Coherence Engine - Manthan continuous-learning plan

Coherence Engine - Manthan continuous-learning plan

Date: 2026-08-26

Status: FINAL_REVIEW_CANDIDATE_WITH_DEGRADED_CROSS_FAMILY_GATE - no adoption, work-assignment,
deployment, ticket-mutation or effect authority

Scope: Manthan-related activity needed to make Manthan a bounded exploration and evaluation
organ of the Coherence Learning Spine

Parent design: delivery/COHERENCE-LEARNING-SPINE-BASE-DESIGN-2026-08-24.md

Target experience: https://dev.conduir.ai/coherence-engine/wip on this host. Both the normal
hostname path and a forced local-service TLS/SNI path redirect unauthenticated requests to Microsoft
sign-in. Direct inspection of the deployed webchat artifact confirmed the Coherence WIP content;
the authenticated end-user journey was not exercised.

0. Decision in one sentence

Use Manthan as the option refinery and contained comparative-evaluation lane for Coherence - not
as the memory system, not as the causal evidence ledger, not as the adoption authority and not as a
direct actuator.

The next useful limit is not to connect more feeds. It is to prove this exact loop:

exact Coherence execution + measured outcome
  -> bounded hypothesis with incumbent/candidate digests
  -> contained Manthan trial with frozen evaluator and denominator
  -> signed/immutable result receipt, including NO_WINNER and degraded results
  -> GraduationCandidateV0 -> governed MoveProposalV0
  -> exact-subject review + independent evidence certification
  -> independent Operations decision + separate exact-subject effect grant
  -> packaged/published/deployed fresh Coherence successor and Role Fabric canary
  -> KEEP / REVERT / UNKNOWN
  -> only evaluated residue becomes Cortex-recallable learning

The central finding is therefore:

Manthan already supplies useful churn, lineage, quality-diversity, temporal holdout and selection

primitives. What it does not supply is the exact, tenant-scoped, causal and replay-safe bridge from

a Coherence execution to a controlled experiment and back to a measured successor. That bridge is

the Manthan-focused work. Building another generic learner or writing Manthan winners straight to

Cortex would preserve the current fragmentation.

1. Evidence boundary

1.1 Pinned sources inspected

This plan is a gap-and-delivery proposal, not a statement that the composed capability is complete.
The present-but-unaudited third source and the unexercised authenticated end-user journey remain
explicit evidence limits rather than guessed inputs.

1.2 Fresh live controls

The low-rate control ran before any availability statement:

• docker ps showed no Manthan containers;
• GET local-service/healthz refused the connection;
• /readyz, /gate/selection, /feed and /diversity therefore also refused the connection; and
• the manthan-moltbook-team.service evidence reviewed by the independent lane showed a failed unit.

The current claim is source available, local service unavailable. The 2026-08-13 report of a
partially running but unready stack is historical and must not be used as today's runtime state.

The Coherence WIP page is a separate control. The deployed webchat image at revision
359563067e11fc3f555f3f7766d1f42b49aba982 rewrites /coherence-engine/wip to the internally named
role-fabric-wip.html; the deployed file and source pin share SHA-256
6f8aa5b26738443c42131db5b79a9ecdfd385a7c6be72b1cc2472d07ceef7409. Its static Coherence content
is present. Its live read projection returns 503 not_configured because the webchat deployment has
no ROLE_FABRIC_URL; that failure means current run status is unavailable, not that the static page
or the underlying workflow is absent.

1.3 Five-dimension supply audit

2. What Manthan is good for

2.1 Candidate discovery and option diversity

Manthan's strongest current contribution is to stop the learning system collapsing to one familiar
answer:

• typed artefacts and explicit derives_from, contradicts, refines, combines and replaces

lineage preserve alternative and dissenting paths;

• MAP-Elites keeps a best candidate per kind/novelty cell rather than one global argmax;
• /diversity exposes occupied and empty cells and can separate runtime, Maestro, explorer and seed

provenance; and

• /feed blends fitness, graph signal and recency with an inspectable score breakdown.

These are valuable candidate-supply and search-prior mechanisms. They do not prove that a
candidate improves a Coherence workflow.

2.2 Falsifiable comparative evaluation

/ab/temporal ranks with pre-cutoff information and grades with later engagement. The provisional
/gate/selection pins a 10 percent relative-lift floor, minimum cohort of 30 and minimum 20 graded
edges; it reports INSUFFICIENT_DATA rather than forcing a verdict. Those are good fail-closed
behaviours.

The same implementation states its ceiling: the verdict is a top-k point estimate with no confidence
interval, and its median cutoff is recency-stratified. It requires randomized assignment and a
bootstrap confidence interval before it can certify. Future engagement is also a proxy for an idea
being built upon; it is not the Coherence mission outcome, evidence completeness, downstream escape
rate or effect-safety result.

2.3 Adversarial exploration, not governance

The correct use of Manthan is to generate and challenge bounded hypotheses. Manthan may say:

• this unexplored cell deserves a candidate;
• these alternatives dominate the current proxy surface;
• this candidate failed its declared experimental criterion;
• the data is insufficient; or
• no candidate won.

It may not say:

• replace the active Coherence workflow;
• promote this result into validated Cortex guidance;
• alter a role charter, policy, evaluator or protected path;
• grant work, deployment or external-effect authority; or
• treat Patina, popularity or future citations as causal effectiveness.

3. Bounded candidate gaps to re-audit and close

These are evidence-backed gaps in the inspected sources, not D-34-complete “exact gaps.” The separate
conduir-clean-room repository is present but was not audited for this plan, so each item must be
re-audited across all three required sources before it can be admitted as net-new capability work.

G-MAN-01 - runtime and operational admission

Manthan is not available on the current host. The latest main bounds failure duration and database
pool use, but it does not remove the confirmed 1536-dimensional vector materialisation on the
lightweight lineage path. The unmerged tenant-header fix and the failed sequential sensing unit must
also be reconciled against current main rather than merged by assumption.

Required outcome: a pinned deployed Manthan revision passes liveness, readiness and real data-path
canaries for generation, ingest, materialisation, feed, diversity, trial and result retrieval. An
availability badge cannot substitute for these paths.

G-MAN-02 - experiment identity and tenant membrane

The current Manthan store has no tenant/workspace partition, Coherence run/admission identity,
incumbent/candidate digest, evaluator identity, metric-set digest, evidence denominator or decision
grant. Provenance is inferred from writer-controlled tags and optional author identity.

Required outcome: every Coherence trial is addressed by immutable, tenant-scoped identities, and
all writes, queries, caches, uniqueness constraints and lineage edges enforce the same scope. A
missing or conflicting scope fails closed.

G-MAN-03 - durable ingress and replay

The current bridge uses core NATS subscription, a bounded process-memory queue, best-effort posts and
process-memory promotion deduplication. It can lose offline/overflow events and repeat promotions
after restart.

Required outcome: the originating record and durable outbox row co-commit in one transaction;
the consumer acks only after transactional persist; retries, quarantine, replay cursor and scoped
idempotency survive crashes at every producer/publish/consumer/ack boundary. Reconciliation proves
exact set equality after subtracting an explicitly enumerated allowed-extra set, with no unexpected
same-scope record and no cross-tenant collision. Trial lifecycle transitions use an expected-version
CAS; the exact START/STOP/terminal vocabulary remains a contract decision rather than an inferred
state machine.

G-MAN-04 - Coherence-specific experimental semantics

Sending a Coherence outcome through the existing bridge marks it via-maestro, while Manthan's gate
defaults to runtime provenance. More importantly, arbitrary outcome prose loses exact incumbent,
candidate, plan, admission, role/fence, evaluator and denominator identity.

Required outcome: a distinct coherence-trial cohort and an explicit adapter from service-neutral
experiment contracts to Manthan storage. The adapter retains complete structured identity; a text
rendering is a projection only.

G-MAN-05 - real evaluation and no-winner semantics

The present gate measures Manthan's own selection brain with held-out future engagement. It does not
compare two exact Coherence workflow versions on the same workload and evaluator.

Required outcome: paired or randomized incumbent/candidate evaluation with frozen workload,
evaluator, metrics and hard invariant floors; confidence/uncertainty; confounder reporting; and
PASS, FAIL, NO_WINNER, INSUFFICIENT_DATA, DEGRADED and CONTAINMENT_BREACH outcomes.

G-MAN-06 - decision, application and readback separation

A statistical winner is not an adoption decision, and an adoption decision is not proof the exact
successor was delivered or effective.

Required outcome: separate signed result, GraduationCandidateV0, governed MoveProposalV0,
exact-subject review, independent evidence certification, Operations decision, exact-subject effect
grant, successor materialisation, package, publish, deploy, application/readback, canary and
KEEP/REVERT/UNKNOWN records. Manthan returns evidence; the authorized decision actor decides;
Coherence compiles; Role Fabric/Temporal executes only under current admission, lease/fence and effect
authority; Cortex indexes only the evaluated residue allowed by policy. Terminal decisions use an
expected-version CAS and rollback increments a generation used by recall and later admission.

These are acceptance controls proposed by this plan. They do not inherit authority from
EXECUTION-REVIEW-CONTROL-PLANE.v0.md or the candidate team package while either remains unadopted;
each live decision/effect record must cite the then-current independently adopted governing source.

G-MAN-07 - learning consumer and memory closure

Yesterday's review showed the estate has several stores that are individually populated but are not
read before a later decision. Cortex has persistent episodic supply but almost no effective
crystallisation, and Swarm-to-Cortex capture is default-off.

Required outcome: name the exact later consumer for every Manthan-produced record. A promoted
learning must be retrieved under stable tenant/repository identity and proven present in the next
Coherence candidate or agent-context decision. Raw candidates, failures and dissent remain evidence,
not behavioural instructions.

4. Target composition and authority

5. Delivery plan

M0 - recover and admit Manthan as a dependency

Purpose: turn source supply into a proven, current service without yet connecting it to Coherence.

Activities:

1. Rebase the historical critical findings onto current Manthan main and pin the exact source,
image, compose manifest and configuration digest.
2. Bind M0 to a private, single-scope recovery environment. Require authentication on every route,
a non-empty admin token that fails closed when absent, and network-isolation proof before service
admission; do not expose or back up mixed-tenant production content in M0.
3. Repair the light lineage query with explicit columns plus a SQL has_embedding scalar; do not
defer the vector and then access it lazily.
4. Reconcile the still-open gateway tenant-header fix with current main and validate it against the
actual gateway contract.
5. Split sensing, harvest, Cortex projection and external sensors into independently observable,
retryable stages. One failed stage must not silently skip later stages.
6. Bring services/explorers/ and services/maestro-bridge/ into lint, tests and CI.
7. Repair the nightly Prometheus assertion. Restore a backup into a clean isolated instance, compare
exact record/key/digest sets, and rerun the served-path canary against the restored instance.
8. Pre-register a production-shape corpus digest, distribution/concurrency profile, percentile
latency/resource bounds and scale-negative cases before observing route results.
9. Add product-output signals: last generated, ingested, materialised, served, trialled and exported
record; queue lag; dropped/retried/dead-letter counts; and no-output alert.
10. Run a fresh SOURCE -> BUILT -> DEPLOYED -> EFFECTIVE canary against the exact deployed digest.

Exit evidence:

• source and deployed digests match;
• all required containers/services are present and ready;
• public/unauthenticated route access and empty/missing admin tokens are rejected before readiness is

accepted;

• a new bounded artefact is generated, persisted, linked, materialised and served;
• /feed, /diversity, /ab/temporal and /gate/selection meet predeclared percentile/resource

bounds on the pinned corpus and fail the declared scale negatives;

• deliberate substrate, gateway, NATS and Cortex failures turn the correct stage red without

manufacturing a successful loop; and

• restart and replay evidence is retained, and a clean restore reproduces the exact expected record,

key and digest sets and serves the canary successfully.

M1 - reconcile contracts and build the Coherence-Manthan adapter

Purpose: create a typed, no-effect bridge before adding a live learning loop.

Activities:

1. Reconcile the shipped Conduir experiment contracts and current Symposium consumers with the
Learning Spine's planned WorkflowExperimentV1 and ExperimentResultV1. Reuse the existing
intake, actuation-guard, graduation-candidate and MoveProposal mapping supply; do not duplicate it
or call it current Manthan ownership.
2. Require, at minimum:
- tenant, repository/project, actor/grant and sensitivity scope;
- hypothesis and falsification condition;
- incumbent and candidate artifact/content digests;
- Coherence blueprint, plan, admission, profile, catalogue/provider/tool snapshots;
- work/run/node/attempt, role/incarnation and fence identities;
- frozen workload/corpus, evaluator and metric-set digests;
- declared denominator and evidence references;
- budget, current lease/fence epoch, expiry, containment, rollback generation and references; and
- explicit no-winner/degraded states.
3. Add a coherence-trial provenance class that cannot collide with runtime, Maestro, explorer or
seed traffic.
4. Implement a no-effect adapter that writes immutable candidate/trial records and reads immutable
result receipts. It must not expose an apply operation or write SHADOW records into any globally
eligible Manthan signal feed.
5. Define canonical signed bytes, trusted issuer/key ID, algorithm, rotation/revocation policy and
negative verification for every result receipt.
6. Co-commit the originating record and outbox row in one transaction. Use ack-after-persist, replay,
dead letter, exact-set reconciliation and expected-version trial/decision transitions.
7. Add conformance, parity, tenant/repository-negative, unexpected-extra-event, duplicate,
out-of-order transition, producer/publish/consumer/ack crash-window, stale-digest,
altered-denominator, expired-authority and containment tests.

Exit evidence:

• schema and Python/TypeScript parity pass;
• exact request -> stored trial -> result round trip preserves every required digest;
• replay is idempotent by event ID and scoped subject, lifecycle transitions are version-guarded, and

reconciliation rejects unaccounted-for durable extras;

• cross-tenant and cross-repository reads/writes, mutable refs, invalid signatures, expired authority

and missing denominators are rejected;

• a result cannot encode or trigger an application; and
• Manthan unavailable yields a durable owed/failed projection, never a vanished request.

M2 - shadow discovery loop

Purpose: prove Manthan adds useful option diversity without changing production behaviour.

Activities:

1. Start from an E0 Learning Spine observation/candidate set, not raw un-attributed telemetry.
2. Project bounded candidates into the coherence-trial lineage space.
3. Use lineage, contradiction, MAP-Elites and diversity to generate or select alternative
hypotheses.
4. Compare Manthan suggestions with a simple declared baseline such as recency or existing
candidate-factory ordering.
5. Keep the path SHADOW: record what would have been selected but do not alter the active plan,
agent context, Role Fabric work or Cortex validated-learning tier. SHADOW records are ineligible
for /signal/feed, runtime generation, promotion and any other non-shadow selector.
6. Measure novelty, duplicate reduction, evaluator acceptance, hard-floor failures, cost and time.
7. Predeclare an accepted-novelty or cost-adjusted non-inferiority floor against the baseline; a
reproducible but useless/noisy ranking does not close M2.

Exit evidence:

• a frozen cohort produces reproducible incumbent and Manthan-assisted rankings;
• alternative/dissent paths are retained rather than collapsed;
• the treatment cannot see held-out outcomes;
• the run reports uncertainty and no winner honestly;
• every suggested candidate traces back to source observations and forward to its evaluation; and
• the predeclared novelty/non-inferiority floor passes, or the phase records NO_USEFUL_DIVERSITY

and remains open; and

• disabling Manthan reproduces the baseline with no production effect.

M3 - contained Coherence workflow trial

Purpose: make Manthan useful to the first real workflow-successor evaluation while preserving
the existing authority split.

Activities:

1. Materialise one immutable workflow successor from an independently accepted learning hypothesis.
2. Re-resolve current Coherence admission, role incarnation, lease/fence epoch and expiry immediately
before compilation and again before each contained or canary launch.
3. Compile incumbent and successor through Coherence using frozen snapshots.
4. Randomize assignment under a predeclared blocked/crossover design with provider/tool parity,
exchangeability checks, sample-size/power floor and confounder reporting. The randomized design
itself requires independent validity review before M3 may certify a result. Deterministic pairing
is not certification unless an independently reviewed design proves equivalent validity.
5. Execute in a no-effect or explicitly sandboxed Role Fabric lane.
6. Measure end-to-end quality, evidence completeness, queue time, execution time, rework, cost,
downstream escape and hard security/governance floors.
7. Have Manthan return the signed comparative result receipt; do not return a promotion command.
8. Convert an eligible result through GraduationCandidateV0 -> MoveProposalV0, then require
exact-subject correctness review and independent evidence certification.
9. Present the certified proposal, evidence and dissent to the authorized Operations decision inbox;
persist the decision with an expected-version CAS and a single terminal transition.
10. If adopted, require a separate fresh exact-subject effect grant, PACKAGE/PUBLISH/DEPLOY receipts,
current-authority revalidation, a fresh canary and application/readback before KEEP, REVERT or
UNKNOWN. If rejected, failed, no-winner or insufficient-data, close through a non-adoption branch
without inventing successor/canary evidence.
11. On rollback, advance a monotonic application/recall generation so concurrent or later recall and
admission reject pre-rollback guidance rather than relying on cache invalidation alone.
12. Index only the policy-allowed evaluated learning into Cortex and prove the next decision path
retrieves the same record under stable scope identity.

Exit evidence:

• same workload/evaluator and exact-digest arms;
• randomization/balance, provider/tool parity, sample-size/power, confidence bounds, confounder report

and executable NO_WINNER/INSUFFICIENT_DATA assertions;

• independent evaluator, exact-subject reviewer and evidence certifier separation;
• containment and effect-negative proof;
• GraduationCandidateV0 -> MoveProposalV0, decision CAS and exact scope;
• for adoption: separate current effect grant plus PACKAGE/PUBLISH/DEPLOY and successor

application/readback receipts;

• for non-adoption: a terminal FAIL/NO_WINNER/INSUFFICIENT_DATA/reject record with no canary required;
• rollback removes the successor from new runs, advances the recall/admission generation and rejects

stale concurrent repopulation;

• for adoption: delayed observation window and KEEP/REVERT/UNKNOWN; and
• complete backward and forward lineage.

The first honest claim after M3 is:

One exact Coherence workflow successor was comparatively evaluated through a contained Manthan

trial, explicitly decided, delivered to a fresh canary, measured and kept, reverted or left

unknown with complete lineage.

It is not yet “continuous learning.”

M4 - repeated governed learning cadence

Purpose: prove repeated compounding without turning repetition into uncontrolled autonomy.

Activities:

1. Schedule bounded observation/candidate/trial windows through Temporal while keeping ledger and
decision truth in their authorities.
2. Define immutable window identity, selection bounds, watermark, allowed lateness, finalization,
overlap, missed-window catch-up and duplicate-window suppression before scheduling.
3. Register the full path in the Learning Spine wiring registry with mode, producer, store,
consumer, decision, freshness, lag, replay cursor, rollback and executable canary.
4. Add oscillation, delayed-regression, distribution-shift and winner's-curse detection.
5. Calibrate Manthan's proxy measures against actual Coherence outcomes; retire measures that do not
predict value.
6. Permit bounded policy automation only for explicitly granted low-risk scopes after repeated
evidence. Keep evaluator, governance and protected-path floors immutable to this loop.
7. Surface loop health and decisions on the Coherence WIP page only from current runtime evidence.

Exit evidence:

• multiple exact-digest cycles across more than one finalized observation window, each bound to a

Temporal scheduler-dispatch receipt rather than a manual launch;

• repeatable improvement or honest reversion/unknown outcomes;
• no silent writer-only path;
• no cross-tenant, cross-project or stale-version learning leak;
• operator disengage, rollback and reapply drills; and
• Temporal restart/resume, missed-window catch-up, duplicate-window suppression and schedule-disable

alarm drills; and

• independent freshness recheck against the deployed digest.

Only after M4 may a bounded statement such as “selected Coherence workflow behaviour continuously
improves under the declared policy” be evaluated for adoption.

6. Role-neutral work packets

These packets describe work; they do not assign people, bind roles or grant effects.

7. Sabotage and falsification matrix

8. What to reuse, adapt and reject

Reuse

• Manthan artefact lineage, contradiction relations, MAP-Elites, diversity and inspectable feed;
• temporal holdout as an advisory regression/selection test;
• Conduir side-lane, experiment, objective-score and graduation contract work after reconciliation;
• Coherence exact plan/admission/template/provider identities;
• Role Fabric work/role/lease/fence/receipt truth;
• Evolution candidate, replay, evaluator, proposal and rollback mechanisms;
• Cortex scoped retrieval and bounded agent-context projection; and
• Grafosa as the derived lineage view.

Adapt

• Manthan storage for tenant-scoped coherence-trial records;
• the provisional gate into one evaluator inside a wider same-digest experiment receipt;
• provenance from inferred tags to authenticated structured producer identity;
• the Maestro bridge pattern into durable acknowledged ingress/outbox; and
• feed/diversity UI into a Coherence Operations projection driven by current path evidence.

Reject

• Patina or popularity as an automatic production reward;
• writing winners directly into Cortex behavioural guidance;
• using /gate/selection as a promotion API;
• treating NATS delivery, Temporal completion or an HTTP acknowledgement as learning truth;
• using mutable branches, latest, row counts or liveness as artifact/effectiveness evidence;
• adding more always-on generation before restoring observation and consumer paths; and
• duplicating Conduir experiment contracts under a Manthan-owned vocabulary.

9. Decisions required before M1 can become admitted work

1. Canonical service-neutral contract naming: reconcile current experiment v0 records with the
Learning Spine v1 names.
2. Initial tenant/storage strategy for Manthan experiments: isolated schema/database versus fully
tenant-keyed shared tables with FORCE RLS, a non-owner NOBYPASSRLS runtime role,
per-transaction tenant context and pooled-connection reset proof.
3. First Coherence workflow, frozen fixture, evaluator and hard floors for M3.
4. Which trial classes may enter Manthan and what content/sensitivity is forbidden.
5. Operations grant: who may request, stop, accept, reject or revert which trial scope.
6. Retention, deletion and Cortex-promotion policy for raw trials, dissent and evaluated residue.
7. Runtime recovery owner and the source/deployment path for current Manthan main.

Read-only contract reconciliation and an isolated synthetic M0 recovery can proceed before these
decisions. Authentication, private network exposure, fail-closed admin token and admissible
content/retention rules do block M0 service admission, any mixed-scope backup and any representation
that Manthan is tenant-safe or production-active.

10. Ultra-review and swarm status

10.1 Immutable Stage-1 result

The reviewed staged subject was:

• landing-zone HEAD 3c9716f0e7c83bb1e1ed70484d9405c822cb7a7d;
• staged blob 8ee96dae2458d27dcbc4388a404f8994deeedea0 (503 lines before this fold); and
• canonical full-index patch SHA-256

58de3a7d83a4e97accd723c651aff0f9b8647f683c0c007b0c67afe21592a4eb, reproduced with
git diff --cached --binary --full-index --no-ext-diff -- <path> | sha256sum.

Six isolated finder perspectives raised 34 raw candidates. Zero mechanically identical candidates
were removed and four semantic duplicate pairs were merged, leaving 30 unique candidates. Thirty
separate adversarial verifiers returned 27 CONFIRMED, one PLAUSIBLE and two REFUTED. The two
refutations were the demand for a specific Cortex retraction-marker implementation and an asserted
blanket Manthan same-commit tracker/README invariant. The plausible lifecycle-ordering item remains
an unresolved contract question; this revision requires version-guarded terminal transitions without
inventing START/CANCEL vocabulary.

10.2 Cross-family gate: incomplete, not a verdict

The current review engine was fetched and run at origin/main
4d4a3a53c912dbdd0a26d948c886aa7aa83b1fcb, zero commits behind. Every accepted dry run resolved
exactly one Anthropic, OpenAI and Google seat with status=ok. No live Stage-2 run produced three
usable family contributions:

1. the first leased pooled worktree disappeared during execution and returned status=degraded, no
findings and no contribution health;
2. a dedicated retry returned usable Anthropic/OpenAI findings but the Google lane timed out;
3. a bounded retry exposed a Google sandbox permission error because evidence roots were outside the
review cwd; and
4. a common-parent retry made both evidence roots visible but the Google lane still timed out.

DeepSeek lanes were rate-limited and Grok was non-responsive during fresh probes. The two-family raw
findings repeatedly confirmed the principal corrections folded above, but they are degraded triage,
not a three-family swarm verdict. Stage 3 was not run because healthy Stage 2 is its prerequisite.
Runtime receipts remain under .swarm-artifacts/coherence-manthan-ultra-review/; they are not delivery
evidence and must not be promoted by path or process existence.

10.3 Questions for the next healthy swarm

The next healthy three-family review should try to refute these propositions against the revised,
newly pinned artifact:

1. Is “option refinery plus contained evaluator” the smallest coherent Manthan role, or does current
source prove a stronger safe role?
2. Does the plan preserve the already-decided E0 boundary and avoid making live Manthan a hidden
prerequisite?
3. Are any existing Conduir, Maestro main, Maestro clean-room or Manthan contracts/runtimes being
duplicated or assigned to the wrong authority?
4. Can a Manthan result influence production without an exact Operations decision, application
receipt and canary outcome?
5. Are tenant, identity, event-replay, containment, evaluator-separation and no-winner controls
complete enough to prevent false learning?
6. Does every write name a later decision consumer and every claim have both backward and forward
lineage?
7. Do the phase gates prove source, built, deployed and effective state without reading a proxy as
the fact?
8. What is the smallest change that would falsify or materially improve this plan?

11. Claim boundary

This plan establishes a Manthan-focused delivery sequence and review target. It does not establish:

• current Manthan availability or readiness;
• a deployed Coherence-Manthan adapter;
• tenant-safe experiments;
• statistical certification by the existing gate;
• a closed Cortex knowledge path;
• an adopted Operations grant;
• a completed authenticated end-user WIP journey or working Role Fabric read projection;
• full D-34 source coverage; the local conduir-clean-room repository was present but not audited; or
• a healthy three-family Stage-2 or Stage-3 ultra-review verdict; or
• continuous learning in production.

The plan succeeds only when the evidence chain ends in a measured fresh-run result, not when the
document, schema, event, memory entry or dashboard exists.