A continuation of PANTHEON-COMPLETE.REDACTED.txt. The complete source identity, every recorded occurrence, and linked equations are on its first page.
Source transcript · continued
Download complete source text ↓3,729 lines · 194.4 KB
4.7 THE CODE CORPUS, /opt/mm/code on megamind-agi
--------------------------------------------------------------------------------
1.2 GB. The "learn every programming language" work.
mmcode :9524 GET /recall?q=&k= | /reindex | /health | /caps
caps code programming language software rust python java javascript cpp
golang function api syntax library algorithm compiler framework stdlib
Same mmwikix binary as the Bubbles recall organs, holding code not prose.
mmparse :9526 GET /recognize?code= | /comprehend?code=&lang= | /stress |
/health | /caps
caps understand comprehend parse lex tokenize grammar recognize language
code structure ast function deterministic
/recognize identifies the language. /comprehend builds an AST.
mmscheme :9527 GET /execute?code= | /stress | /health | /caps
mmpy :9528 GET /execute?code= | /health
mmc :9529 GET /execute?code= | /health
TOGETHER these form recall, comprehension, then execution. mmcode finds code,
mmparse builds an AST, and the three executors actually RUN it to verify
semantics. That is what separates this from a code search index.
Supporting files in /opt/mm/code:
langs_master.txt the Wikidata programming language list
proof_langs.txt, proof_run.sh
crawl.py, crawl.log, recognize_det.py, recognize_proto.py
measure_parse.py, c_fuzz.py, py_fuzz.py, launch_full.sh
--------------------------------------------------------------------------------
4.8 DOMAIN ORGANS
--------------------------------------------------------------------------------
atlas Bubbles :8095
caps agency brain seo eo domain expert marketing status live
Returns {"error":"unauthorized"} without credentials. ORACLE reports health 2
and atlas is one of the organs that actually WINS coalitions, so it is
genuinely reachable by the federation.
hephaestus Bubbles :8770
caps forge build website site create landing page web design deploy publish
craft builder studio seo discoverable business homepage generate site
portfolio creative art gallery showcase photographer designer artist
agency architect illustrator case study masonry visual brand identity
generative
Self describes: {"ok":true,"organ":"hephaestus","service":"forge",
"god":"Hephaestus","role":"builds discoverable websites",
"uses":["atlas","oracle/pantheon"]}
It consumes ORACLE, making it both an organ and a client of the federation.
bubbles-brain Bubbles :8893
caps brain recall insight memory node megamind associate knowledge
Returns {"status":"healthy"}.
codex VALKYRIE :8790
caps codex lore bible fantasy creature monster deity item place spell
semantic search recall knowledge entity world
In BOTH mmbroad's fan out and mmorch's execute list.
mm-gateway M4 :8090
caps generate language model llm answer question reason completion continue
paragraph text write
A voice. Qwen via MLX. NOT the only floating point component: VALKYRIE runs a
second MLX Qwen server on :8081, unregistered. See PART XIII section 13.4.
--------------------------------------------------------------------------------
4.9 TIMER ORGANS
--------------------------------------------------------------------------------
No port. Registered so ORACLE knows they exist and can reason about them.
mmheart mm-heart.service heartbeat, liveness, homeostasis, watch pri 8
mmgate mm-gate.service safety gate, policy, approve, reject pri 12
mmeval mm-eval.service evaluation, test, score, benchmark pri 12
mmcap mm-cap.service capacity monitor, growth, saturation pri 12
mmharvest mm-harvest.service continuous multi source knowledge harvest pri 24
Plus, not registered but running on Bubbles:
mmcurator curator.sh every 4 min, the federation self knowledge loop
mmcurator-db curator-db.sh database curation
mmteach-cycle teach-cycle.sh periodic teaching cycle
arxiv-crawl arxiv-crawl.sh OAI-PMH harvest into arxiv.bsig
gutenberg-crawl, wiki-harvest
mm-live-reindex, mm-estate-reindex, mm-legacy-reindex
================================================================================
PART V COMPLETE API REFERENCE
================================================================================
--------------------------------------------------------------------------------
5.1 CONVENTIONS
--------------------------------------------------------------------------------
Almost every organ answers GET with a query string, not POST with JSON.
The sole exception found: mmreason :9520 takes POST /reason {"query":...}
Almost every organ exposes /health and /caps.
Query values must be URL encoded. Spaces as %20.
THE DISCOVERY TRICK
Every organ lists its own verbs when you call an invalid route. That error IS
the documentation and it is always current:
$ curl -s "http://100.90.97.104:9503/__anything__"
{"error":"GET /recall?q=&k= | /health | /caps"}
When unsure, call a nonsense route and read the answer.
--------------------------------------------------------------------------------
5.2 VERB TABLE, EVERY ORGAN
--------------------------------------------------------------------------------
BUBBLES 100.90.97.104
9500 mmk GET /health | /caps | /answer?q= | /root
9501 mmk2 GET /health | /caps | /answer?q=CLAIM s=..|p=.. | /root
9502 mmk3 GET /health | /caps | /answer?q=CLAIM s=..|p=.. | /root
9503 mmbroad GET /recall?q=&k= | /health | /caps
9504 mmcert GET /ask?q=&k= | /certify?q=&e= | /verify?q=&e= | /health | /caps
9505 mmorch GET /ask?q=&k= | /trace?q= | /health | /caps
9506 mmdb GET /sources | /health | /caps
9507 mmcog GET /q | /tick | /health | /caps
9508 mmreshit GET /select?q=&k= | /health | /caps
9509 mmepona GET /update?q=&k=&eta=&lambda= | /health | /caps
9510 mmstdp GET /wire?ctx=&cert=1 | /edges?top= | /neighbors?node=&n= |
/health | /caps
9511 mmact GET /choose?actions=a|b|c | /reset | /health | /caps
9512 mmevolve GET /evolve?state=&input=&steps= | /health | /caps
9513 mmlattice GET /project | /activate?q=&k= | /cell?i= | /health | /caps
9514 mmloop GET /run?cycles=N | /reset | /health | /caps
9515 mmteach GET /teach | /health | /caps
9516 wikipedia GET /recall?q=&k= | /reindex | /health | /caps
9517 books GET /recall?q=&k= | /reindex | /health | /caps
9518 mmask GET /ask?q= | /health | /caps
9520 mmreason POST /reason {"query":...} | GET /health | GET /capability
9521 arxiv GET /recall?q=&k= | /reindex | /health | /caps
9522 legacy GET /recall?q=&k= | /reindex | /health | /caps
9523 live GET /recall?q=&k= | /reindex | /health | /caps
9525 estate GET /recall?q=&k= | /reindex | /health | /caps
MEGAMIND-AGI 100.69.120.61
9096 mmsem GET /sem?q=text
9097 mmgame /game/remember?agent=&imp=&text= | /game/recall?agent=&k=&text=
/game/social?agent=&k= | /game/health
9098 mmpredict GET /gen
9099 mmserve /recall?q=&k= | /health | /reload
9401 mmworld GET /world/recall?q=tok1|tok2 | /world/health
9402 ORACLE see 5.3
9403 mm-narrate GET /narrate?sents=&theme=&seed=
9404 eye (rejects unknown powers by name)
9410 loom /health | /organ?q= | /world/overview | /wheel | /chronicle?k=
/narrate?beat= | /narrategate?prose=&source=&decoy=
/worldmap?k= | /heritages
/forge?heritage=&sex=&background=&seed=
/play?heritage=&sex=&allegiance=&action=&target=
/enter?name=&heritage=&sex=&place= | /act?player=&action=&ta...
9524 mmcode GET /recall?q=&k= | /reindex | /health | /caps
9526 mmparse GET /recognize?code= | /comprehend?code=&lang= | /stress |
/health | /caps
9527 mmscheme GET /execute?code= | /stress | /health | /caps
9528 mmpy GET /execute?code= | /health
9529 mmc GET /execute?code= | /health
--------------------------------------------------------------------------------
5.3 ORACLE ENDPOINTS 100.69.120.61:9402
--------------------------------------------------------------------------------
GET /oracle/health
GET /oracle/organs
GET /oracle/register?name=&caps=&url=&mode= name+caps required
GET /oracle/orchestrate?q= route only
GET /oracle/execute?q= route and call
GET /oracle/coalition?q= form, call, score by trust
GET /oracle/learn write into the entity graph
GET /oracle/ask?q= entity centric, NOT free text
GET /oracle/recall /inverse /hop /analogy /similar /facts /path /set
/consistent /unlearn /revise /counterfactual /induce /predict
/abduce /factor /map
GET /oracle/orchstress /execstress /coalstress /fedstress
/federation /pantheon /pantstress
GET /oracle/seferim-* ~40 staged endpoints S0..S16
--------------------------------------------------------------------------------
5.4 WORKED EXAMPLES, all actually run
--------------------------------------------------------------------------------
ASK THE ESTATE CORPUS DIRECTLY
curl -s "http://100.90.97.104:9525/recall?q=missionMATCH%20cleared%20workforce&k=2"
{"organ":"estate","binary":true,"corpus":156229,"dimension":512,
"hits":[{"id":153691,"hamming":119,"similarity_permille":767,
"body":"Cleared Workforce Intelligence missionMATCH, screenMATCH..."}]}
ASK THE WHOLE FEDERATION
curl -s "http://100.90.97.104:9503/recall?q=PANTHEON%20organs&k=3"
{"power":"broad-recall","organs_consulted":9,
"organs_reached":[{"organ":"megamind-m4","reached":false},...],
"reached_count":6,"hits":[...]}
ROUTE A QUESTION
curl -s "http://100.69.120.61:9402/oracle/orchestrate?q=missionMATCH"
-> winner, winner_url, ignited, salience, coalition[], math{}
REGISTER AN ORGAN
curl -s "http://127.0.0.1:9402/oracle/register\
?name=mm-estate\
&url=http%3A%2F%2F100.90.97.104%3A9525\
&mode=remote\
&caps=estate+own+projects+clients+missionMATCH+ATLAS+PANTHEON+..."
-> {"registered":{...},"persistent":true}
FULL ORCHESTRATED ANSWER
curl -s "http://100.90.97.104:9505/ask?q=missionMATCH&k=3"
-> oracle{}, organs_executed[], verdict, reason
INGEST A DOCUMENT TREE
/home/user/mmk/mmdoc /path/to/docs --corpus live
CHECK AN ORGAN
curl -s "http://100.90.97.104:9525/health"
{"status":"ok","organ":"estate","role":"binary Hamming recall over a signature
store (mmap popcount, exact NN)","corpus":156229,"binary":true}
================================================================================
PART VI CORPORA
================================================================================
--------------------------------------------------------------------------------
6.1 THE FOUR FILE PATTERN
--------------------------------------------------------------------------------
<name>.bsig signature store. 64 bytes (512 bits) per record. APPENDED.
<name>.blob evidence. the actual passage text. APPENDED.
<name>.idx the index mmwikix mmaps to serve. REBUILT, never appended.
<name>.seen dedup keys. mmbsig checks this so re-ingest is idempotent.
.bsig and .blob grow together on ingest.
.idx must be REBUILT after every ingest. If you skip it, the organ serves the
old index and your new knowledge is invisible while every check looks green.
--------------------------------------------------------------------------------
6.2 EVERY CORPUS MEASURED
--------------------------------------------------------------------------------
books / gutenberg THE LARGEST
gutenberg.bsig 9,496,573,632 bytes 9.5 GB
gutenberg.blob 20,670,709,067 bytes 20.7 GB
gutenberg.idx 1,187,071,704 bytes 1.19 GB
records 148,383,963
served by books :9517
source gutenberg-crawl.sh
content Project Gutenberg. prose, philosophy, history, science.
wiki THE ENCYCLOPEDIA
wiki.bsig 3,137,986,944 bytes 3.1 GB
wiki.blob 6,340,461,661 bytes 6.3 GB
wiki.idx 392,248,368 bytes 392 MB
wiki.seen 1,656 lines
records 49,031,046
served by wikipedia :9516
source wiki-harvest.sh, wiki-crawl.sh, wiki-dump.sh
legacy THE SALVAGE
legacy.bsig 207,600,832 bytes 208 MB
legacy.blob 117,460,248 bytes 117 MB
legacy.idx 25,950,104 bytes 26 MB
legacy.seen 3,243,763 lines
records 3,243,763
served by legacy :9522
content salvaged huggingface models, tensor fingerprints, wikidata
concepts, the maddie archive.
9,908 models, 346K tensors, 485K urls.
arxiv THE RESEARCH
arxiv.bsig 193,881,216 bytes 194 MB
arxiv.blob 433,413,544 bytes 433 MB
arxiv.idx 24,235,152 bytes 24 MB
arxiv.seen 3,029,394 lines
records 3,029,394
served by arxiv :9521
source arxiv-crawl.sh, OAI-PMH across sets
cs math stat eess econ q-bio physics q-fin
live THE FRESH
live.bsig 5,978,752 bytes 6.0 MB
live.blob 17,879,327 bytes 17.9 MB
live.idx 747,024 bytes 747 KB
live.seen 93,418 lines
records 93,418
served by live :9523
source mm-harvest.service, continuous multi source
NOTE also received the 2026-07-28 Desktop ingest, jumping from 66,419 to
93,418. It is inside mmbroad's fan out, which is the ONLY reason estate
content is reachable through broad recall at all.
knowledge / estate YOURS
knowledge.bsig 9,998,656 bytes 10 MB
evidence.blob 39,517,700 bytes 39.5 MB SHARED, see note
knowledge.idx 1,249,832 bytes 1.25 MB
seen.k
records 157,152
served by estate :9525
source mmdoc over ~/Desktop
content your projects, clients, manuals, architecture docs,
PDFs, OCR'd images, RTF, and this documentation set.
Includes credential material by explicit instruction.
NOTE this corpus uses evidence.blob rather than knowledge.blob, breaking the
naming pattern. That is by design: mmbsig's own usage string names them
"<knowledge.bsig> <evidence.blob> <seen.k>".
TOTALS
records ~204,938,736 passages
signatures ~13.0 GB
evidence ~27.6 GB
indexes ~1.63 GB
grand total ~42.2 GB
--------------------------------------------------------------------------------
6.3 THE PASSAGE CONTRACT, non negotiable
--------------------------------------------------------------------------------
one passage per line
40 to 400 characters
NO pipe character
must contain real words
WHY 40 TO 400
Capacity is roughly K* = D/43. At D=512 that is about 11 items bundled before
the majority vote stops reliably recovering any single member. A passage
longer than a few sentences bundles too many tokens and becomes unrecallable.
WHY NO PIPE
Pipe is the claim delimiter:
s=kb:substrate | p=learned_count | o=#231577 | pol=pos
A pipe inside a passage corrupts claim parsing. It is also the argument
delimiter for mmact /choose?actions=a|b|c and mmworld /world/recall?q=t1|t2.
--------------------------------------------------------------------------------
6.4 ADDING KNOWLEDGE CORRECTLY
--------------------------------------------------------------------------------
1. SHAPE passages 40 to 400 chars, no pipe, deduped
2. INGEST mmbsig ingest <lines> <bsig> <blob> <seen>
appends, dedups automatically, returns records_new and deduped
3. REINDEX systemctl start mm-<name>-reindex
writes .idx.new then mv into place, atomic
4. RESTART systemctl restart mm-<name>
so the organ mmaps the new index
5. VERIFY curl the organ /health, confirm corpus count rose,
then /recall a passage you know is in there
6. REGISTER if the organ is NEW, register with ORACLE using caps written in
the vocabulary of the questions you expect
mmdoc does 1 through 4 automatically with --corpus live.
It does NOT do 6. Registration is always manual and deliberate.
SKIPPING 3 OR 4 the most common failure. Ingest reports success, the file
grows, nothing is findable.
SKIPPING 6 the knowledge is servable but never routed to.
WHICH CORPUS FOR WHAT
your own work, clients, projects, docs -> knowledge.bsig, estate
external continuous harvest -> live.bsig
encyclopedic content -> wiki.bsig
papers -> arxiv.bsig
books -> gutenberg.bsig
code -> /opt/mm/code on megamind-agi
================================================================================
PART VII DATA FLOW
================================================================================
--------------------------------------------------------------------------------
7.1 A QUESTION, END TO END
--------------------------------------------------------------------------------
YOU
|
| GET http://100.69.120.61:9402/oracle/orchestrate?q=...
v
+-------------------------------------------------------------------+
| ORACLE megamind-agi :9402 |
| |
| a. tokenise |
| b. task_sig = majority_bundle(token hypervectors) D = 2048 |
| c. for each of 51 organs: |
| hamming = popcnt(task_sig XOR capability_sig) |
| score = (2048 - hamming) + 96*overlap + 32*live |
| + 16*local + priority |
| d. sort, winner = highest, ignited if salient |
| e. broadcast_sig = majority_bundle(task, winner_caps) |
+----------------------------+--------------------------------------+
|
+---------------------+---------------------+
| |
ROUTE ONLY EXECUTE
returns winner_url ORACLE calls it itself
caller fetches it ONLY FOR ORGANS WITH A
| COMPILED CONTRACT
v v
+--------------------+ +----------------------+
| KNOWLEDGE ORGAN | | COALITION |
| e.g. estate :9525 | | top members called |
| | | scored by trust: |
| q -> 512 bit sig | | route_score |
| popcount scan over | | +768 pass |
| the mmap'd .bsig | | +256 valid_json |
| argmin hamming | | +128 called |
| return k hits | | +min(bytes,2048)/4 |
| body + permille | | -256 not_called |
+---------+----------+ +----------+-----------+
| |
+--------------------+---------------------+
|
v
+-----------------------+
| CERTIFICATION |
| mmcert :9504 |
| mmk / mmk2 / mmk3 |
| |
| CERTIFIED iff query |
| content tokens are a |
| SUBSET of evidence |
+-----------+-----------+
|
+------------+------------+
| |
GROUNDED NOT GROUNDED
| |
v v
+-----------------+ "The indexed and verified sources
| mm-gateway | do not support an answer."
| M4 :8090 |
| Qwen renders | REFUSAL IS A FEATURE.
| evidence into | mmorch returns:
| prose | "refusing rather than fabricate"
+--------+--------+
|
v
ANSWER
THE LANGUAGE MODEL IS LAST AND ONLY EVER SEES CERTIFIED EVIDENCE.
It is never asked what it knows. That is the entire anti fabrication design.
--------------------------------------------------------------------------------
7.2 THE THREE INGESTION DOORS
--------------------------------------------------------------------------------
DOOR 1 mmdoc, documents
/home/user/mmk/mmdoc <path> [--corpus live|knowledge] [--dry] [--no-ocr]
[--dpi N] [--max-mb N] [--exclude a,b] [--out F]
file -> extractor -> passages -> OCR gate -> mmbsig -> reindex -> restart
EXTRACTORS
.txt .md .csv .tsv .json .sql .js .ts .css .yaml .sh .py .go .conf .ini .log
read directly
.html .htm .xhtml tag strip and entity decode
.rtf .rtfd textutil, else control word strip
.pdf pdftotext text layer. under 60 words it is treated
as scanned, rasterised with pdftoppm and OCR'd
images tesseract OCR
everything else reported unsupported and skipped
THE OCR PROSE GATE
Text derived from OCR must additionally read like prose: at least 6 tokens,
60 percent of them plausible words, and none of the tesseract artefact
signatures. Real documents bypass the gate entirely.
Without it, diagrams contributed "Lj eALMStracusto/e Source ALvis(p,q.cu,tl)".
With it they contribute "Atlas fixes real websites, verifies the result".
DOOR 2 the crawlers, external harvest
arxiv-crawl.sh OAI-PMH -> arxiv.bsig
gutenberg-crawl.sh -> gutenberg.bsig
wiki-harvest.sh -> wiki.bsig
mm-harvest.service -> live.bsig
All share the same passage contract.
DOOR 3 the curator, federation self knowledge
curator.sh, every 4 minutes as mmcurator.timer
PHASE A DRAIN
read the next 5000 lines of darkstore.txt from the cursor
mmingest into the M4 substrate at 100.94.8.94:9440
advance the cursor
PHASE B REFRESH
once drained, re-scan LIVE federation self knowledge
append only genuinely new claim lines to darkstore.txt, dedup by exact line
darkstore.txt 290,592 lines
cursor 290,592, fully drained
WARNING PHASE A targets 100.94.8.94:9440 which is DEAD. Nothing is being
lost right now only because the queue is empty. See Part XI issue 2.
--------------------------------------------------------------------------------
7.3 THE FOUR LEARNING LOOPS
--------------------------------------------------------------------------------
LOOP 1 PLASTICITY, Bubbles
co-activation -> mmstdp :9510 wires a directed weighted graph
-> integer weight accumulates on repeated co-firing
-> mmteach :9515 takes edges with w >= 200
-> POSTs them into ORACLE's graph via /oracle/learn
The w >= 200 threshold stops weak correlation becoming asserted fact.
This is the only path by which Bubbles cognition writes into megamind-agi.
LOOP 2 COGNITIVE CLOSURE, Bubbles
mmloop :9514 turns 8 kernels into one cycle:
mmreshit choose a path Omega = argmin[S+U-V]
mmevolve evolve the state dx/dt = D+N+F+T
mmact select an action a* = argmax J
mmepona update toward value
mmstdp rewire
mmlattice project to R^5000
mmcog measure Q
-> back to mmreshit, hill climbing Q
LOOP 3 ORACLE GRAPH
74,271 entities, 510,124 edges, persisted to oracle.wal which is actively
written. Powers: learn, unlearn, revise, induce, abduce, counterfactual.
Symbolic relations, not passages.
LOOP 4 HARVEST
mm-harvest plus the crawlers continuously grow live.bsig and the external
corpora, with the reindex units rebuilding indexes atomically as they grow.
--------------------------------------------------------------------------------
7.4 WHERE A FACT PHYSICALLY GOES
--------------------------------------------------------------------------------
Teach PANTHEON one sentence today:
1. mmdoc shapes it into a passage, 40 to 400 chars
2. mmbsig hashes it to a 512 bit signature, 64 bytes
3. signature appended to knowledge.bsig
raw text appended to evidence.blob
dedup key appended to seen.k
4. mmwikix reindex rebuilds knowledge.idx over the whole store
5. mm-estate restarts and mmaps the new index
6. the organ reports corpus 157,153 instead of 157,152
7. ORACLE already knows mm-estate exists, so a matching question routes there
8. the passage returns as evidence with a hamming distance and a permille
NO RETRAINING. No embeddings recomputed. No model touched. Seconds, not hours.
That is the practical payoff of the binary design.
--------------------------------------------------------------------------------
7.5 THE INVISIBILITY FAILURE MODE
--------------------------------------------------------------------------------
Knowledge can be present and unreachable. THREE conditions must ALL hold:
1. the passage is IN a .bsig store ingest succeeded
2. an ORGAN SERVES that store unit exists, index built, restarted
3. the organ is REGISTERED with ORACLE and its caps contain the vocabulary of
the question token_overlap > 0
Break any one and the knowledge is invisible while every individual check looks
green. This happened twice on 2026-07-28:
FIRST knowledge.bsig was ingested with 21,029 passages and no organ served
it. mmbsig recall by hand worked perfectly. Every federation query
missed it entirely.
SECOND mm-live served the content, but its caps described news and academic
material, so ORACLE routed estate questions to mm-narrate and mmgame.
ADD A FOURTH CONDITION FOR AGGREGATOR REACHABILITY:
4. the organ is on the aggregator's compiled in list
mmbroad has 9 organs, mmorch has 6. Newly registered organs are on neither.
================================================================================
PART VIII MATHEMATICS
================================================================================
Every formula quoted from the system's own self description.
--------------------------------------------------------------------------------
8.1 THE CORE IDEA
--------------------------------------------------------------------------------
Everything is a bit vector. Meaning is a point in a very high dimensional binary
space. Similarity is Hamming distance. Composition is majority bundling.
D = 512 knowledge signatures, every passage in every corpus
D = 1024 mmreason shadow organ
D = 2048 capability hypervectors, ORACLE routing
D = 5000 mmlattice consciousness projection target
A 512 bit signature is 64 bytes. mmbsig reports "width":64 on every ingest.
--------------------------------------------------------------------------------
8.2 RETRIEVAL
--------------------------------------------------------------------------------
"binary Hamming recall over a signature store (mmap popcount, exact NN)"
query text -> 512 bit signature q
for each stored signature s_i:
d_i = popcount(q XOR s_i)
return argmin d_i
EXACT true nearest neighbour, not approximate. no ANN index, no
quantisation, no recall/precision tradeoff.
MMAP the store is memory mapped. a 9.5 GB corpus is not loaded into
RAM; the kernel pages in what the scan touches.
DETERMINISTIC popcount is integer. the same query returns the same neighbour
on every machine, forever.
similarity_permille = 1000 * (D - hamming) / D
observed: hamming 119 at D=512 -> 1000*(512-119)/512 = 767
--------------------------------------------------------------------------------
8.3 ROUTING
--------------------------------------------------------------------------------
signature binary majority bundle of token hypervectors
distance hamming = popcnt(task XOR capability)
score (D-hamming) + 96*token_overlap + 32*live + 16*local + priority
broadcast majority_bundle(task, winner_capabilities)
D - hamming base similarity, 0 to 2048, in practice ~1000 to 1100
96 * overlap literal token overlap. USUALLY DECISIVE.
32 * live live beats dead
16 * local local beats remote
priority 8 to 80 per organ
--------------------------------------------------------------------------------
8.4 COALITION TRUST
--------------------------------------------------------------------------------
member_trust = route_score + 768*pass + 256*valid_json + 128*called
+ min(bytes,2048)/4 - 256*not_called
An organ that routes best but cannot be called is punished by 256 and forfeits
1152 more. Observed: mm-estate route_score 1869 -> trust 1613, losing to atlas
at 2877.
--------------------------------------------------------------------------------
8.5 BUNDLING
--------------------------------------------------------------------------------
majority_bundle(a, b, c, ...) = bitwise majority vote per bit position
The standard VSA superposition operator. The bundle is similar to all of its
inputs, which is what makes a composed task signature comparable to each
organ's composed capability signature.
mmlattice uses the same operator spatially: each of the 5000 cells is the
majority bundle of its neuron region.
--------------------------------------------------------------------------------
8.6 EVERY KERNEL EQUATION
--------------------------------------------------------------------------------
mmreshit Omega = argmin[ S + U - V ]
S surprise (hamming), U uncertainty (midpoint), V value (mmcert)
active inference: minimise surprise and uncertainty, maximise value
mmepona binary separable energy descent
belief converges toward certified (value), away from refused (loss)
parameters eta (rate) and lambda (regularisation)
mmstdp integer bit-shift exponential window
directed weighted graph from grounded co-activation
threshold w >= 200 before mmteach will assert it
mmact a* = argmax J
J = base + explore + theta - conflict - resource
+ entropy - repetition
the -repetition term is the anti-loop mechanism
mmevolve dx/dt = D + N + F + T
diffusion + emergence + force + transport
binary majority integration
mmlattice Pi : R^N -> R^5000
each cell = majority-bundle of its neuron region
the experiential coupling lattice X in Psi = f(A, X, S)
mmcog Q = coherence x integration x field-change x scale
Phi_pred = 1/(1 + dH) predictive coherence
Psi_self recursive self coherence
THE ORGAN ITSELF SAYS: "a COHERENCE diagnostic, NOT a consciousness
proof". Quote that whenever this metric is discussed.
mmloop choose -> evolve -> update -> wire -> measure -> reward
hill-climbs Q across 8 kernels
--------------------------------------------------------------------------------
8.7 BINARY END TO END, AND THE ONE EXCEPTION
--------------------------------------------------------------------------------
BINARY, all the way:
text -> mmsem -> 512 bit signature
signature -> mmbsig -> .bsig store, 64 bytes per record
query -> signature -> popcount scan
caps -> 2048 bit hypervector
task XOR capability -> popcount
majority bundle
STDP integer bit-shift weights
mmk/mmk2/mmk3 merkle roots, Ed25519 signatures
Every organ reporting "binary":true asserts this. All 24 Bubbles organs do.
mmorch additionally reports "no_float":true.
NOT BINARY, deliberately and only here:
mm-gateway on M4 a floating point language model, Qwen via MLX.
A downstream CONSUMER. It receives already retrieved,
already certified evidence and renders it into prose.
It never participates in retrieval or routing.
confidence_permille integers per thousand, not floats, precisely so they
cross machine boundaries exactly.
THE ARCHITECTURAL RULE: retrieval and decision are exact and reproducible.
Only the final rendering into human language is allowed to be probabilistic.
That is what "zero fabrication" in mmk2's caps actually means.
--------------------------------------------------------------------------------
8.8 CAPACITY
--------------------------------------------------------------------------------
K* ~= D / 43
At D=512 that is roughly 11 items bundled per signature before the majority
vote stops reliably recovering any single member. This is WHY passages are 40
to 400 characters: a passage must stay inside the capacity of one signature.
The frozen result of 2026-07-17: at D=1024 with L+S+R the system resolves to a
four outcome switch ANSWER, AMBIGUOUS, CONFLICT, ABSTAIN. Representation,
reasoning and routing were considered solved. Grounding remains evidence
bounded at roughly 49 percent certified.
WHAT MAY HONESTLY BE CLAIMED
measured retrieval, routing, relational reasoning, and safe refusal.
WHAT MAY NOT
universal truth, or zero hallucination as a general property.
--------------------------------------------------------------------------------
8.9 WHY THIS MATTERS PRACTICALLY
--------------------------------------------------------------------------------
1. To make PANTHEON find something you must win token_overlap. 96 points per
overlapping token beats almost any hamming improvement. Write caps in the
vocabulary of the questions you expect.
2. Passages must be short. Capacity is D/43.
3. Nothing needs retraining. Adding knowledge is appending 64 byte signatures
and rebuilding an index.
4. Two machines always agree. popcount is exact, which is what lets ORACLE on
megamind-agi trust a distance computed on Bubbles.
================================================================================
PART IX FILE MAP
================================================================================
--------------------------------------------------------------------------------
9.1 BUBBLES 100.90.97.104
--------------------------------------------------------------------------------
/home/user/mmk/ THE ACTIVE TOOLCHAIN AND ALL CORPORA
BINARIES
mmbsig ingest and recall. THE core tool.
usage: ingest <textlines> <bsig> <blob> <seen>
| recall <bsig> <blob> <query> | stress
mmwikix the recall organ. serve <port> | reindex
reads WX_BSIG, WX_BLOB, WX_IDX, WX_NAME from env
mmingest pushes into the M4 substrate. TARGET IS DEAD.
usage: <file> [max] [ip] [port]
mmreencode re-encode a store at a different dimension
hdc hyperdimensional computing helper
mmk mmk2 mmk3 mmcert mmbroad mmorch mmask mmreason mmdb
mmcog mmreshit mmepona mmstdp mmact mmevolve mmlattice mmloop mmteach
SCRIPTS
arxiv-crawl.sh gutenberg-crawl.sh wiki-harvest.sh wiki-crawl.sh
wiki-dump.sh curator.sh curator-db.sh teach-cycle.sh
DOCUMENT INGESTION, added 2026-07-28
mmdoc bash wrapper
mmdoc.js the tool. pdf, ocr, rtf, html, text.
MMDOC.md its documentation
CORPORA
gutenberg.bsig .blob .idx
wiki.bsig .blob .idx .seen
legacy.bsig .blob .idx .seen
arxiv.bsig .blob .idx .seen
live.bsig .blob .idx .seen
knowledge.bsig + evidence.blob + knowledge.idx + seen.k
STATE
darkstore.txt 290,592 claim lines
cursor 290,592
arxiv.state arxiv.setcursor arxiv.seen arxiv.log
gutenberg.cursor
corpus/federation.knowledge 24 triple lines
mmreason_corpus.tsv sources.tsv
baseline_seo.tsv codemind.tsv codex_lore.tsv gen2.txt
BACKUPS
backup-pre-desktop-ingest/ pre 2026-07-28 corpus snapshot
/etc/systemd/system/ UNIT FILES
mmk mmk2 mmk3 mmbroad mmcert mmorch mmdb mmcog mmreshit mmepona
mmstdp mmact mmevolve mmlattice mmloop mmteach mmask mmreason
mmwikix (wiki :9516) mmwikix-books (gutenberg :9517)
mmarxiv (:9521) mm-legacy (:9522) mm-live (:9523) mm-estate (:9525)
mm-legacy-reindex mm-live-reindex mm-estate-reindex
arxiv-crawl gutenberg-crawl forge
mm-heart mm-gate mm-eval mm-cap mm-harvest
mmcurator mmcurator-db mmteach-cycle
THE RECALL ORGAN UNIT PATTERN, copy this to make a new one:
Environment=WX_BSIG=/home/user/mmk/<name>.bsig
Environment=WX_BLOB=/home/user/mmk/<name>.blob
Environment=WX_IDX=/home/user/mmk/<name>.idx
Environment=WX_NAME=<name>
ExecStart=/home/user/mmk/mmwikix serve <port>
WorkingDirectory=/home/user/mmk
User=user
Restart=always
/var/www/ ~60 CLIENT VHOSTS. REVENUE HOST.
--------------------------------------------------------------------------------
9.2 MEGAMIND-AGI 100.69.120.61, inside WSL2
--------------------------------------------------------------------------------
/opt/mm/
bin/ 94 MB, ~200 binaries
oracle THE ORCHESTRATOR
oracle.bak.p1 .. p10 staged version ladder
oracle.bak.sefS0 .. sefS16 the Seferim capability ladder
oracle.bak.preAtlas .preP1fed .preP2 .preP3 .preTeach .wire1 .orch .exec
oracle_mt oracle.prefix5
mmcortex mmcortex2 .. mmcortex7
eye eye.bak.p28 .. p40 eye.bak.book1 .x1 .x2
loom loom.bak.book1 .x1 .x2
mmserve mmserve.phase4.contiguous mmserve.phase4.permuted_union
mmworld mmsem mmgame mmpredict mm-narrate mmshard
mmheart mmgate mmeval mmcap mmcapacity mmctl mmstatus mmboot mmdbg
reason reason3 reasoner reasoning reasond reason_v2 csreason recall2
hdc hdc2 hdcask hdcembed hdcgrow hdcrecall hdctrain
embgrow semgrow semgrow2 semgrow3 grow
selfimprove selfmodel introspect surprise psi realcog connectome
constitution safegate memberd mindd thoughtd swarmd
agent magent talk voice wave sdm stdp shard
answer ask code codegen extract_defs multi_extract datalog
NOTE the .bak.* names are a DELIBERATE VERSION LADDER, not clutter.
Each corresponds to a build stage. Do not delete without checking.
code/ 1.2 GB the code corpus and its five organs
data/ 1.9 GB substrate data
oracle/ 18 MB oracle.wal (10 MB, live), oracle.pcw, et_reason3*.txt
pyembed/ 211 MB embedded python runtime
src/ 17 MB C++ sources, Makefile, commit_s4.sh .. commit_s16.sh
THIS IS WHERE THE ORACLE CONTRACT FIX WOULD GO
logs/ 72 KB
backups/
/home/joseph/
mmk/ a SECOND toolchain copy
oracle/ oracle_serve binary and oracle_serve.cpp, et_reason3*.txt
mm-snapshots/ mm.tgz
--------------------------------------------------------------------------------
9.3 M4 AND VALKYRIE
--------------------------------------------------------------------------------
M4 node :8090 (mm-gateway, LIVE), :8096
megamind- :8091, :8092
Python :8082, :8899
NOT LISTENING :9440, :8099
VALKYRIE codex :8790
--------------------------------------------------------------------------------
9.4 DOCUMENTATION, your Mac
--------------------------------------------------------------------------------
~/Desktop/PANTHEON-COMPLETE.txt THIS FILE, the combined reference
~/Desktop/01-AI-Federation/ the same material in 13 focused files
00-MASTER-INDEX.txt start here, hand this to Claude
01-ARCHITECTURE-MAP.txt 02-DEVICE-MAP.txt 03-ORGAN-MAP.txt
04-ORGAN-DETAIL-BUBBLES.txt 05-ORGAN-DETAIL-MEGAMIND-AGI.txt
06-ORGAN-DETAIL-M4-VALKYRIE.txt 07-API-MAP.txt 08-DATA-FLOW.txt
09-MATHEMATICS.txt 10-CORPORA-MAP.txt 11-FILE-MAP.txt
12-KNOWN-ISSUES.txt
plus pre existing planning documents:
ATLAS_ARCHITECTURE.txt FEDERATION_BUILD_PROGRAM.md
FEDERATION-ARCHITECTURE-SCAN-2026-07-21.txt
MEGAMIND_AGI_BRAIN_ROADMAP.md MEGAMIND_RESUME_2026-06-25.md
CODEX_HANDOFF_depth.md Living_World_Build_Plan.txt
Living_World_Frontend_Plan.txt codex_handoff_code/
Those describe INTENT AND HISTORY. This document describes MEASURED REALITY
as of 2026-07-28. Where they disagree, this document is what the machines
actually report.
================================================================================
PART X OPERATIONS
================================================================================
--------------------------------------------------------------------------------
10.1 REACHING EACH MACHINE
--------------------------------------------------------------------------------
BUBBLES
ssh user@100.90.97.104
[REDACTED: credential-associated original line 2380; original retained outside this bundle.]
MEGAMIND-AGI, the awkward one
ssh lands in Windows cmd.exe, not Linux. Key auth fails with
"Too many authentication failures" because the agent offers too many keys.
[REDACTED: credential-associated original line 2386; original retained outside this bundle.]
[REDACTED: credential-associated original line 2387; original retained outside this bundle.]
"wsl -d Ubuntu -- bash -lc 'command'"
cmd.exe CONSUMES PIPES before WSL sees them. Anything with a pipe must be
base64 encoded:
B64=$(printf '%s' "$SCRIPT" | base64 | tr -d '\n')
ssh ... "wsl -d Ubuntu -- bash -lc \"echo $B64 | base64 -d | bash\""
Without this you get: "'grep' is not recognized as an internal or external
command".
[REDACTED: credential-associated original line 2399; original retained outside this bundle.]
[REDACTED: credential-associated original line 2400; original retained outside this bundle.]
VALKYRIE beandaddy@100.89.0.6
--------------------------------------------------------------------------------
10.2 TEACHING PANTHEON A DOCUMENT
--------------------------------------------------------------------------------
# dry run first, always
/home/user/mmk/mmdoc /path/to/docs --dry
# then for real, into the estate corpus
/home/user/mmk/mmdoc /path/to/docs --corpus knowledge
# verify
curl -s "http://127.0.0.1:9525/health"
curl -s "http://127.0.0.1:9525/recall?q=<something%20you%20know>&k=1"
For files on your Mac, run mmdoc there with --dry, scp the passage file to
Bubbles, then ingest:
mmbsig ingest /tmp/passages.txt knowledge.bsig evidence.blob seen.k
systemctl start mm-estate-reindex
systemctl restart mm-estate
--------------------------------------------------------------------------------
10.3 CREATING A NEW ORGAN
--------------------------------------------------------------------------------
1. build the corpus mmbsig ingest into <name>.bsig
2. build the index WX_BSIG=... WX_IDX=... mmwikix reindex
3. write the unit copy the pattern in 9.1
4. pick a free port ss -tln | grep :95
5. systemctl daemon-reload && systemctl enable --now mm-<name>
6. verify curl 127.0.0.1:<port>/health
7. REGISTER WITH ORACLE see 10.4. WITHOUT THIS IT IS INVISIBLE.
--------------------------------------------------------------------------------
10.4 REGISTERING WITH ORACLE
--------------------------------------------------------------------------------
curl -s "http://127.0.0.1:9402/oracle/register\
?name=<organ-name>\
&url=<url-encoded-base-url>\
&mode=remote\
&caps=<url-encoded-capability-string>"
Returns {"registered":{...},"persistent":true}
THE CAPS STRING IS THE ROUTING CONTROL SURFACE.
Write it in the vocabulary of the QUESTIONS you expect, not in the vocabulary
of how the organ works. Every overlapping token is worth 96 points.
GOOD "estate own projects clients missionMATCH ATLAS PANTHEON heritage
sgearbox manuals handover credentials my work my sites"
BAD "binary hamming signature store mmap popcount exact nearest neighbour"
(describes the mechanism, matches no real question)
Verify routing actually changed:
curl -s ".../oracle/orchestrate?q=<a%20real%20question>" | grep winner
--------------------------------------------------------------------------------
10.5 HEALTH CHECKS
--------------------------------------------------------------------------------
WHOLE FEDERATION
curl -s "http://100.69.120.61:9402/oracle/organs" | grep -o '"live":false'
counts dead organs
ONE ORGAN
curl -s "http://100.90.97.104:<port>/health"
ALL BUBBLES ORGANS AT ONCE
for p in $(ss -tln | grep -oE ':9[0-9]{3}' | tr -d ':' | sort -un); do
echo -n "$p "; curl -s -m3 "http://127.0.0.1:$p/health"; echo
done
ORACLE STATE
curl -s "http://100.69.120.61:9402/oracle/health"
--------------------------------------------------------------------------------
10.6 BACKUP AND ROLLBACK
--------------------------------------------------------------------------------
/home/user/mmk/backup-pre-desktop-ingest/
live.bsig, live.blob, live.idx, live.seen,
knowledge.bsig, evidence.blob as of pre 2026-07-28
THERE IS NO AUTOMATIC CORPUS BACKUP. gutenberg alone is 30 GB across bsig and
blob. Bubbles has 40 GB free. A naive full backup does not fit. Plan before
any bulk corpus operation.
TO ROLL BACK AN INGEST
systemctl stop mm-<name>
cp backup-pre-.../<name>.* /home/user/mmk/
systemctl start mm-<name>
curl 127.0.0.1:<port>/health confirm the corpus count dropped back
================================================================================
PART XI KNOWN ISSUES
================================================================================
ISSUE 1 ORACLE CANNOT AUTO CALL MOST REGISTERED ORGANS SEVERITY HIGH
/oracle/execute returns called:false reason:"no_contract" for most organs.
wikipedia-x called:true status 200 supervisor_pass:true reason ok
mm-estate called:false no_contract
mm-live called:false no_contract
mmbroad called:false no_contract
Not the URL path: registering with a bare base URL changed nothing.
Not simply builtin vs registered: mmbroad is builtin and still fails.
ORACLE holds compiled in execution contracts. Strings "contract_satisfied",
"contract_fingerprint" and "no_contract" appear in the binary.
THREE INDEPENDENT HARDCODED LISTS EXIST:
1. ORACLE execution contracts only some organs callable
2. mmbroad fan out 9 organs
3. mmorch knowledge list 6 organs
A newly registered organ passes ORACLE ROUTING and fails ALL THREE.
WORKAROUND read winner_url from /oracle/orchestrate and fetch it yourself,
or query the organ directly.
REAL FIX make the lists registry driven rather than compiled, so
registering an organ is sufficient. Requires a source change in
/opt/mm/src and a rebuild of a cortex running six days. NOT DONE.
ISSUE 2 TWO REGISTERED ORGANS ARE DEAD SEVERITY MEDIUM
megamind-m4 :9440 and world-brain :8099 on M4. Confirmed by lsof.
ORACLE already reports both live:false health:0.
WHY IT MATTERS MORE THAN A DEAD ENTRY: curator.sh PHASE A ingests into
100.94.8.94:9440 every 4 minutes. That is a SILENT DATA LOSS PATH. It has not
fired only because the darkstore is fully drained. The moment new claim lines
are appended, PHASE A writes into a void and advances the cursor anyway.
Also mmbroad lists megamind-m4 FIRST in its fan out and reports reached:false
on every single broad recall.
ISSUE 3 mm-estate NOT IN THE AGGREGATOR LISTS SEVERITY MEDIUM
Not in mmbroad's 9 organ fan out. Not in mmorch's 6 organ execute list.
Mitigated by also ingesting the same passages into live.bsig, which IS in the
fan out, at the cost of the two corpora overlapping.
ISSUE 4 CAPABILITY WORDING SILENTLY DECIDES ROUTING SEVERITY HIGH
96 points per overlapping token routinely decides the winner. mm-live held the
missionMATCH knowledge and scored token_overlap 0, so ORACLE routed to
mm-narrate and mmgame. Knowledge present, unreachable.
ISSUE 5 INGESTED KNOWLEDGE CAN BE INVISIBLE SEVERITY HIGH
Three conditions must all hold, four for aggregator reachability. See 7.5.
ISSUE 6 ORGANS BIND 0.0.0.0 WITH NO LOCAL FIREWALL SEVERITY LOW
Protected by something upstream, not by host configuration. Verified closed
from the public IP. If that upstream filter ever changes, ~205 million
passages including credential material become publicly queryable with no
authentication. Binding 127.0.0.1 or a ufw default deny would make the
protection intentional rather than incidental.
ISSUE 7 TWO mmk TOOLCHAINS DIVERGE SEVERITY LOW
/home/user/mmk on Bubbles (active), /home/joseph/mmk on megamind-agi,
/opt/mm/code on megamind-agi. Separate copies. mmdoc exists only on Bubbles.
ISSUE 8 DIAGRAM OCR IS INHERENTLY PARTIAL SEVERITY LOW
Mitigated by the prose gate in mmdoc. Expect ~70 percent fewer passages from
diagram images. That is the gate working, not data loss.
SUMMARY (status updated 2026-07-29; see the UPDATE block at the top of this file)
# ISSUE SEVERITY STATUS
1 ORACLE cannot auto call most organs HIGH FIXED — shapeReq/pathFor now
registry-driven (FIX 1). This
is also the drift correction.
2 megamind-m4 and world-brain dead MEDIUM FIXED — m4 revived; world-brain
started+re-registered (FIX 6);
silent-loss path closed; dead
organs health-gated (FIX 4)
3 mm-estate not in aggregator lists MEDIUM FIXED — mmbroad + mmorch now
fan out 10 organs (FIX 2, FIX 3)
4 caps wording silently decides routing HIGH MITIGATED — new organs given
question-vocabulary caps; dead
organs can no longer win (FIX 4)
5 ingested knowledge can be invisible HIGH FIXED — registration now suffices
for callability (FIX 1); the two
aggregators cover the corpora
6 organs bind 0.0.0.0 LOW open — network policy, by choice
7 two mmk toolchains diverge LOW open — housekeeping
8 diagram OCR partial LOW mitigated — working as designed
================================================================================
PART XII DATABASES
================================================================================
Full standalone version in ~/Desktop/01-AI-Federation/13-DATABASE-CENSUS.txt.
Scanned 2026-07-28. This part exists because the federation stores roughly 725
million records and only about 30 percent of them are reachable by a live organ.
--------------------------------------------------------------------------------
12.1 THE HEADLINE
--------------------------------------------------------------------------------
3 PostgreSQL clusters holding 19 databases, ~45 GB
418 SQLite files, of which ~110 hold real application data, ~940 MB
6 binary signature corpora on Bubbles, 204.9 million records, 42.2 GB
1 sharded binary corpus on M4, 374.5 million records, 184 GB, UNSERVED
4 BadgerDB key value stores, ~8.2 GB, two marked corrupt
1.6 TB on a removable volume that M4's PostgreSQL depends on to boot
TOTAL STORED RECORDS, all technologies ~725 million
RECORDS REACHABLE BY A LIVE ORGAN ~217 million 30 percent
RECORDS KNOWN TO mmdb 115,203 0.016 percent
A SCANNING LESSON WORTH KEEPING
The first pass used find -xdev, which does not cross filesystem boundaries,
and missed /Volumes/MEGAMIND on M4 entirely. That volume holds the single
largest database in the federation. Enumerate mount points before scanning.
--------------------------------------------------------------------------------
12.2 WHAT mmdb ACTUALLY HARVESTS
--------------------------------------------------------------------------------
mmdb :9506 is the organ whose entire job is pulling databases into the
substrate. From /sources:
SOURCE MACHINE COLUMNS ROWS SEEN
codex-lore VALKYRIE category, name, summary 73,698
seferim-events bubbles agent_name, item 35,338
baseline-seo bubbles url, title, meta_description, h1 6,138
strategymind VALKYRIE content, source 14
codemind VALKYRIE content, source 15
TOTAL 115,203
Every recent pass reports embedded_new 0 and deduped equal to rows_seen.
mmdb has fully absorbed these five and is idle. It is not failing. It has
nothing else registered to look at. Five sources out of roughly 130 real
databases.
--------------------------------------------------------------------------------
12.3 THE BIGGEST DATABASE IN THE FEDERATION IS UNSERVED
--------------------------------------------------------------------------------
/Volumes/MEGAMIND/wiki on M4, 184 GB, four shards.
SHARD passages.txt semsig.bin offs.bin RECORDS
shard0 10,027,787,806 2,996,462,528 749,115,632 93,639,454
shard1 7,038,365,018 2,996,462,528 749,115,632 93,639,454
shard2 6,958,251,791 2,996,462,528 749,115,632 93,639,454
shard3 7,037,104,645 2,996,462,432 749,115,608 93,639,451
-------------- ------------- ----------- -----------
31.07 GB 11.99 GB 3.00 GB 374,557,813
Records = offs.bin / 8 bytes per offset.
Signature width = semsig.bin / records = 32 bytes, so 256 BIT signatures.
The Bubbles corpora use 512 bit signatures. THE TWO FORMATS ARE NOT
INTERCHANGEABLE. See Part VIII for why the width matters to the maths.
SOURCE CORPORA ALSO PRESENT, ~99 GB
29 GB all374_passages.txt 24 GB enwiki.xml.bz2 24 GB so_passages.txt
11 GB quality_passages.txt 6.0 GB so_quality_passages.txt
5.2 GB enwiki_passages.txt 333 MB simplewiki.xml.bz2 120 MB passages.txt
NOTHING SERVES IT. Port 9440, registered with ORACLE as megamind-m4 with
caps "megamind deep knowledge recall wikipedia encyclopedia facts corpus
document memory million binary hamming nearest", is not listening. That
organ was the consumer. See Part XI issue 2.
M4's actual listeners, verified by lsof:
node *:8090 mm-gateway, the voice, LIVE and registered
node *:8096 Atlas Orchestrator conductor
node 127.0.0.1:8081
megamind-gateway 100.94.8.94:8091 not in ORACLE registry
megamind-gateway-staging 127.0.0.1:8092 not in ORACLE registry
external-federation 127.0.0.1:10099
ollama *:11434
postgres -D /Volumes/MEGAMIND/missionmatch-pg/data
megamind-fedbridge.py a Python process
SCALE
six Bubbles corpora, all served 204,938,736 records
this one M4 corpus, served by nothing 374,557,813 records
There are 1.8x more records unreachable on that volume than in the entire
live PANTHEON substrate.
--------------------------------------------------------------------------------
12.4 POSTGRESQL, THREE CLUSTERS, 19 DATABASES
--------------------------------------------------------------------------------
CLUSTER 1 megamind-agi PG 18, cluster "main", :5432, bind 127.0.0.1
DATABASE SIZE CONNS STATUS
megamind 26 GB 0 DORMANT, see 12.5
livingworld 2,462 MB 10 ACTIVE, the Living World
postgres 7,862 kB 1 cluster maintenance
CLUSTER 2 megamind-agi PG 18, cluster "missionmatch", :5433
bind 127.0.0.1 AND 100.77.114.40, tailnet exposed
missionmatch 14 MB 0 idle
missionmatch_staging 14 MB 0 idle
postgres 7,678 kB 1 cluster maintenance
CLUSTER 3 M4 postgresql@16 via homebrew, :5432
data dir /Volumes/MEGAMIND/missionmatch-pg/data, 16 GB
THE DATA DIRECTORY IS ON A REMOVABLE VOLUME, see 12.9
megamind 15 GB 0 DORMANT, see 12.5
agency_brain 335 MB 1 ACTIVE, this is ATLAS
missionmatch 95 MB 0 LARGEST missionmatch copy
missionmatch_restore_drill 73 MB 0 restore rehearsal
brain_staging 38 MB 0 idle
missionmatch_staging 35 MB 0 idle
tenant_w3_staging 11 MB 0 per tenant
tenant_alpha 10 MB 0 per tenant
tenant_beta 10,239 kB 0 per tenant
cp_staging 10,073 kB 0 idle
tenant_staging_demo 9,673 kB 0 per tenant
postgres 7,623 kB 1 cluster maintenance
LIVINGWORLD, the healthiest database in the federation, 10 connections
events 1,880 MB 10,319,394 nodes 8,568 kB 3,271
memories 436 MB 665,242 edges 2,944 kB 12,824
codex 59 MB 73,698 opinions 2,736 kB 7,469
chronicle 30 MB 146,853 relationships 2,392 kB 6,268
entities 26 MB 2,236 node_relhv 1,872 kB 2,569
powers 240 kB 597 places 136 kB 94
Ten connections means eye :9404 and loom :9410 are actively using it.
NOTE the codex table holds 73,698 rows, exactly the count mmdb reports for
codex-lore which it attributes to VALKYRIE. The same lore lives in this
table AND in VALKYRIE's endless.db. Which is authoritative is not recorded
anywhere and should be.
--------------------------------------------------------------------------------
12.5 THE TWO MEGAMIND DATABASES, 41 GB, ZERO CONNECTIONS
--------------------------------------------------------------------------------
megamind-agi:5432 megamind 26 GB M4:5432 megamind 15 GB
nbin 17 GB 32,404,944 knowledge 12 GB 22,623,502
knowledge 8,836 MB 24,364,578 nbin 3,262 MB 22,491,524
sembond 360 MB 3,418,102 sembond 39 MB 75,448
game_memory 48 kB - plus small gstage staging tables
THEY ARE NOT COPIES OF EACH OTHER. nbin differs by 9.9M rows, knowledge by
1.7M, sembond by 3.3M. pg_stat_user_tables on the megamind-agi copy shows
n_tup_ins 0 and n_tup_upd 0.
BUT ONE OF THEM HAS ALREADY BEEN EXPORTED TO BINARY.
/Volumes/MEGAMIND/m4brain/sigs512.bin is 1.5 GB.
1.5 GB / 64 bytes per signature = 24,364,578 records.
That is EXACTLY megamind-agi's knowledge table row count.
m4brain, 5.4 GB, is the finished binary export of the 26 GB database:
3.5 GB knowledge_full.tsv the passage text
1.5 GB sigs512.bin 24,364,578 x 512 bit signatures
186 MB offsets.bin
175 MB knowledge_export.tsv
42 MB learned.tsv
92 KB megamind_m4_idx compiled index builder
59 KB megamind_m4 compiled server binary
megamind_m4.cpp, megamind_m4_idx.cpp the source
The migration to binary succeeded. Its consumer, :9440, is dead. The server
binary and its C++ source sit right beside the data. Restoring that organ is
plausibly running one binary, not rebuilding a pipeline.
--------------------------------------------------------------------------------
12.6 MISSIONMATCH EXISTS FIVE TIMES
--------------------------------------------------------------------------------
M4:5432 missionmatch 95 MB <- largest, 0 conns
M4:5432 missionmatch_restore_drill 73 MB
M4:5432 missionmatch_staging 35 MB
megamind-agi:5433 missionmatch 14 MB
megamind-agi:5433 missionmatch_staging 14 MB
Three running services: mm-api on Bubbles, mm-api-staging on Bubbles,
mm-api on megamind-agi.
megamind-agi:5433 tables: audit_logs, auth_sessions, term_expansions,
opportunity_records, login_attempts, solicitation_extractions, users,
candidates, requisitions, sam_opportunities, applications, tenant_directory
The largest copy has zero connections; the smallest is on the tailnet
exposed port, which is what a live API would connect to. If mm-api serves
the 14 MB copy while 95 MB sits idle, the running product is missing data.
CONFIRM FROM THE SERVICE CONFIG.
--------------------------------------------------------------------------------
12.7 MADDIE CORRUPTED. 3.6 GB IN, 263 MB OUT, NEVER MERGED BACK.
--------------------------------------------------------------------------------
/Volumes/MEGAMIND/MADDIE/data on M4
3.6 GB maddie_corrupted.db the original, unreadable
263 MB maddie_recovered.db what came back out
6.2 MB maddie.db the current working database
128 KB w_know.bin the Hebbian weight matrix
4.0 KB brain.db and knowledge.db, both empty
224 B badger a BadgerDB directory
RECOVERY RATE 7.2 PERCENT BY BYTES.
RECOVERED CURRENT WORKING maddie.db
patterns 3,360,543 models 2,806
sources 485,193 tensors 38,147
tensors 346,069 sources 0
models 9,908
chunks, discovery_queue, discovered_models, legacy_models, legacy_tensors,
query_phi all 0
THE WORKING DATABASE HAS LESS DATA IN EVERY TABLE THAN THE RECOVERED FILE.
models 2,806 now vs 9,908 recovered losing 7,102
tensors 38,147 now vs 346,069 recovered losing 307,922
sources 0 now vs 485,193 recovered losing 485,193
Nearly half a million source URLs and 3.36 million patterns exist in
maddie_recovered.db and in no live system. This is the single most
recoverable data loss found anywhere in the scan and it needs a merge,
not new code.
--------------------------------------------------------------------------------
12.8 BADGERDB, A FOURTH DATABASE TECHNOLOGY, TWO STORES MARKED CORRUPT
--------------------------------------------------------------------------------
3.8 GB /Volumes/MEGAMIND/unified/data/badger the live one
4.3 GB /Volumes/MEGAMIND/tmp-corpus-badger "tmp" in the name
56 MB /Volumes/MEGAMIND/unified/data/badger_backup
224 B /Volumes/MEGAMIND/MADDIE/data/badger empty
PLUS:
unified/data/badger.corrupt-orig-20260624/ 128 MB 00001.mem
unified/data/badger_startup_artifacts_20260317_1401/ 22 MB 00007.mem
17 badger-named directories exist under unified/data.
A corruption event dated 2026-06-24, the same date carried by the maddie.db
timestamps in 12.7. Two different database technologies showing corruption
or sync artefacts on the same day is worth investigating.
unified/data is 76 GB: substrate.bin 2.7 GB plus several thousand JSON and
JSONL research artifacts (arc_benchmark v1 to v4, arc_qualia_telemetry,
antibiotic_resistance_*, api_mcp_deep_scan, autonomous_exploration). It is
an uncleaned research working directory, not a served database.
--------------------------------------------------------------------------------
12.9 THE MEGAMIND VOLUME, 5.5 TB ON M4, AND A SINGLE POINT OF FAILURE
--------------------------------------------------------------------------------
/dev/disk6s1 at /Volumes/MEGAMIND. 5.5 TB capacity, 1.6 TB used.
554 GB kimi_k25 Kimi K2.5 weights, 64 safetensors shards
485 GB tcb movie video, not federation data
184 GB wiki THE 374M RECORD CORPUS, 12.3
77 GB unified Go source + 76 GB data dir, 12.8
66 GB titan_learning bloom, deepseek-r1, deepseek-v3, kimi-k2,
kimi-k2-thinking, kimi-k2.5, ling-1t,
llama-405b, ring25 (160 shards)
66 GB backups
45 GB bubbles-backups the ONLY off-device backup found anywhere
38 GB conflict_learning sam-vit-huge, clip-vit-large, whisper-large-v3
29 GB m4-internal-archive
16 GB missionmatch-pg THE LIVE POSTGRESQL DATA DIRECTORY
12 GB _scan_scratch
5.4 GB m4brain 24.3M signature export, 12.5
4.3 GB tmp-corpus-badger
3.9 GB MADDIE 12.7
3.0 GB MediaBrain 2.7 GB data 1.1 GB hf_cache
512 MB megamind-train 68 MB rel_all.tsv 52 MB BATMAN
18 MB wikidx Go source for the sharded index
7.0 MB knowledge per model .bin signatures: BAAI_bge,
deepseek-coder, santacoder, DeepSeek-V2
0 B megamind-data empty
Also present, unsized: curiosity-crawler, DASHBOARD, hf_tmp_kimi, hf-deploy,
marketing, megamind-infer (Go source), queue, state, goal.go
~658 GB of the 1.6 TB is model weights, not databases. A further ~23 GB of
ollama models sits on M4's internal disk at /Users/joseph/.ollama.
THE SINGLE POINT OF FAILURE
M4's PostgreSQL runs with -D /Volumes/MEGAMIND/missionmatch-pg/data.
That includes agency_brain, which is ATLAS, the only database on M4 with a
live connection. If this removable volume unmounts or fails, PostgreSQL on
M4 stops and ATLAS loses its brain. No second copy of agency_brain was
found anywhere. The same volume holds bubbles-backups, the only off-device
backup in the federation, so one failure takes out both.
--------------------------------------------------------------------------------
12.10 SQLITE BY DEVICE
--------------------------------------------------------------------------------
BUBBLES 263 files, 174.5 MB. 58 real data (75.2 MB), 165 ops, 40 noise.
51.2 MB 2026-04-18 /home/user/tdg-pipeline/data/baseline.sqlite
6.7 MB 2026-01-11 complete-agi-core/seferim/federation/remote_memory.db
6.7 MB 2026-01-11 complete-agi-core/seferim/memory.db
6.6 MB 2026-01-11 complete-agi-core/seferim/federation/shared_memory.db
652 KB 2026-02-14 docker volumes designmind-data/_data/brain.db
496 KB 2026-06-24 /home/user/goatcounter/db/goatcounter.sqlite3
384 KB 2026-02-14 docker volumes financemind-data/_data/brain.db
384 KB 2026-02-14 docker volumes marketingmind-data/_data/brain.db
232 KB 2026-05-13 /opt/twhg-platform/db/twhg.db
180 KB 2026-07-27 /opt/groovesteps-accounts/data/accounts.db
172 KB 2026-03-30 /var/www/tcbcombatsports.com/portal/api/data/tcb.db
148 KB 2026-07-28 /var/lib/lead-ops/lead_ops.db
88 KB 2026-07-20 /opt/arcw-admin/data/arcw.db
80 KB 2026-05-05 /opt/handledtax-admin/data/handledtax.db
60 KB 2026-04-15 /home/user/uptime-kuma/db/kuma.db
48 KB 2026-06-24 /home/user/megamind/data/maddie.db
32 KB 2026-02-20 /home/user/rotator-site/backend/queue.db
32 KB 2026-07-26 /opt/homesongrand-admin/data/hog.db
24 KB 2026-06-15 /opt/fgtp-cms/data/cms.db
24 KB 2026-07-26 /opt/marshallese-blog/data/blog.db
4 KB 2026-07-28 /opt/sgearbox-panel/data/panel.db
MEGAMIND-AGI 31 files, 9.1 MB, NO real application data. Only snapd and
command-not-found. All real data is in PostgreSQL and /opt/mm
binary stores. The only device with a clean separation.
M4 60 files internal, 1.16 GB, ALL macOS system. On the volume only
4 sqlite outside the MADDIE set, of which 2 are Spotlight.
VALKYRIE 790.3 MB 2026-07-28 21:09 endless-table-server/data/endless.db
5.4 MB 2026-06-24 13:00 /Users/beandaddy/MADDIE/data/maddie.db
endless.db was written TODAY. It is the live Endless Table
database, almost certainly what codex :8790 serves, and the
largest single SQLite file in the federation.
--------------------------------------------------------------------------------
12.11 DUPLICATE AND VERSIONED DATABASES
--------------------------------------------------------------------------------
megamind x2 DIFFERENT CONTENT, 41 GB, 0 connections between them
missionmatch x5 largest copy idle
wiki x2 374,557,813 records at 256 bit on M4, UNSERVED
49,031,046 records at 512 bit on Bubbles, served :9516
the Bubbles corpus is 13 percent the size of the M4 one
maddie.db x5 6.2 MB M4, 5.4 MB VALKYRIE, 1.6 MB M4/M4-MEGAMIND,
48 KB Bubbles, 4 KB Bubbles, plus corrupted 3.6 GB and
recovered 263 MB
brain.db x9 3 real in docker volumes (652/384/384 KB), 4 stubs in
/home/user (16/16/16/4 KB), 2 empty on M4
seferim mem x3 6.7 / 6.7 / 6.6 MB, all 2026-01-11 within two seconds.
One dataset written three times, not three datasets.
badger x4 plus 17 badger-named directories, 2 marked corrupt
--------------------------------------------------------------------------------
12.12 ARE THEY BEING USED
--------------------------------------------------------------------------------
DEFINITELY IN USE
livingworld 10 conns; agency_brain 1 conn; endless.db written today;
all 6 Bubbles .bsig served and verified via /health; lead_ops.db and
panel.db written today; accounts.db, hog.db, blog.db, arcw.db, cms.db,
twhg.db, tcb.db all recent
HARVESTED INTO THE SUBSTRATE
the five mmdb sources, 115,203 rows
DORMANT
megamind 26 GB and 15 GB; all 5 missionmatch copies; brain_staging,
cp_staging and the 4 tenant_* databases; the seferim trio since 2026-01-11;
brain.db x9 since 2026-02; baseline.sqlite since 2026-04-18
STORED AND UNREACHABLE, the category that matters
wiki shards on M4 374,557,813 records, no server
m4brain sigs512.bin 24,364,578 records, no server
maddie_recovered.db 4,201,713 rows, never merged back
unified/data/badger 3.8 GB, no consumer identified
tmp-corpus-badger 4.3 GB, no consumer identified
NOT DATABASES, correctly excluded
kimi_k25, titan_learning, conflict_learning, hf_cache, ollama ~681 GB
tcb movie 485 GB
macOS system sqlite on M4 and VALKYRIE ~2.3 GB
--------------------------------------------------------------------------------
12.13 WHAT SHOULD HAPPEN, IN PRIORITY ORDER
--------------------------------------------------------------------------------
Findings, not actions taken. The scan changed nothing.
1. MERGE maddie_recovered.db INTO THE WORKING maddie.db. 485,193 sources,
3,360,543 patterns and 307,922 tensors exist only in the recovered file.
Cheapest recovery available, needs a merge not new code.
2. GET agency_brain OFF REMOVABLE STORAGE, OR BACK IT UP OFF IT. ATLAS is
the only live database on M4 and its data directory shares a volume with
the only off-device backup of Bubbles. One failure takes both.
3. DECIDE WHAT TO DO WITH 374 MILLION UNSERVED WIKI RECORDS. The compiled
server and its C++ source sit beside the data at m4brain/megamind_m4.
Restore :9440, re-ingest through mmbsig, or archive deliberately.
4. FIX THE SILENT WRITE PATH INTO A DEAD ORGAN. curator.sh PHASE A ingests
into 100.94.8.94:9440, which is not listening. The darkstore is drained
so nothing is lost today. The next appended claim line writes into a void
and the cursor still advances. Part XI issue 2.
5. CONFIRM WHICH missionmatch THE LIVE API SERVES.
6. REGISTER MORE SOURCES WITH mmdb. Candidates already structured and
relevant: livingworld.codex, livingworld.entities, livingworld.chronicle,
agency_brain, endless.db, maddie_recovered.models.
7. RESOLVE THE megamind PAIR. 41 GB, zero connections, one already exported
to binary. Compare sigs512.bin against the .bsig corpora, then archive.
8. LOOK AT WHAT HAPPENED ON 2026-06-24. A BadgerDB store and a MADDIE
database both carry corruption or sync artefacts from that date.
9. DELETE THE STUBS. Three 16 KB brain.db files shadow the real docker
copies. A 4 KB maddie.db shadows a 5.4 MB one. Keep one seferim memory db.
================================================================================
PART XIII ORCHESTRATION GAP
================================================================================
Federation-wide rescan 2026-07-29. Full standalone version in
~/Desktop/01-AI-Federation/14-ORCHESTRATION-GAP.txt.
This part exists because Parts I to XII were built from a four device view of a
seven device federation, and because three of their claims are wrong.
--------------------------------------------------------------------------------
13.1 THREE CORRECTIONS TO PARTS I THROUGH XII
--------------------------------------------------------------------------------
CORRECTION 1 PANTHEON IS ABOVE ORACLE, NOT EQUAL TO IT
Parts I to XII describe ORACLE as the top of the stack. The layering is:
PANTHEON the intelligence
|