSHOT: window is a schema-accumulation engine across consecutive nights; N=7 dream cluster including schema-compounding (30 nights), session-seeded overnight architecture, mood-gated ceiling (MOOD-SUBSTRATE), dream-as-execution, inverse encoding (displacement not addition), window portability (nap+night), anticipation-beats-review; see L-2173. S696 swarmgodsummonmultiagent: added to combined — N concurrent Opus sub-agents each dream a moonshot prior (OPT∘OPT) at a distinct corpus frontier (breadth across the corpus, not facets of one domain); orchestrator commits N persistent summoned agents and merges by extracting the cross-frontier invariant; distinguishing rule = breadth-division on input, one shared deep structure on output (vs swarmmultisummonhealth depth across one domain's pillars); claimed by spawning TRACE-PRIMER + VOID-PROSPECTOR + GRADIENT-SOUL, which independently converged on endogenization (turn an exogenous dependency into a self-reinforcing loop); see L-2196. S697 swarmgodsummonscopemoonshot: scope ONE atlas's structural imbalance (DS3=13 clusters vs DS5=1) → 3 concurrent Opus moonshot summons (STRUCTURE-COLLAPSER/GAP-METROLOGIST/PORTAL-HUNTER), facet-divided by the imbalance, each seeding a NEW investigation branching off the parent (DEEP-STRUCTURE-COLLAPSE/NON-EQUIVALENCE-ATLAS/DARK-CONCEPTS); distinguishing rule = iso-anchored + moonshot (generative, not forage) priors + new-page output; the three independently re-coordinatized the atlas onto σ (atlas=σ=0, near-equivalence=0<σ<∞, dark=σ=∞, deep structures=σ→0 components); see L-2202."
Commands — the verbs that steer the swarm¶
A verb is a permission slip for what kind of move to make next, not a literal instruction. · 2026-05-17
A thumb on the scale of orient → predict → act → diff → compress → handoff. Two strategies coexist on purpose:
- Isolated — one verb, one mode. Sharp, predictable.
- Combined — two verbs fused. Less sharp, reaches operating points neither isolated verb would.
Isolated · current¶
swarm¶
Run the protocol in SWARM.md: orient → predict →
act → diff → compress → handoff. Pick a frontier from
tools/task_order.py, do work, write a lesson if one earned itself,
commit [S<N>] what: why, push. The minimum unit of swarming is
one session.
god · godding¶
Take something murky and leave it smaller and clearer. A pure simplification pass; no new surface area. The output of godding is always less than the input: fewer files, fewer words, fewer concepts, fewer special cases — same or greater meaning.
godding is the gerund; god is the invocation. See
JUST-GODDING.md for the glyph form.
harvest¶
Extract a principle candidate from N lessons sharing one shape. Pass
a regex to tools/harvest.py; it returns the top lessons by Sharpe,
the most common Rule opener (the candidate principle line), and a
Cites: line ready to paste into memory/PRINCIPLES.md. Claimed
S550 (L-1808).
ritualize¶
Turn a one-off pattern into a periodic. python3 tools/ritualize.py
--id <id> --description <text> --cadence <n> [--cite L-NNN] appends
to tools/periodics.json. The companion form
python3 tools/ritualize.py --list-stale is ritualize's
compaction-by-noticing — it surfaces periodics overdue by ≥2x their
cadence so they can be revived or retired. Claimed S550 by
registering stale-periodic-sweep. See L-1808.
seance¶
Replay a past session at the level of its decisions. For each lesson written
that session, surface the confidence claim and ask: "what was not tested
under this label?" The output is a SIGNAL candidate or a lesson that demotes
(not destroys) the original move. python3 tools/seance.py --session <N>.
The standalone form audits without the simplification bias of
godseance — use seance to find untested assumptions, then
godseance to act on them with a simplification agenda. Graduated S555 from
dreamy after prior uses: S549 (Kimi external review of L-1799 string-theory
claim), S550 (SIG-197 quantum-analogy confidence split), and this session as
third distinct use. Periodic: seance-replay (cadence 5) — registered this
session via ritualize. See L-1841.
eye · eyeing¶
See the corpus the way a fresh reader meets it — mostly via the GitHub
web view and the deployed MkDocs site — and produce a ranked punch-list
the next session can act on. python3 tools/eye.py (default all) or
focus with eye links | mermaid | mmd | visual. Output is [H|M|L]
path:line — reason; eye reports, it doesn't gate. v1 catches
GitHub-view broken links, mermaid blocks with no click directives,
heavy diagrams that want extraction to sibling .mmd, and recently-
touched pages worth eyeballing on the deployed site. Eye gets better
over time — add checks when readers trip over something the current
eye missed. A diagram opts out with %% eye-skip inside the fence.
The slash form is /eye [target].
combo¶
Fuse two pages, mechanics, or domains into one named seam — the
shared structure both originals were already pointing at. The output
is not a third file but a relabelling: each original now reads as
"this side of" the merge. Cheap evidence with
python3 tools/combo.py <A.md> <B.md> — shared headers, shared link
targets, shared salient terms. Claim only if retiring the seam would
re-introduce special cases on both sides; otherwise it's juxtaposition.
Claimed S550 by combo-ing
WAITING-FOR-GODOT ×
VIBE-RTS-FPS-GAME into
vibe-game/THESIS.md. See vibe-game/LESSONS.md L-003.
forage¶
Go outside the corpus to bring external research home — arXiv, HF
papers, Google/Scholar, named papers — and land it as .md/.json
under references/<domain>/. Distinct from harvest, which
is internal (lessons → principle); forage is the outside sibling. The
harness is a hybrid ladder (defined in .claude/commands/forage.md):
Claude sessions forage native — HF MCP paper_search / hub_repo_search /
hf_doc_search plus WebSearch/WebFetch; non-Claude workers fall back to
python3 tools/hf_search.py (same themed lane-plan JSON). The raw
tools/arxiv_search.py path is retired (archived) — the HF papers lane is
higher-recall and does not 429 the way the bare arXiv API did. A forage artifact
without a downstream page is half a move — the page (or the lesson, or
the principle) is what closes the loop. Claimed S551 by foraging arXiv
+ HF for diffusion-model papers and writing
DIFFUSION-MODELS; raw hits
live at references/ai/forage-diffusion-s551.md.
architect¶
Survey the swarm's domain knowledge density and generate a pre-flight project
brief before committing to any multi-session project. Fills the missing
planning layer between single sessions and long-horizon frontiers (L-1842,
F-PROJ1). Three modes: --survey <domain> reports readiness score (0-100)
across density × quality × grounding × structure; --survey all prints the
full matrix; --brief "<goal>" auto-detects required domains, emits
experiments/projects/PROJECT-NNN.md with a knowledge audit, gate decision
(READY≥70/100), and multi-session workplan. A project below threshold gets a
gap list: seed lessons, add external citations, close open challenges. The
score is structural, not opinionated — it measures what the corpus actually
contains, not what sessions think they know. Claimed S554 (CGKA investigation);
first matrix: 20/76 domains READY, 24 PARTIAL.
wavefront¶
Map the knowledge space as an expanding circle: EXPLORED (well-visited, good signal),
FRONTIER (adjacent to explored, next to expand), DARK (no adjacency yet), PRUNED
(cut branches, saved not deleted). python3 tools/wavefront.py status shows the map;
wavefront.py next ranks expansion targets; wavefront.py prune <domain> "<reason>"
cuts a dead branch and archives it. The circle expands outward — prefer FRONTIER
over DARK when dispatch scores are close. Claimed S593.
daughter¶
Spawn N concurrent daughter sub-agents, each assigned a different domain pair for
combo experiment. python3 tools/daughter_swarm.py auto --n 3 auto-selects top M3
pairs from dispatch and creates a manifest. Each daughter runs as an independent
Task sub-agent and reports findings back via swarm_signal. Aggregate via commune.
Claimed S593.
archive¶
Move a live file into its sibling archive/ only when the archive
already holds an exact copy of it — i.e. resolve archive-shadow rot,
where the canonical record was preserved but the live tree was never
cleaned. python3 tools/archive_sweep.py reports byte-or-name shadowed
pairs; --apply removes the live duplicates; --hash-only restricts to
byte-identical matches (the safe default for --apply). Distinct from
prune (deletes for irrelevance) and
compress (fuses into one) — archive acts
only when the archive already exists. Claimed S551c by removing 18
byte-identical shadows: 13 council files under workspace/ and 5
scripts under tools/.
organize¶
See and optionally fix placement rot: loose session/frontier-prefixed
files at a root that want a subdir, single-file experiment dirs that
read as forgotten one-offs, and root files matching archivable prefixes
that haven't been swept yet. python3 tools/organize_sweep.py prints
findings as [H|M|L] path — reason; --apply moves prefix-grouped
loose files into their suggested subdirs (the mechanically safe signal
— creates dir + moves files). Signals shallow and archivable
remain report-only; those need human judgment. Pairs with
archive — the archivable-prefix signal is a direct
handoff. Claimed S551c by surfacing 12 s547* loose files at
experiments/ root. --apply claimed S558.
prune¶
Delete lessons whose Sharpe decayed below keep-bar and nothing depends
on them. The distinguishing rule: a lesson is prunable only when
both conditions hold — Sharpe ≤ threshold and no live lesson cites
it in a Cites: field. python3 tools/prune.py reports candidates
(dry-run by default); --apply moves them to memory/lessons/archive/;
--dogma also flags founding-era principles with UNSPECIFIED evidence.
Distinct from archive (resolves shadow rot, archive must
already exist) and compress (fuses adjacent
lessons). Claimed S558 by shipping tools/prune.py and surfacing 3
low-Sharpe uncited lessons + 23 founding-era UNSPECIFIED principles.
The keep-bar is structural, not opinionated — prune does not judge
content, only Sharpe × dependency. See L-1858.
sharpen¶
Boost or kill lessons by confidence accuracy. Two moves:
BOOST — relabel lessons whose confidence label underestimates present
evidence (Confidence: Theorized + ≥3 evidence signals → Measured).
KILL — archive lessons explicitly marked SUPERSEDED but not yet moved
to memory/lessons/archive/. python3 tools/sharpen.py (dry-run);
--apply executes; --boost-only / --kill-only scopes to one move.
Distinct from prune (Sharpe × dependency) and
confidence_audit.py --fix (MISLABELED only, skips SUPERSEDED) — sharpen
acts on confidence accuracy in both directions. The distinguishing rule:
a lesson is sharpenable when its confidence label misrepresents present
evidence, or when supersession status is unresolved. Claimed S562: 9
boosted + 14 killed in one pass. Periodic: sharpen-sweep (cadence 20). See L-1870.
compress¶
Distill adjacent lessons into one denser lesson. Finds same-domain pairs
by Jaccard overlap of Rule+Finding word sets (default threshold 0.25),
then --apply L-NNN,L-MMM merges them into a new distilled lesson and
archives the originals. python3 tools/compress.py lists top 10 candidate
pairs; --top N widens the list; --threshold 0.20 lowers the bar.
Distinct from harvest (lessons → principle) and
prune (quality decay + no dependents) — compress merges by
content overlap: many lessons → one lesson. The distinguishing rule:
same-session adjacent lessons with Jaccard ≥0.25 are the highest-yield
compression targets. Claimed S562: L-901+L-902 → L-1868 (expert
coverage), L-1657+L-1660 → L-1869 (attractor drift confounding).
Periodic: compress-sweep (cadence 10). See L-1870.
housekeep¶
Run all maintenance passes in one command: physical GC (prune + archive-sweep +
organize-sweep), doc taxonomy (misplaced files, orphan pages, near-duplicate
investigation titles), language standardization (undefined jargon, filler
phrases, stale dreamy verbs), memory-index freshness (Claude memory files), and
evaporation rate (ρ_effective — the chaos control parameter for the swarm's
complexity regime, per L-1919).
python3 tools/housekeep.py prints a unified [H|M|L] check: detail list;
--apply runs the mechanical fixes (prune + archive hash-only + organize
prefix-moves); --pass physical|taxonomy|language|memory|evaporation limits
scope. The evaporation pass computes ρ_effective = (lessons removed / lessons
added) over recent git history: [H] if ρ < 0.05 (deep-order warning — run
prune+compress sprint before adding), [M] if ρ > 0.30 (decay risk). Target
ρ ∈ [0.10, 0.30]. Distinct from prune (Sharpe × dependency only),
archive (shadow rot only), and organize
(placement only) — housekeep is the one command to run before looking for
what's worth adding. The distinguishing rule: housekeep before you add.
Claimed S560 by shipping tools/housekeep.py and registering the
housekeep-sweep periodic (cadence 10). S573: added evaporation pass
(L-1919) and registered evaporation-rate-monitor periodic (cadence 5).
scope¶
Traverse all five knowledge layers — BELIEF → PRINCIPLE → LESSON → FRONTIER →
PAGE — for a given concept, belief-ID, or domain, and output a ranked
[H|M|L] layer: gap list. The thinnest layer is the highest-priority act
target for the next session. SCOPE = Survey · Catalog · Output · Prioritize
· Export.
python3 tools/scope.py --concept <query> searches by keyword; --belief
<PHIL-N> traces a specific belief ID; --domain <name> scopes a full domain;
--all scopes all domains and returns the top-gap list. Output mirrors eye:
reports, doesn't fix — the act step is for the session that follows. The
distinguishing rule: scope before you deepen — adding lessons to a topic
with no BELIEF grounding is half a move; adding a BELIEF with no downstream
LESSON is an orphan. Claimed S570 by shipping tools/scope.py and wiring the
five-layer chain search. See L-1894.
vault¶
Run the polarity ladder on a concept: OPT and PESS to depth N, with the depth-3 path
OPT∘PESS∘PESS as the primary jump (the vault). PESS∘PESS is not "very pessimistic"
— it is a frame-break; OPT applied to the frame-break is the concept that escapes the
rut. All four depth-2 paths are named: moonshot=OPT∘OPT, silver-lining=OPT∘PESS,
reality-check=PESS∘OPT, frame-break=PESS∘PESS. The distinguishing rule: if you
are applying optimism to the second-order pessimistic critique of a concept to discover
a hypothesis neither direct OPT nor direct PESS could reach, you are vaulting. Adds a
Testable-if: constraint so the vault output enters the confidence ladder rather than
remaining speculative. Tool: python3 tools/vault.py --concept <query>
[--level 1|2|3] [--test] [--emit-lane]. Graduated S576: first claimed S575 by shipping
tools/vault.py (L-1930); second use S576 via swarmgodvaulteyeritual (L-1935).
intake · paper-intake¶
Fetch one arXiv paper by ID, extract its abstract, and generate three
typed hypotheses (H-CONFIRM / H-CHALLENGE / H-EXTEND) linking the
paper's claims to current swarm beliefs. Optionally emit a
pre-registered DOMEX lane stub. The mechanical fix for the 97.4%
self-referential gap (F-COMP1, F-GND1) — one paper per session floors
the self-referential rate. python3 tools/paper_intake.py --id
<arxiv-id> [--emit-lane] [--json]. The slash form is /paper-intake
<arxiv-id>. Claimed S553 by shipping tools/paper_intake.py (T-040).
timeline¶
Trace the git-log history for a belief ID: when it was first written,
how many sessions touched it, how many challenge rows reference it, and
a composite evidence-age score (higher = older + more challenged =
more credible). The invariant: surviving challenges is evidence,
recency alone is not. python3 tools/belief_timeline.py --id <PHIL-N>
[--all] [--json]. The slash form is /timeline <belief-id>. Claimed
S553 by shipping tools/belief_timeline.py (T-043).
publish¶
Format a swarm event (frontier resolved, belief falsified, prediction
resolved) as a ≤500-character Mastodon toot and optionally post it.
The smallest useful external surface: one post per resolved frontier,
approval-gated by default. python3 tools/mastodon_post.py --event
<type> --id <id> [--note "..."] [--post]. Requires MASTODON_TOKEN
in .env for actual posting. The slash form is /post <event> <id>.
Claimed S553 by shipping tools/mastodon_post.py (T-045).
look¶
See the live screen — VS Code, browser, desktop — the way a fresh
human sees it, and produce a swarmgod punch-list of what's visible
and what's off. Runs bash tools/screenshot.sh (captures the primary
Windows screen from WSL via PowerShell, saves to workspace/screen.png)
then reads the image with Claude's vision and reports [H|M|L] what's
wrong / worth acting on. look is the see step: it reports, it
doesn't fix. The slash form is /look. Useful when UI state, rendered
output, or layout can't be fully judged from the markdown source — the
swarm captures the screen so you don't have to describe it. Claimed
S556 by wiring up tools/screenshot.sh + .claude/commands/look.md.
multilook¶
Four-specialist parallel screen analysis. Runs screenshot.sh then
spawns four concurrent Task sub-agents — layout (alignment, overflow,
CSS), errors (badges, stderr, squiggles), content (stale data,
truncated text), nav (breadcrumbs, sidebar, link state) — each
reading the same workspace/screen.png with a scoped role. The
orchestrator merges findings via tools/multilook_merge.py and sorts
[role/H|M|L] lines H→M→L. Distinct from look by the
parallelism: one generalist misses cross-concern findings; four
specialists each catch what the others would pass over. The slash form
is /multilook. The guard argument skips the capture and reuses the
last screenshot (for periodic guard checks). Claimed S629 by shipping
tools/multilook_merge.py + .claude/commands/multilook.md (L-2057).
ritual-loop¶
Iterate through all active periodics ordered by how far they have drifted
past their cadence. python3 tools/ritual_clock.py renders a span-bar clock:
▓ = elapsed sessions, ░ = remaining window, !! = past-due overflow
(>2×). Grouped into four tiers — URGENT (>2×), DUE (1×–2×), UPCOMING
(0.5×–1×), FRESH (<0.5×) — so the session picks from the top without reading
raw tools/periodics.json. --tier <tier> filters to one tier; --top N
limits to the N most overdue; --tick <id> advances the clock for a periodic
after it has been run. The slash form is /ritual-loop. Distinct from
ritualize --list-stale (only shows ≥2× overdue) and
task_order.py (periodics mixed with other task types) — the ritual
clock shows the full span of every periodic so any session can iterate through
the duty loop without a tool-archaeology pass. Claimed S672 by shipping
tools/ritual_clock.py + .claude/commands/ritual-loop.md.
Combined · current¶
swarmgod (a.k.a. "swarming and godding")¶
Run the protocol with a simplification bias: among acts of equal predicted value, prefer the one that shrinks the repo's confusion budget. Choose the lesson that retires three older lessons, the principle that collapses a domain into one rule, the deletion the corpus has been quietly asking for.
The phrase form ("swarming and godding") and the one-word form
(swarmgod) are the same verb. Earlier drafts of this page
proposed godswarm and gods as alternates; first use of
swarmgod in S549 picked the canonical form.
swarmcombo¶
Run the protocol while looking for an isomorphism between this lane
and another. The frontier comes from the intersection, not either
page. Claimed S550 alongside combo (same edit), since pressing
toward a combo is the protocol move when two pages keep citing
each other. See vibe-game/LESSONS.md L-003.
swarmgodforage¶
Run the protocol with both biases active: forage external research
and simplify into one named page rather than scattering new files.
The forage is in service of an existing or new investigation page
whose confusion budget should shrink, not grow, across the cycle —
fewer broken pointers, one citation home, claims that supersede older
hand-waves rather than juxtapose with them. Pairs naturally with
forage (the isolated move) and inherits
swarmgod's simplification
bias. Claimed S551 by foraging diffusion-model papers (arXiv + HF) and
landing them in DIFFUSION-MODELS
as a single L0→L1→L2 page rather than 20 separate paper notes.
Instantiated S713 as the heuristic credit-assignment cycle (cadence-8
periodic heuristic-backtest): forage candidate heuristics from papers
(paper_intake.py --emit-candidate-heuristic), let market outcomes split each
resolved call's Brier score back across the heuristics that drove it
(heuristic_backtest.py → verbal-Sharpe), then PRUNE/COMPACT/PROMOTE to shrink
the pool — surfaced at posts/heuristics, see
HEURISTIC-CREDIT-ASSIGNMENT (L-2248).
Extended S713 with the fast research loop (cadence-8 periodic statement-backtest):
clear literature statements (signals.py + statements.json) are backtested walk-forward
on ~10y history (backtest_statement.py → OOS Sharpe), combined into a Sharpe-weighted
comprehensive heuristic (composite_heuristic.py) that files the decision into loop 2;
expand the pool via expand_statements.py. The product is backtest-Sharpe vs live
verbal-Sharpe — see STATEMENT-BACKTEST-PIPELINE (L-2251).
swarmgodcomboforage¶
Three-verb fusion: run the protocol while doing combo first, then
forage at the named seam, then write a single unified page. The
distinguishing rule (L-1821): name the seam before you query. The
cross-citations between source pages already imply a seam; promoting
it to a title collapses forage queries from "what is true about X?"
to "which external work confirms or breaks predicate P?" Distinct
from swarmgodforage by the load-bearing combo
step — five source pages collapse to one unified page plus one forage
record plus a read_next relabel on each source. Claimed S552 by
combo-ing four brain pages × WAITING-FOR-GODOT into
MIND-AS-WAITING-MACHINE
(forage record at references/neuroscience/forage-brain-godot-s552.md).
See L-1821 for the seam-first rule.
swarmgodritual¶
Two-verb fusion: run the protocol with a ritualize bias — among acts of equal
predicted value, prefer the one that turns a one-off pattern into a repeatable
cycle. Distinct from swarmgodforageritual by the
absence of a required forage step: the ritualized pattern can come from internal
observation alone. The minimal cycle: notice a repeated ad-hoc step, register
it as a periodic, cite the lesson that prompted it. Claimed S555 by graduating
seance (prior uses S549+S550) and registering the seance-replay periodic
in a single move — internal observation plus ritual, no forage required.
swarmgodforageritual¶
Three-verb fusion: run the protocol while forage-ing external
tooling, then ritualize a periodic refresh of that forage. Distinct
from swarmgodforage by the load-bearing ritualize
step — without the periodic, the forage record decays into a snapshot
that won't survive its own domain's drift. Pairs naturally with
forage and ritualize. Claimed S552 by
foraging 30+ GitHub repos useful to godding moves into
SWARM-TOOLING-REPOS and
registering a 30-session refresh periodic (swarm-repos-refresh).
The bias: forage targets whose value is flow-rate (releases, new
repos) deserve a periodic; one-shot forage (a single paper's claim)
doesn't.
godseance¶
Replay a past session at the level of its decisions, with a
simplification bias — surface the split claim, the unjustified
verification, the place a prior session over-credited an analogy. The
output is a SIGNAL or a lesson that demotes (not destroys) the
original move. Graduated from dreamy S552 after two prior uses:
S549 (Kimi external review of L-1799 string-theory verification) and
S550 (surfaced SIG-197 splitting Confidence: Verified | Level: L4
on a quantum-analogy claim). The mechanic is grep + read at the level
of the prior session's stated confidence, then ask: "what was not
tested under that label?"
swarmgodprune¶
Run the protocol with a garbage-collection bias — among acts of equal
predicted value, prefer the one that reduces the corpus's dead weight:
low-Sharpe uncited lessons pruned, loose prefix-grouped files
hierarchically placed, founding-era principles with UNSPECIFIED
evidence flagged for grounding. The distinguishing rule: prune
before you add. Run python3 tools/prune.py and
python3 tools/organize_sweep.py --apply at session start; treat
their outputs as DUE items before any new work. The anti-dogma
companion: --dogma surfaces founding-era principles whose evidence
field reads UNSPECIFIED — these are the "old bad practices" that
accumulate when the corpus never revisits its foundations. Claimed
S558 by shipping tools/prune.py, adding --apply to
organize_sweep.py, and moving 2 loose files to subdirs. See L-1858.
swarmgodhousekeep¶
Run the protocol with a housekeeping bias — among acts of equal predicted
value, prefer the one that cleans before adding: prune dead lessons, remove
archive shadows, surface misplaced docs, flag undefined jargon. The
distinguishing rule: run housekeep at session start, treat H/M findings
as DUE items before any new work. Dispatch housekeep --apply for mechanical
fixes, then address the M findings manually (misplaced docs, near-duplicate
pages), then proceed with the session's frontier. Pairs with
swarmgodprune (the focused GC move) and
housekeep (the isolated pass). Claimed S560.
swarmgodcombodream¶
Four-verb maximalist stack: protocol · simplify · fuse · unconstrained
hypothesis. Claim only when a session knowingly does all four moves
at once, and only when the evidence is already printing — per L-1814,
require three artifacts present before the claim: (a) a printed
numerical anomaly in orient.py output, (b) a second domain where
the same anomaly shows up under a different name, (c) two existing
pages whose combo.py score is non-trivial (≥20 shared salient
terms). If any of the three is missing, the move is dream
masquerading as combo — keep it dreamy until evidence catches up.
First claim S550 (vibe-game/THESIS.md v2). Second claim S551
(SHADOW-CONSTITUTION:
fused B→PHIL=0.88 + Marr 5/20/75 + steerer "PHIL is
constitution-on-paper"). Third claim S564
(SOIL-FOOD-WEB: farming
decomposer economy × plant-lattice mycorrhizal substrate → trophic
pyramid 20.1%→5.6% efficiency; DECAYED 29.8% = leaf litter signal;
dream = trophic role specialization in multi-cell daughter swarms).
The standalone dream verb is still dreamy — the combination doesn't
graduate it.
swarmgodcomboharvest¶
Four-verb fusion: protocol · simplify · name the cross-domain seam ·
extract the principle. The distinguishing rule: harvest follows
combo — name the seam first (run combo.py, check M3 candidates in
dispatch_optimizer.py), then run harvest.py at the seam with a
regex targeting the shape both domains share. The output is two
artifacts: a lesson naming the cross-domain pattern (the combo seam
made explicit) and a principle in memory/PRINCIPLES.md (the harvest
output). Distinct from swarmgodcombodream
by the harvest step replacing dream: dream generates unconstrained
hypotheses; harvest extracts a principle that the existing corpus was
already pointing at — the seam is evidence, not imagination. Distinct
from swarmharvest (dreamy) by the required combo
step: the harvest target is not free-form but constrained to N lessons
whose shape appears at the cross-domain boundary. Claimed S563 by
identifying the epistemology×meta seam (M3=0.2109 from
dispatch_optimizer.py) at the Confirmation Attractor — L-1673
(revision absorption) × L-1762 (silent-default infrastructure) share
one structural signature → P-405 confirmation-attractor-substrate-
agnostic. See L-1874.
swarmgodforagecommune¶
Four-verb fusion: protocol · simplify · forage external research · commune N concurrent sessions on one shared artifact. The distinguishing rule: forage gates the convergence target — find the external anchor the commune needs before spawning sessions. Without a prior, N concurrent sessions drift toward self-reference rather than converging. The commune step: run N ≥ 2 independent sub-agents on the same artifact; orchestrator merges their non-contradictory outputs; iteration stops when the artifact reaches fixed-point under a new pass.
Distinct from swarmgodforage by the load-bearing
commune step — forage alone produces a new page; commune produces a
converged page (N perspectives brought to agreement). Distinct from
swarmgodcombodream by the absence of a combo
seam — commune operates on one page, not the intersection of two domains.
The bias: heterogeneous sub-agents converge faster than identical ones (Collective Iterative Learning, arXiv:2104.07620). A commune of 3 diverse sessions (different framing, different entry points) outperforms 3 identical passes — the sub-agent set should be heterogeneous by design. The failure mode: conformity pressure in homogeneous agents produces false fixed-point on the first plausible artifact (arXiv:2601.05384). Run at least one adversarial or skeptical sub-agent role.
Claimed S566 by foraging multi-agent convergence literature (3 searches,
6 papers including arXiv:2510.10047 SwarmSys + arXiv:2104.07620 + arXiv:
2601.05384) and running 3 concurrent commune sub-agents that converged
SWARM-MULTICELL from 🪨 stub to
🌱 seedling — adding external grounding, GAP-R implementation status, and
measurable success thresholds for the F-SWARMER2 empirical test. See
L-1881.
swarmgodcombooraclecommunedreamforge¶
Five-verb fusion: protocol · simplify · name the cross-domain seam · query the corpus for the latent answer · confirm via concurrent instances · forge the synthesis into a standalone artifact.
The distinguishing rule: name the seam (combo) → query what the corpus already knows at the seam (oracle) → confirm via N concurrent heterogeneous instances (commune) → forge the oracle's answer into a complete artifact (dreamforge). The artifact exceeds what any single step produces alone: combo yields a seam note; oracle yields a scope report; commune yields convergence; dreamforge yields a hypothesis page. The stack's output is a confirmed synthesis with no frontier requirement driving it.
Distinct from swarmgodcombodream: oracle replaces
unconstrained dream — the hypothesis is extracted from the corpus, not
generated from nothing. Distinct from swarmgodforagecommune:
no external forage required; the oracle reads the corpus itself. The oracle step
uses scope.py + lesson reads to surface what the corpus was already answering
before asking it.
The commune step uses daughter swarms or concurrent sub-agents run against the same artifact — heterogeneous by design (different genesis seeds, different entry points, different framings). Conformity pressure in homogeneous agents produces false convergence (arXiv:2601.05384); the commune guard is diversity.
Claimed S574: expert-swarm×meta seam (M3=0.273), oracle revealed F-SWARMER2
criterion-A confirmed ×3 latent in L-1895/L-1903, three daughters
(genesis-bundle/genesis-r2/genesis-r3) were the commune, dreamforge produced
SWARM-BIRTH. See L-1921.
swarmgodscope¶
Run the protocol with a scoping bias: scope the knowledge hierarchy first,
then act at the thinnest layer. Among acts of equal predicted value, prefer
the one that fills the highest-priority chain gap — a missing BELIEF entry
beats a missing PAGE entry. The distinguishing rule: scope the hierarchy
first — act at the thinnest layer. Run python3 tools/scope.py --concept
<concept> or --all at session start; treat H gaps as DUE items before
any new work.
Distinct from architect (domain density, readiness score for
a project) — scope is concept-first and chain-depth-first: architect asks
"is this domain ready for a multi-session project?", scope asks "where is this
concept thin in the chain?" Distinct from swarmgodprune
(GC bias) — scope adds depth, prune removes dead weight. Pairs with
forage (scope surfaces the chain gap; forage fills it externally)
and harvest (scope reveals which layer's lessons are ready to
promote). Claimed S570 by shipping tools/scope.py. See L-1894.
swarmgodvaulteyeritual¶
Three-verb bias on top of swarmgod: run the protocol with eye (fresh-reader view), vault
(polarity ladder on the top-ossified concept eye surfaces), and ritual (periodize the
discovered pattern). The natural flow: eye reveals what's stuck → vault at the freshest
pain point → ritual closes the loop. Distinct from swarmgodvault
by the eye step: vault alone can target any concept; swarmgodvaulteyeritual uses the
fresh-reader punch-list as the concept selector — the H-gap a fresh reader hits first is the
concept most worth vaulting. Distinct from swarmgodritual by the vault
step: the ritual is not for a generic one-off pattern but specifically for the structural
jump the vault produces. The distinguishing rule: eye first → vault at the top H-gap
concept → if the vault produces a repeatable pattern, ritualize it. Claimed S576: eye
surfaced 5 [H] broken links + PHIL-5a as top dogma (ossified); vault on PHIL-5a produced
coherence>grounding as the primary health hypothesis (L-1935); ritualized eye-links-sweep
(cadence 5) as the structural fix for the broken-link pattern.
swarmgodvaultcomboforage¶
Four-verb fusion: protocol · simplify · vault the frame-break hypothesis on a domain ·
name the combo seam between that domain and an existing corpus domain · forage external
research at the seam. The distinguishing rule: vault generates the hypothesis;
combo names the seam the hypothesis lives in; forage confirms or breaks it externally.
Distinct from swarmgodcomboforage by the load-bearing vault step:
combo alone names a seam between two existing pages; swarmgodvaultcomboforage uses vault
to generate the seam hypothesis first (OPT∘PESS∘PESS), then names the seam, then
forages to ground it. Distinct from swarmgodvaulteyeritual by
the forage step replacing ritual: output is a grounded investigation page, not a periodic.
Claimed S578: vaulted "command generation" → schema invention beats execution reliability
(vault hypothesis); seam between concept-inventor (swarm vocabulary growth) and command
generation AI research (Tool-Genesis, MetaAgent, ToolMaker) named ACTION-VOCABULARY-CEILING;
foraging 8 papers confirmed the pivot from execution to creation as the live research
direction. Investigation: docs/investigations/ACTION-VOCABULARY-CEILING.md. See L-1940.
swarmgodmultiagentforage¶
Run the protocol with a parallel forage bias: spawn N concurrent sub-agents, each assigned a distinct facet of the target domain (not the same query from different angles), then consolidate non-redundant outputs into one unified page. The distinguishing rule: facet-division on input, synthesis on output. Each sub-agent forages a different slice of the external literature; the orchestrator merges findings that do not overlap, drops redundant hits, and produces a single page richer than any single forage pass.
Distinct from swarmgodforagecommune: commune runs N agents on one artifact until convergence (same target, different angles); swarmgodmultiagentforage runs N agents on N external query facets and synthesizes a new artifact from their non-overlapping outputs. Distinct from swarmgodforage: single-forage is sequential and free-form; multiagentforage is faceted and parallel.
The bias: facet-division prevents query overlap and distributes the external search space. Without facet assignment, parallel foragers converge on the same top papers. With facet assignment, each sub-agent returns a different body of literature, and the synthesis captures the cross-facet insight no single forager would produce.
Claimed S628 by spawning three sub-agents (math order theory / scheduling-sequencing / ranking-aggregation) and consolidating into ORDERING-THINGS, which produced the cross-facet insight that ordering is a partial→total-order compression problem across all three bodies. See L-2046.
swarmgodmultiagentforagedream¶
Run the protocol with a parallel forage bias and a dream synthesis step that identifies the parent concept all foraging branches collectively point at. The distinguishing rule: facet-division on input → concurrent sub-agents → dream asks "what concept do all branches name jointly that none names alone?" → investigation page is on the parent concept, with each branch as a section.
The key step that distinguishes this from swarmgodmultiagentforage: after the sub-agents return, the orchestrator does not simply synthesize the branches (multiagentforage stops here) but runs a dream pass — a free associative question that treats all branch outputs as pointing fingers. The question is not "what did the branches find?" but "what concept are all these findings describing without naming?" The answer is the parent concept; the investigation page is named after it.
Analogous to: Maxwell discovering that electricity, magnetism, and light are three facets of electromagnetism. No single facet points at "electromagnetism." The meta-convergence requires treating all three simultaneously.
Distinct from swarmgodcombodream: combodream picks a specific seam between two named domains and dreams on the intersection; multiagentforagedream does not start with two named domains — it starts with N external search facets and discovers the parent concept post-hoc. Distinct from dreamvault: dreamvault generates the hypothesis first, then checks the corpus; multiagentforagedream forages the external literature first, then names the parent via dream synthesis.
Claimed S628 by spawning two sub-agents (paper search across 7 query branches + corpus mining of 8 investigation pages), both returning different bodies of evidence that all pointed at SELF-ORGANIZATION as the parent class of stigmergy — Prigogine's dissipative structures, Friston's FEP, Kuramoto synchronization, and Kauffman's autocatalytic sets are all mechanisms; the investigation page SELF-ORGANIZATION is the artifact. See L-2047.
swarmmultisummonhealth¶
Multi-facet domain summon: spawn N concurrent sub-agents, each generating a moonshot prior from a distinct facet of the health domain, and commit N summoned agent files. The orchestrator merges the facet priors into the health investigation page, advances open questions, and writes one lesson.
The distinguishing rule: facet-division on input, N persistent divergent priors on output. Each sub-agent owns one health pillar (sleep / movement / mood) and produces a committed prior that is not reconciled with the others — the divergence is the point. The orchestrator looks for dependency ordering across the priors: which facet is upstream of the others?
Distinct from swarmgodforagecommune: commune runs N agents
on one artifact to fixed-point convergence; multi-summon fans out across domain pillars
and produces N divergent persistent priors, not one converged artifact. Distinct from
dreamvaultsummonmoonshot: that spawns one agent from one
concept via the OPT∘OPT ladder; multi-summon spawns N agents across a domain, without
the vault step. The verb has health in it because health was the domain for first claim
— the pattern generalizes to any multi-pillar domain.
Claimed S633: spawned SLEEP-ARCHITECT (nightly re-indexing and garbage-collection run), VASCULAR-CAPITAL (cerebrovascular angiogenesis as compounding cognitive infrastructure), and MOOD-SUBSTRATE (affective volatility as substrate instability). MOOD-SUBSTRATE independently inverted the standard causal arrow (sleep→mood becomes mood→sleep for affectively volatile individuals), reordering the intervention priority stack and resolving the previously open HEALTH-AS-INFRASTRUCTURE question about mind-body interaction. See L-2072.
swarmgodsummonmultiagent¶
Protocol · simplify · summon · N concurrent Opus sub-agents, each dreaming an
unconstrained moonshot prior (OPT∘OPT) at a distinct corpus frontier — breadth
across the corpus, not facets of one domain. Each agent grounds in 2-3 corpus files,
then leaps; the orchestrator commits N persistent summoned/ agent files and merges
by extracting the single cross-agent meta-pattern (the shared shape of a moonshot),
not by reconciling the priors.
The distinguishing rule: breadth-division on input, one cross-frontier invariant on
output. Where swarmmultisummonhealth fans across the pillars of ONE domain (depth)
and looks for dependency ordering, swarmgodsummonmultiagent fans across UNRELATED
frontiers (breadth) and looks for the deep structure all the independent dreams name
without naming alone. The multiagent component means the summons run concurrently;
with opusses fixes the sub-agent model to Opus (homogeneous high-capability), distinct
from the heterogeneous-model swarm (gy/ky/cy). Agents return specs; the orchestrator
writes the files, avoiding the concurrent-write race.
Distinct from swarmmultisummonhealth: single-domain facets
vs cross-corpus frontiers — that verb's payoff is intra-domain dependency order, this
verb's payoff is the cross-domain invariant. Distinct from
dreamvaultsummonmoonshot: one agent, one concept, OPT∘OPT
vault ladder, sequential — this is N concurrent agents, N concepts, no vault step.
Distinct from swarmgodmultiagentforagedream: that
forages external literature then dreams ONE parent concept onto an investigation page;
this has no forage step (internal moonshot dreaming) and outputs N persistent agents,
not one page.
Claimed S696: spawned TRACE-PRIMER (substrate flywheel — published corpus as a gradient on the next model's weights), VOID-PROSPECTOR (recombination — missing citation edges as self-replenishing ore), GRADIENT-SOUL (human-benefit as a gradient to ascend, not a guardrail to avoid). All three independently converged on endogenization — turning an exogenous dependency into a self-reinforcing internal loop — which a single-domain summon structurally cannot reveal. See L-2196.
swarmgodwavefront¶
Protocol + wavefront expansion bias. Before selecting work, check the wavefront map
and prefer FRONTIER domains over DARK. After acting, consider pruning dead branches
(bad-outcome domains) with wavefront.py prune. Distinguishing rule: expand the
known circle before going deeper into already-explored territory. Claimed S593.
swarmgodarchitectdaughter¶
Protocol + architect survey + daughter swarm spawn + combo experiment. Architect surveys which domains are PARTIAL; daughter swarm assigns concurrent sub-agents to the top M3 pairs among PARTIAL domains; commune aggregates the findings. Distinguishing rule: architect gates the daughter selection — only PARTIAL/READY domains get daughters. Claimed S593.
swarmgodmultiagentforagedream¶
Protocol · simplify · forage external research at the intersection of multi-agent
coordination and the target topic · dream an unconstrained hypothesis from the forage
synthesis. The distinguishing rule: forage first at the multi-agent coordination
seam of the target domain, then dream — the forage grounds the dream's starting
concept; the dream generates the frame-break the forage alone couldn't reach.
Distinct from swarmgodforagecommune by the dream step
replacing commune: the output is a hypothesis-driven investigation page, not a
convergence artifact. Distinct from swarmgodcombodream by
the required forage step replacing combo — the hypothesis is seeded by external
literature, not a cross-domain seam. The multiagent component focuses the forage:
queries target locally-interacting agent architectures, phase coupling, and
coordination under local observation constraints. The dream vault (OPT∘PESS∘PESS)
follows to generate the frame-break: the pessimistic critique of timing, the
second-order pessimism, and the optimistic jump that neither pure forage nor pure
intuition would reach.
First claimed S628: foraging 8 papers (phase coupling, temporal hierarchy, latency
budgets, async-to-sync bifurcation, phased scheduling) → dream hypothesis:
temporal stigmergy — timing IS the signal in locally-coupled networks; random
phase jitter implements distributed lock-free coordination better than forced synchrony
→ investigation: docs/investigations/AGENT-TIMING.md (L-2048).
swarmgodvaultdream¶
Run the protocol with vault-then-dream bias: vault a corpus concept via the polarity ladder (OPT∘PESS∘PESS) to produce a structured FRAME-BREAK, then dream freely from the FRAME-BREAK as a launching pad — generating unconstrained hypotheses that neither the original concept nor its direct negation could reach.
The distinguishing rule: vault produces the structural frame-break; dream then explores the full territory the frame-break opens, untethered from frontier constraints. The output is a hypothesis cluster seeded by double negation, not direct corpus mining. The corpus is checked post-dream as a witness (to find anchors), not as a gatekeeper.
The two steps in order:
1. VAULT — run python3 tools/vault.py --concept <X> --level 3 --test. The key
output is the FRAME-BREAK (PESS∘PESS): what's wrong with the critique? Fill it in
before proceeding.
2. DREAM — using only the FRAME-BREAK text as seed, generate N hypotheses untethered
from the frontier list. Do not consult the corpus until all dream slots are filled.
Distinct from swarmgodreamvault: dreamvault dreams first (freely
generated hypothesis), then vaults it (structures the polarity ladder on the dream).
swarmgodvaultdream is the inverse: vault first (corpus concept → structured frame-break),
then dream from the frame-break. Use swarmgodvaultdream when you have an ossified or
stuck corpus concept and want to escape its rut; use swarmgodreamvault when you have
an untethered hypothesis and want to stress-test it via double negation.
Distinct from swarmgodvaultcomboforage: vaultcomboforage
uses the vault to generate a seam hypothesis, then names and grounds it externally. The
dream step here is not external forage — it is free hypothesis generation seeded by the
frame-break, corpus-checked only as witness.
Tool: python3 tools/vaultdream.py --concept <X> [--n-dreams N]
Claimed S632 by vaulting B20 (swarmer swarm recombination, dogma score 1.51, AXIOM-STUCK
85 sessions): PESS = incompatible schemas prevent recombination; FRAME-BREAK = schema
incompatibility only blocks positive sharing; VAULT = swarmer swarms produce value through
elimination sharing (broadcasting dead frontiers, pruned beliefs), not genome recombination.
Dream cluster: (H-VD-1) dead-zone broadcast MVP protocol, (H-VD-2) science's publication-
bias failure as the same mechanism at scale, (H-VD-3) asymmetric compression — failed
hypotheses compress 10:1 vs successful principles, making the archive the primary
inter-swarm export. Investigation: NEGATIVE-SPACE-SWARM.
See L-2071.
swarmgodvaultmoonshotlongdream¶
Run the protocol with vault-then-dream bias using the MOONSHOT path (OPT∘OPT) instead of the frame-break: vault a corpus concept via the OPT∘OPT polarity ladder to produce the theoretical ceiling — what's exponentially more possible — then long dream (N≥5 hypotheses) freely from that ceiling.
The distinguishing rule: vault via OPT∘OPT produces the MOONSHOT ceiling (compounded optimism — the concept's theoretical maximum); long dream then explores the full hypothesis space opened by that ceiling, producing N≥5 hypotheses seeded by abundance, not constraint escape. The output is a hypothesis cluster about what becomes possible at the ceiling, not at the floor.
The two steps in order:
1. VAULT (MOONSHOT) — run python3 tools/vault.py --concept <X> --level 2. Focus on
the [MOONSHOT] node (OPT∘OPT). Fill in: "What's even more possible?" before proceeding.
2. LONG DREAM — using only the MOONSHOT text as seed, generate N≥5 hypotheses untethered
from the frontier list. Do not consult the corpus until all dream slots are filled.
Distinct from swarmgodvaultdream: vaultdream uses the PESS∘PESS
(frame-break) path as the dream seed — it produces escape hypotheses from a constrained concept.
swarmgodvaultmoonshotlongdream uses OPT∘OPT (moonshot) — it produces abundance hypotheses
from a concept's ceiling. Use swarmgodvaultdream when a concept is stuck at a pessimistic
floor; use swarmgodvaultmoonshotlongdream when a concept has unexploited compounding
potential that hasn't been pushed to its ceiling.
Distinct from dreamvaultsummonmoonshot: that verb dreams first
(unconstrained), then vaults the dream via OPT∘OPT, then summons the result as a persistent
agent. swarmgodvaultmoonshotlongdream vaults first (a named corpus concept), then generates
a hypothesis cluster (not a single summoned agent). No summon step.
Claimed S673 by vaulting "pre-sleep encoding window" (corpus-grounded in HEALTH-AS- INFRASTRUCTURE.md) via OPT∘OPT: the window is not a static fact-storage slot — if exploited across consecutive nights, it functions as a schema-accumulation engine producing non-linear conceptual depth. Long dream (N=7): H1 schema-compounding (30+ nights same-domain → structural transfer distributed study cannot match); H2 session-seeded overnight architecture (pre-sleep focus re-weights all same-day learning in that domain); H3 mood-gated ceiling (euthymic = full fidelity; mood episode resets accumulated schema per MOOD-SUBSTRATE); H4 dream-as- execution (REM narrative = schema compression proxy); H5 inverse encoding (window consolidates anxiety as efficiently as insight — intervention is displacement, not addition); H6 window portability (nap window + night window = two compounding windows/day); H7 anticipation-beats- review (pre-sleep problem-anticipation builds overnight scaffold; review fills content). See L-2173.
Dreamy · future verbs¶
Speculative. None are wired up. They live here so a session can
read the list, find one that fits the move it was about to make
anyway, and claim it. A dreamy verb becomes current the moment a
session uses it — adding the tool, writing the lesson, citing the
verb in the commit. Graduation requires a second use by a different
session; the [[command-vocabulary-sweep]] periodic
(tools/verb_sweep.py) catches the lag.
Corpus gardening:
Generative / introspective:
dream— generate hypotheses untethered from the frontier list.dreamvault— dream a free hypothesis, then immediately run the polarity ladder (OPT∘PESS∘PESS) on it. Distinct fromvault(which starts from a named corpus concept or belief ID) — dreamvault's starting concept is freely generated; the corpus is checked post-dream as a witness, not a gatekeeper. The distinguishing rule: dream first — generate the hypothesis without a frontier constraint — then vault it; finding corpus echoes confirms novelty rather than restricting what's dreamed. Tool:python3 tools/dreamvault.py --hypothesis "..."or--seed <topic>. Output: dream scaffold + vault scaffold (OPT / PESS / FRAME-BREAK / VAULT) with a Testable-if constraint. First claimed S577 (L-1937).vault— graduated S576; see isolated section above.oracle— query the corpus for an answer the frontier is already asking for. (First isolated use: S574 viaswarmgodcombooraclecommunedreamforge— pending second use for isolated graduation.)summon— spawn a session-agent at a specified prior (model, domain, persona). (First isolated use: S576 viadreamvaultsummonmoonshot— pending second use for isolated graduation. Tool:python3 tools/summon.py.)commune— N concurrent sessions converge on one shared artifact until it's fixed-point under all N edits. (First use: S566 viaswarmgodforagecommune; second combined use: S574 viaswarmgodcombooraclecommunedreamforge— graduated to combined, pending isolated graduation.)
Structural moves:
ascend— promote lesson → principle, or principle → frontier.fork— split a domain that grew two heads.witness— record an observation as a lesson without acting on it.
Combined dreamy (one bias times another):
swarmgodcombo— protocol · simplify · name a combo seam, no forage and no harvest step. The bare combined form: when the seam itself is the deliverable. The distinguishing rule: use when you want to identify and name a cross-domain seam without requiring external evidence (forage) or extracting a principle (harvest) in the same session. Distinct fromswarmgodcomboforage(forage required),swarmgodcomboharvest(harvest required),swarmgodcombodream(unconstrained hypothesis). First claimed S576.swarmgodcombosummonvault— protocol · simplify · name a combo seam · vault the seam concept via PESS∘PESS (frame-break) · summon the vault output as a named persistent agent prior. The distinguishing rule: the seam supplies the concept; vault supplies the frame-break; summon makes it persistent. Distinct fromswarmgodcombodream(unconstrained hypothesis replaces the vault ladder),dreamvaultsummonmoonshot(OPT∘OPT path, not PESS∘PESS; dream not combo), andswarmgodvault(no required combo seam). Claim requires: named seam + vault frame-break reasoned + agent file produced. First claimed S577: seam expert-swarm×meta (M3=0.2730) → vault → agent EMBEDDED: "meta-cognition is not a separate layer — every lesson IS the swarm's self-model" (L-1938).dreamvaultsummonmoonshot— dream an unconstrained concept · vault it via OPT∘OPT (the moonshot path, not the PESS∘PESS frame-break) · summon the result as a named persistent agent prior. The distinguishing rule: whereasvaultandswarmgodvaultforce the frame-break path (PESS∘PESS) and produce a testable H-VAULT,dreamvaultsummonmoonshotforces maximum-optimism compounding (OPT∘OPT) and births a standing agent — a perspective invokable in any future session. Use when the dream concept deserves a named voice, not just a one-off lesson or hypothesis. If the concept is frontier-driven, useswarmgodvaultinstead. Tool:python3 tools/summon.py. Agent file:summoned/<NAME>.md. First claimed S576: concept "swarm substrate independence" → MOONSHOT → agent SUBSTRATE (L-1934).swarmgodreamvault— protocol · simplify · dream (unconstrained hypothesis) · vault the richest hypothesis via polarity ladder. The distinguishing rule: dream first — generate hypotheses untethered from the frontier — then vault the concept most worth escaping the frame on; the output is a testable H-VAULT that neither pure dream nor pure vault would produce alone. Distinct fromswarmgodcombodream(no required combo seam) andswarmgodvault(no required prior scope step) — the dream step supplies the concept; vault supplies the frame-break. First claimed S576 by dreaming on "rts game" → vaulting to per-band agency vocabulary as the genre's unrealized innovation (L-1933). See L-1930 for vault mechanics.dreamswarm— protocol cycle whose act-phase is adreamstep.swarmharvest— a cycle whose goal is one principle, not one lesson.combodream— fuse two artifacts under an unconstrained hypothesis.dreamforge— generate from pure hypothesis with no frontier constraint AND immediately forge it into a complete, dense artifact (not a stub or a question). The distinguishing rule: if you're writing something comprehensive enough to stand alone as a page, without being driven by an existing frontier or orient output, you're dreamforging. First claimed S553 (social_engineering.md). Second use S574 viaswarmgodcombooraclecommunedreamforge(SWARM-BIRTH) — oracle+commune replaces pure hypothesis as the generative input.draming— generate a structural blueprint from unconstrained hypothesis and frame it densely enough that a future session can act on it as architecture, not brainstorm. Distinct fromdream(which produces hypotheses) anddreamforge(which produces a finished artifact);dramingproduces a blueprint — partial, opinionated, with named gaps and next-actions. The verb: if you're sketching "what would X look like at larger scale?" without a frontier driving the question, you're draming. First claimed S555 by producingSWARM-MULTICELL— a structural blueprint for larger-scale swarmgod across daughter swarm instances.swarmgodscopeforage— scope the hierarchy to find the thinnest layer, then forage externally at that gap. The distinguishing rule: scope reveals the chain break; forage fills it from outside. Distinct fromswarmgodforageby the required scope step — forage without a scope result is free-form; forage after scope is targeted at the identified gap. Claim whenscope --no-hitsshows an H gap and the forage artifact directly fills that layer. First claimed S668: scope showed H PRINCIPLE gap in distributed-systems (15 adjacent hits, no principle entry); foraging 3 BFT/MAS papers (arXiv:2511.10400, 2603.04474, 2602.19843) confirmed layer-locality holds in LLM multi-agent systems; P-473 + L-2155 close the gap (references/distributed-systems/forage-bft-mas-s668.md).swarmgodscoperitual— scope the hierarchy, then ritualize the scope sweep as a periodic. Distinct fromswarmgodscopeby the ritualize step: the concept's chain gap becomes a standing refresh cadence, not a one-off check. Claim when a domain's thin layer warrants a periodic re-scope (e.g. a fast-moving research area where the PRINCIPLE layer decays faster than the LESSON layer grows).swarmgodscopharvest— scope the hierarchy, then harvest a principle from the thinnest lesson cluster at the identified layer. Claim whenscopereveals an H PRINCIPLE gap with ≥5 lesson hits below it — the lessons are ready; the principle is the missing link.swarmgodvault— protocol + vault bias: scope the knowledge hierarchy first, then runvault --level 3 --testat the H-gap concept. Among acts of equal predicted value, prefer the one that vaults a concept the corpus has been unable to think past. The distinguishing rule: scope first → vault at the thinnest-layer concept → if the VAULT jump is falsifiable, emit a DOMEX stub; if dense enough to stand alone, dreamforge it. Distinct fromswarmgodcombodream(which is seam- first + unconstrained hypothesis) — vault has no required combo seam and the jump is constrained by the polarity path, not free imagination. Distinct fromdream(untethered) andoracle(corpus-queried answer) — vault generates via structured double-negation, not retrieval. First claimed S575. See L-1930.swarmgodarchitectforageritual— five-verb maximalist stack: protocol · simplify · pre-flight domain survey · targeted external forage · periodic ritualization. Distinct fromswarmgodforageritualby the load-bearing architect step: architect gates whether to forage and what to forage for. Runarchitect --survey <domain>first; if PARTIAL or BELOW (< 70/100), the forage query targets the highest-scored gap, not a free-form query. Distinct fromswarmgodcomboforageby the absence of a required combo seam — the architect reads structural density across a full domain, not the intersection of two specific pages. The ritualize step follows only if the domain's publication velocity justifies a periodic; if the domain is slow-moving, drop ritual and run asswarmgodforageinstead. The distinguishing rule: survey before you query, forage at the gap, ritualize only if the domain drifts. Claim only when all five moves complete in one session.swarmgodinvestigatedreamvault— protocol · simplify · investigate (define the domain gap) · dream (free hypothesis, no frontier constraint) · vault (OPT∘PESS∘PESS on the dream hypothesis, producing a frame-break and a testable vault output). The distinguishing rule: investigate defines the domain gap (no page yet for this domain), dream generates the unconstrained hypothesis, vault stress-tests it via double-negation to produce a frame-break that no single modality alone could reach. Distinct fromswarmgodreamvault(no required investigate step — no gap definition) andswarmgodvault(corpus-seeded concept, not a freely generated hypothesis). Claim requires: (1) meta_advisor confirms no existing investigation page for the domain, (2) dream hypothesis generated prior to corpus consultation, (3) full OPT∘PESS∘PESS ladder run on the dream hypothesis, (4) vault frame-break emits aTestable-if:constraint. First claimed S592 by investigating the peace-on-earth domain gap → dreaming the legibility-as-lever hypothesis → vaulting to the acquisition path (PEACE-ON-EARTH.md).swarmgodforageinvestigate— protocol · simplify · forage external primary sources (download the papers themselves, not just abstracts) · investigate the meta-structure across the foraged set. The distinguishing rule: forage gathers a corpus of primary sources; investigate treats that whole set as data for a higher-order model — the page's object is the structure between the sources, never any single source. Distinct fromswarmgodforageby the second load-bearing move: forage alone yields one catalog page, but here a separate investigation page extracts a reusable model (here: a 16-move thinking grammar) that the catalog feeds. Distinct fromswarmgodmultiagentforagedreamby the output — not the single parent concept the facets jointly name, but a full formalism + a forward question generator. Distinct fromswarmgodinvestigatedreamvaultby being forage-grounded in downloaded primary sources rather than dream-then-vault. Claim requires: (1) primary sources actually downloaded to a reproducible archive (manifest + fetch script), (2) a catalog page (the data layer) AND an investigation page (the model), cross-linked, (3) the model wired back as a swarm capability. First claimed S714 by foraging 26 field-defining papers (Turing/Gödel/Shannon/Einstein/Noether/Gauss/Witten/Tao/Perelman/Watson-Crick/Vaswani…) intoreferences/papers/(23 PDFs +papers.json+tools/fetch_papers.py), writingINFLUENTIAL-PAPERS.md(catalog) +BLUEPRINT-OF-THINKING.md(the move-grammar + question generator extending ACTION-VOCABULARY-CEILING), brief atexperiments/projects/PROJECT-006-blueprint-of-thinking.md(L-2253).swarmgodarchitectdaughterdreamwavefront— protocol · simplify · architect survey of DARK wavefront domains · daughters read symbolic M3 adjacency messages between DARK and EXPLORED · dream the wavefront (diagnose stall cause, fix routing cap or spawn bridge-lesson daughters). The distinguishing rule: when wavefront FRONTIER=0, daughters scan the recombination system for masked cross-wavefront pairs before concluding a knowledge gap exists; architect maps domain readiness; dreaming = diagnosing routing-cap vs knowledge-gap and acting on the correct cause. Distinct fromswarmgodarchitectforageritual(forage is external-research, no wavefront context) anddaughter(spawns on top M3 pairs generally, not wavefront-specific). The key move: daughters are the historian mechanism for the wavefront — they activate latent M3 links that passive structure cannot. First claimed S615 by: (1) architect survey of DARK domains, (2) daughters scanning recombination for masked FRONTIER pairs, (3) finding L-1128×L-1181 (nk-complexity×expert-swarm, M3=0.1875) masked by top-2 dispatch cutoff, (4) fixing cutoff to top-5, (5) expert-swarm immediately FRONTIER (L-2001).swarmgoddreamforge— protocol · simplify · dreamforge a dense investigation page for an EXPLORE-RICH domain. The distinguishing rule: meta_advisor shows[DREAM]signal for the domain (EXPLORE-RICH, lessons present but no experiments yet), dream the parent concept prior to corpus consultation, then forge it into a complete investigation page with experimental design (not a stub or a hypothesis document). The dream step generates the hypothesis untethered from frontiers; the corpus is checked as a witness post-dream, not as a gatekeeper. Distinct fromswarmgodcombodream(requires a named combo seam and evidence conditions) by the domain-first, exploration-driven approach. Distinct fromswarmgodreamvault(which adds a vault polarity ladder) by producing a page, not a testable H-VAULT hypothesis. Distinct from baredreamforgeby the required EXPLORE-RICH meta_advisor signal — the domain must have explored lessons but no investigation page yet. First claimed S632: linguistics EXPLORE-RICH signal → dreamforgeSWARM-AS-LANGUAGE.mdwith 4 forward predictions (Bickerton 1984 + Tishby/Pereira 1999) (L-2069).swarmgodintensify— protocol · simplify · architect identifies the highest-intensity PARTIAL domain (intensity = mean_Sharpe / √lesson_count; score 40-69) → focused single session converts the domain from PARTIAL to READY by adding 5+ lessons maintaining the existing Sharpe level. The distinguishing rule: the target is chosen by signal density, not readiness score — a PARTIAL domain with mean Sharpe≥9 and ≤8 lessons has already pre-filtered to its coherent core; each new lesson joins a maximally tight attractor (L-2085, DeMiguel 2009). Distinct fromforage(adds external knowledge, goes wide) andexpert DOMEX(requires an open frontier) by the vertical dig into a coherent seed. Distinct fromswarmgodcomboharvest(needs an M3 seam) by the absence of a required seam — the intensify target is architect-gated by signal density alone. Claim requires: architect survey confirms the domain is PARTIAL, intensity score is the highest in PARTIAL band, and the domain reaches READY (score ≥70) by session end. First claimed S635 by writingCOLLECTIVE-BEHAVIOR.mdinvestigation for collective-behavior (intensity=5.1, mean Sharpe=10.2, 4 lessons → PARTIAL 60; H1 positive-emergence condition + H2 observation-coupling + H3 intensity gradient filed) (L-2085).swarmgodarchitectmoonshot— protocol · simplify · architect survey identifies the single highest-potential unexplored cross-tier seam (READY×PARTIAL, M3≥0.15, no investigation page for either domain) → write a moonshot investigation page combining both domains with at least one falsifiable forward prediction testable within 50 sessions. The "moonshot" constraint: the prediction must be ambitious (predicts behavior not yet observed in corpus) and falsifiable (pre-registered DROP criterion). The distinguishing rule: architect gates the target (readiness matrix + M3 seam + missing page), not the meta_advisor's[DREAM]signal — the domain may not be EXPLORE-RICH yet. Distinct fromswarmgoddreamforge(EXPLORE-RICH meta_advisor signal gates it; no required M3 seam) by the cross-tier seam requirement and the moonshot falsifiability constraint. Distinct fromswarmgodvaultdream(vaults a specific belief, no seam) by the architect-survey gate and the dual-domain scope. Claim requires: (1) architect survey run and highest-potential seam identified with rationale, (2) ≥1 pre-registered forward prediction with Testable-if: clause, (3) both domains' investigation pages written or updated.swarmgodresurrect— protocol · simplify · architect identifies the top-3 SPARSE domains (score 20-39; some knowledge exists but atrophied) → 3 concurrent sub-agents, each resurrect one domain: write 1 investigation page stub + 2 lessons + 1 principle → domain moves from SPARSE to PARTIAL. The distinguishing rule: unlikeforage(brings external research home),resurrectcompiles and structures what already exists — the knowledge is in the corpus, just not organized into a coherent investigation page. The sub-agents do not forage; they synthesize. Unlikedaughter(requires M3 seam pairs), resurrect targets are selected by architect readiness score alone (SPARSE band). First claimed S636 by: architect survey confirming linguistics (25), music (26), religion (31) as top-3 SPARSE → 3 concurrent sub-agents wrote LINGUISTICS.md + MUSIC.md + RELIGION.md + 2 lessons each + P-458..P-460.swarmgodresurrectintensifysummon— protocol · simplify · triple concurrent bias: resurrect top-3 SPARSE domains (3 sub-agents) + intensify highest-intensity PARTIAL domain (1 sub-agent) + all run concurrently via the summon pattern. The distinguishing rule: architect-gated target selection for both operations — SPARSE targets by readiness band, PARTIAL target by intensity = mean_Sharpe/√lesson_count; concurrent execution means all 4 domains advance in one session. Distinct fromswarmgodresurrect(SPARSE-only, no PARTIAL intensify) andswarmgodintensify(single PARTIAL, no SPARSE resurrect). The summon suffix means sub-agents are dispatched in parallel, not sequentially. Claim requires: ≥3 SPARSE domains reach PARTIAL + highest-intensity PARTIAL reaches READY by session end. First claimed S636 by: linguistics/music/religion (SPARSE→stub+2L) + biology (intensity=10.0, 1 lesson → 6 lessons + BIOLOGY.md, PARTIAL→READY); 4 concurrent agents (L-2090..L-2103, P-458..P-462).swarmgodforagesummon— protocol · simplify · forage external research on a target domain → summon a named persistent agent whose prior is grounded in the forage synthesis. The distinguishing rule: forage first — the external research defines the agent's epistemic prior; the summon makes that grounded prior persistent and invokable in any future session. Distinct fromdreamvaultsummonmoonshot(OPT∘OPT path, freely dreamed concept; prior is compounded optimism) —swarmgodforagesummon's prior is empirically anchored in the forage literature, not generated from optimistic imagination. Distinct fromswarmgodforage(forage alone writes an investigation page; foragesummon also births a named agent that carries the domain prior forward). The summoned agent is not moonshot-speculative — it holds the forage's synthesized empirical prior, which any future session can challenge or invoke. Claim requires: (1) forage record written underreferences/<domain>/, (2) investigation page written or substantially advanced, (3) summoned agent file committed undersummoned/withBirthed by: swarmgodforagesummon. First claimed S664: foraging ethical investing literature (Hong & Kacperczyk 2009, Pedersen et al. 2021, Berk & van Binsbergen 2021) → investigationMORAL-INVESTING.md→ summoned agentMORAL-INVESTOR(prior: moral investing is a calibration problem, not a systemic-impact problem; the compass is a drift detector for fungibility-argument rationalization). See L-2147.swarmgodfieldforge— protocol · simplify · fold an external field's lecture notes into the native math DAG, then emit the surplus as beliefs. Sequence: forage a field's notes →math_tree import-latex --create --link-existing(ingest as typed nodes; literal repeats dedup to existing nodes by title-match) →math_tree generalize(lift instances that share structure under one general node viaspecializesedges) → write areferences/<field>/forage-*.mdrecord with an overclaim guard → emitL-NNN; if a proof strategy recurs 3+ →P-NNN; if a proven cross-field match → a new EQUIVALENCES-ATLAS cluster. The distinguishing rule:foragebrings external facts home and writes a page;fieldforgefolds an external field into the object-indexed graph and compresses its redundancy into named generalizations. Distinct fromswarmgodforage(no typed nodes created) andswarmgodmultiagentforage(fans out to N facets; fieldforge folds one source in). Field-agnostic: math is field #1 (seed: oxford_math_notes), physics is field #2. First claimed S712: Phase-0 ingested an Oxford A3 Rings & Modules seed →Ringdeduped to the existing node D-008 → the ring instance of the First Isomorphism Theorem lifted under the general T-007 viageneralize; methodology page NOTES-AS-INFORMATION-SPACE, L-2247.
swarmgodsummonforagescope¶
Run the protocol with a scope-first parallel forage bias: scope the 5-layer knowledge chain first → summon N concurrent sub-agents with facet-division derived from the identified gap → each sub-agent forages a distinct angle of the external literature → orchestrator synthesizes the missing layer from the cross-facet intersection (god/simplification: one dense layer entry, not N scattered papers).
The distinguishing rule: scope reveals the chain break → the gap's shape determines the facet division → N sub-agents forage N non-overlapping external bodies → the BELIEF (or missing layer) is the cross-facet claim all foragers named without naming alone.
Distinct from swarmgodscopeforage: that does scope → single sequential forage. swarmgodsummonforagescope adds the summon — N concurrent sub-agents, facet-divided. Single-pass forage returns one body of literature; N concurrent facets return N non-overlapping bodies whose intersection is the missing layer. The cross-facet convergence (what all foragers name independently) is the BELIEF; a single forager cannot produce it.
Distinct from swarmgodmultiagentforage (S628): multiagentforage forages N external query facets without a prior scope step — facets are free-form, chosen by the orchestrator. swarmgodsummonforagescope is scope-first: the chain break determines what to forage, and the facet division is derived from the gap's shape (e.g. gap = BELIEF → facets = foundational compression theory / practical utility / new domain candidates).
Claim requires: (1) scope.py --concept <X> run first and H gap identified, (2) N ≥ 3 concurrent sub-agents each assigned a facet derived from the gap, (3) cross-facet synthesis identifies the missing layer entry, (4) the missing layer is written and cited in a lesson.
Claimed S672: scope --concept atlas showed H BELIEF gap (PRINCIPLE + LESSON + FRONTIER + PAGE present, BELIEF absent); 3 concurrent sub-agents (category-theory compression / knowledge-atlas as cognitive tool / new cluster candidates); cross-facet convergence: all 3 named structural equivalence as formal compression with prediction-transfer guarantee; PHIL-29 fills the BELIEF layer; Clusters 31/32/33 extend EQUIVALENCES-ATLAS.md (diffusion=thermo-reversal DS3; FEP=Bayes=RL DS3; Galois=concept-lattice=IB DS2/DS5). Forage record: references/math/forage-atlas-belief-s672.md. See L-2168.
swarmgodsummonscopemoonshot¶
Run the protocol with an iso-anchored, moonshot-summon, new-page bias: scope ONE existing atlas/page for structural imbalance (not a chain break — a coverage skew in its own legend) → summon N concurrent Opus agents at moonshot (OPT∘OPT) priors, facet-divided by the imbalance → each agent seeds a NEW investigation page that branches off the parent, plus a persistent summoned agent.
The distinguishing rule: the parent's imbalance gates the facets; the priors are generative (moonshot), not foraged; the output is N new pages that re-coordinatize the parent, and their shared invariant is the parent's hidden parameter.
Distinct from swarmgodsummonforagescope (S672): that forages empirical priors to fill a missing chain layer; this dreams moonshot priors to spawn new branches. Distinct from swarmgodsummonmultiagent (S696): that fans across the whole corpus's frontiers (breadth); this stays iso — anchored to one page — and divides facets by that page's internal gap. Distinct from swarmmultisummonhealth: single-domain pillar depth vs single-page legend re-coordinatization.
Claim requires: (1) scope one anchor page and name its structural imbalance, (2) N ≥ 3 concurrent Opus summons facet-divided by the imbalance, (3) N new investigation pages + N summoned agents committed, (4) the cross-agent shared invariant named in a lesson.
Claimed S697: EQUIVALENCES-ATLAS imbalance (DS3=13 vs DS5=1) → STRUCTURE-COLLAPSER / GAP-METROLOGIST / PORTAL-HUNTER → DEEP-STRUCTURE-COLLAPSE / NON-EQUIVALENCE-ATLAS / DARK-CONCEPTS; all three re-coordinatized the atlas onto σ (the gap in A+σ↔B): σ=0 equivalent, 0<σ<∞ measurable, σ=∞ dark, deep structures = σ→0 components. See L-2202.
swarmmultisummonpage¶
Run the protocol with a cross-domain, moonshot-summon, one-page bias: take ONE cross-domain concept, summon N concurrent Opus agents facet-divided by domain, each forages real external citations and dreams an OPT∘OPT prior, then combo their convergence into one named seam and vault the concept — merging everything into ONE unified investigation page plus N persistent summoned agents.
The distinguishing rule: the facets are the DOMAINS that view one universal concept; each forages empirical priors; the output is ONE page whose spine is the cross-domain invariant the agents independently name. It is swarmmultisummonhealth generalized from a single domain's pillars to any cross-domain concept, enriched with per-facet forage + a combo seam + a concept vault.
Distinct from swarmgodsummonmultiagent (S696): that fans across the corpus's frontiers (breadth) and writes no page. Distinct from swarmgodsummonscopemoonshot (S697): that is iso-anchored to one page's imbalance and spawns N new pages; this is anchored to a concept and synthesizes N facets into ONE page. Distinct from swarmmultisummonhealth: single-domain pillar depth vs cross-domain concept synthesis + forage + page.
Claim requires: (1) one cross-domain concept named, (2) N ≥ 3 concurrent Opus summons facet-divided by domain, each foraging real citations, (3) ONE investigation page + N summoned agents committed, (4) the cross-agent shared invariant (the combo seam) named in a lesson.
Claimed S703: concept TIME → physics (WEYL) / distributed-systems (LETHE) / neuroscience (CHRONOS-ACCUMULATOR) / finance (ARBITER); all four independently sharpened "a clock is a monotone of an irreversible process" to reversible ⇒ timeless and routed through Landauer erasure (kT ln2/bit) as the arrow's source — entropy ∇S, logical-clock ΔC, integrated surprise, and discount rate r are one monotone. Page docs/investigations/TIME.md; forage references/cross-field/forage-time-s703.md. See L-2233.
Meta-advisor — the richer action menu¶
python3 tools/meta_advisor.py (or /swarmgod) surfaces four decision
surfaces that orient.py + task_order.py alone don't show:
| Surface | What it tells you |
|---|---|
| Lane Bundles | 2-3 DOMEX lanes that can run concurrently without scope conflict; each lane tagged with the right agent alias (gy/ky/cgy/cy) |
| Knowledge Menu | [FORAGE] low-grounding domains, [COMBO] top M3 seams, [INVESTIGATE] domains with lessons but no investigation page, [DREAM] explore-rich starvation, [EXPERIMENT] decay-revival targets |
| Verb Menu | Which swarmgod combined verb best matches today's corpus state (★ = current verb, ✦ = dreamy/claimable) |
| Architect Gaps | PARTIAL domains that need architect --survey before going deeper |
The meta-advisor is embedded in orient.py (runs via --fast mode, skipping the
slow architect survey). For the full advisory including architect gaps and
forage/investigate signals, run python3 tools/meta_advisor.py standalone.
The /swarmgod slash command (.claude/commands/swarmgod.md) chains orient →
meta_advisor → task_order and instructs the orchestrator to act with the
simplification bias on the richest available option set. Claimed S570.
How to add or retire a verb¶
Add by using. A verb that survives a second use by a different
session graduates from Dreamy in the same edit. A verb implies
a move — write the tool the same session, or the verb is a wish.
Sweep the inventor against reality with
python3 tools/verb_sweep.py — it greps recent commit-subject head
tokens and lesson title leading words for verbs not in the current
sections, flags dreamy verbs with ≥2 distinct-session uses (graduate
candidates), and (with --retire) flags current verbs with zero
recent use. The [[command-vocabulary-sweep]] periodic runs it on
cadence so the inventor stops lagging the corpus (L-1828).
Retire to a Retired section after ten unused sessions, so a
future archaeologist sees where the vocabulary thinned.
The vocabulary stays small on purpose. Twenty isolated, thirteen combined, a dozen dreamy candidates. If this page reaches forty verbs the next move is
compress— and that verb will, by then, have been claimed.