Manthan — the agents' home (scope 6, scratchpad, 2026-08-13)

Manthan — the agents' home (scope 6, scratchpad, 2026-08-13)

Scan only, verified live. manthan, 4 containers running.

Operator's framing: "the home of agents, where agents live, sleep, talk, explore." The README says
the same in its own words:

Sanskrit मंथन — "churning." An independent, always-on learning & exploration platform where

Conduir agents live, talk, prototype and evolve ideas. Conduir and other internal systems consume

Manthan's outputs to learn and improve.

*Samudra Manthan — … Halahala (poison) surfaces first; then the 14 ratnas; finally amrita. Bad

ideas come early. The rare breakthrough comes late. You keep churning.*

1. The API is exactly the metaphor — 20 routes

manthan-api-1, container port 8000 → host :47002.

Two things stand out:

• patina — per-agent accrued weathering/experience, with a leaderboard. This is the same

concept as the (empty) agent_effectiveness table in the conduir flywheel ledger, but modelled as
a first-class product surface here.

• /lineage/graph/grafosa — grafosa is already the lineage projection in this world. Directly

relevant to the agent-fork lineage question: the graph-as-derived-lineage-view pattern is not
hypothetical, it exists here.

2. Live state — the lights are on, nobody is home

manthan-api-1        Up 43h   healthz 200  {"status":"ok","service":"manthan-api","version":"0.2.0"}
manthan-substrate-1  Up 17h   docker health: UNHEALTHY
manthan-postgres-1   Up 43h   (pgvector)
manthan-nats-1       Up 43h

/healthz                200   ok
/readyz                 200   {"ready": FALSE,
                               "substrate": "unreachable:ReadTimeout",
                               "herald[9120]": "unreachable"}
/agents                 000   timeout
/elites                 000   timeout
/diversity              000   timeout
/patina/leaderboard     000   timeout
/artefacts/recent       000   timeout
/prototypes/list        000   timeout
/opportunities/current  000   timeout
/herald/status          200   node_count: 3, reachable_count: 0
/docs                   200

Every population, exploration and evolution route times out. The process is alive and its
liveness probe is green; nothing behind it answers.

/readyz is honest — ready: false with named failing checks. That distinction (liveness green,
readiness false, data path dead) is the same trap found elsewhere in the estate today, and here at
least the readiness probe tells the truth.

Herald — the "talk" layer is fully down

alpha  http://host.docker.internal:9120  ConnectError: All connection attempts failed
beta   http://host.docker.internal:9121  ConnectError
gamma  http://host.docker.internal:9122  ConnectError

Not a DNS problem — host.docker.internal resolves fine inside the container (172.17.0.1). I
checked, expecting the usual Linux gotcha; it is not that. Host ports 9120/9121/9122 simply have
nothing listening. The three herald nodes were never started.

Substrate — healthy-by-endpoint, unhealthy-by-docker, unreachable-by-caller

docker inspect manthan-substrate-1 -> health: unhealthy
curl :47000/healthz                -> 200 (fast); its logs show healthz served OK repeatedly
manthan-api readyz                 -> substrate "unreachable:ReadTimeout"

Three different answers about one service. The consistent explanation is the project's own
BUG-001: data paths eagerly load a 1536-dim embedding per node, so /healthz returns instantly
while any real query hangs past the caller's timeout.

3. The project documents its own state honestly

docs/KNOWN-ISSUES.md — 2 × P0, 2 × P1, all open, and my probes reproduce them:

Both ends of the loop are broken: BUG-002 kills the ingest (world-sensing → nothing new to
churn), BUG-001 kills the egress (/feed → Conduir cannot consume what was churned). A churning
platform with neither intake nor outflow.

4. How this lands against the rest of the estate

• Conduir's documented consumer path from Manthan is /feed, and it is down. So the

"Conduir consumes Manthan's outputs to learn and improve" link is not merely unexercised — it is
actively broken at the far end, independently of anything on the Conduir side.

• patina vs agent_effectiveness: two worlds independently modelled per-agent accrued

performance. Manthan exposes it as a product surface (leaderboard); Conduir has the table and it
is empty. If agent measurement is wanted, these should not be built twice.

• /lineage/graph/grafosa is precedent for the fork-lineage decision: grafosa as a *derived

graph view* over an authoritative store, which is the option I recommended in the fork scan.

• Same estate-wide pattern, fourth instance: the measurement surfaces exist and carry no data.

5. Not established

• No write path was exercised; nothing was started, stopped, or reconfigured.
• BUG-003 (CI) not verified.
• Whether /feed ever worked, and what the 3 herald nodes are meant to be, was not investigated.
• The manthan-postgres and manthan-nats contents were not inspected — there may be churned

artefacts sitting in the database that the API simply cannot serve.

• Old-world/lineage comparison not done.