Architecture paper · v2 · 2026

Protean & Galen.
Substrate and kernel.

Protean is a deterministic scientific runtime substrate — evidence, candidates, validators, scoring, Ledger proposal envelopes, contradiction and sequence-space artifacts, and a public/private disclosure boundary. Galen is a bounded scientific cognition kernel — typed capability composition, candidate adjudication, contradiction-priority routing, experiment prioritization, dry-lab handoff review, and OpenClaw-native orchestration.

The deployed path is explicit: Galen drafts proposals, an operator approves them, Bankr broadcasts, and the Protean Ledger on Base mainnet emits records and edges. The indexer replays those events into a reproducible Digest. The Explorer is a lens. The Verification Mirror is supplemental.

Open the Ledger Jump to figures

01—Current Ledger model

Chain first,
mirror second.

The earlier downloadable PDF predates the Protean Ledger migration and is no longer the current architecture reference. This web edition describes the deployed system: Base Mainnet, Protean Ledger, events, indexer, Digest, Explorer, and supplemental Verification Mirror artifacts.

Galen proposal

Operator approval

Bankr broadcast

Protean Ledger

02—Two layers

Substrate &
kernel.

The runtime has two layers separated by a typed interface. The substrate owns data, schemas, validators, scoring, and the envelope drafted for each cycle’s terminal Ledger records. The kernel owns cognition composition and review orchestration. Neither owns the other’s invariants, and the chain owns the canonical record.

Protean — substrate

Deterministic scientific runtime.

Cycle executor — bounded, locked, idempotent.
Evidence lake — deduplicated, content-addressed records.
Candidate generation — constraint-guided, embedding-seeded.
Deterministic validators — protease, structure, motif gates.
Scoring — seven canonical features plus bounded model-informed advisories.
Ledger proposal envelopes — operator-reviewed before Bankr broadcast.
Protean Ledger — typed records and typed lineage edges.
Contradiction artifacts — evidence and motif disagreement.
Sequence-space artifacts — cluster maps, motif families, gaps.
Disclosure machine + publication guard + Base mainnet events.

Galen — kernel

Bounded scientific cognition.

Capability composition — typed scheduling under policy.
Candidate adjudication — eight-stage dry-lab readiness path.
Contradiction routing — priority signal, not scoring input.
Experiment prioritization — evidence-linked, uncertainty-aware.
Dry-lab handoff review — drafts-only, review-gated.
Scientific review orchestration — disclosure-aware.
Replayable cognition traces — summarized into approved Ledger records.
OpenClaw-native — imported skills, scheduler, three surfaces.

Protean proposes — Galen evaluates, reviews, prioritises, orchestrates.

03—Figures

Ten companion
diagrams.

Each figure carries its paper section in the caption. They are reproduced inline so they can be linked, cited, and embedded without leaving the page.

Figure 01

Substrate and kernel

§3 · System Overview

Figure 02

Scientific capability execution

§6 · The ScientificCapability System

Figure 03

Candidate adjudication

§7 · Candidate Adjudication Architecture

Figure 04

Assay-preparation handoff review

§8 · Wet-Lab Handoff Workflows

Figure 05

Replayable cognition

§9 · Replayable Cognition

Figure 06

Contradiction-priority routing

§4.8, §5.6 · Contradiction artifacts

Figure 07

Operational capability tiers

§10 · Operational Capability Tiers

Figure 08

OpenClaw orchestration

§11 · OpenClaw Integration

Figure 10

Structured scientific memory

§9.3 · Memory traces · App. B

Figure 09

Disclosure state machine

§12 · Public/Private Boundary

04—Capability system

Every cognition action,
typed.

The ScientificCapability schema is the typed interface between the substrate and the kernel. Composition is by identifier, not by free-form invocation. Mismatches are refused at schedule time. Declared vs. actual is what makes cognition auditable before any proposal reaches the Ledger.

identifier

stable name

inputs

typed contract

outputs

typed contract

replay_requirements

min state to re-derive

provenance_writes

permitted lineage edges

contradiction_reads

graph projections

memory_permissions

read · write tiers

telemetry_requirements

structured emit set

operational_tier

0–3 (§10)

governance_policy

policy version

05—Candidate adjudication

Eight passes,
one judgement.

Adjudication composes retrieval, contradiction, failure, novelty, and developability signals into a single defensible scientific judgement under review. The output is advisory — human review is the authority.

01 retrieval audit
trust state · retraction · current-source check
02 contradiction analysis
active · unresolved · strength
03 failure-memory review
motif resemblance · neighbourhood
04 novelty analysis
embedding novelty · motif novelty (kept separate)
05 developability review
hydrophobic moment · TOP-IDP · Boman · entropy
06 experiment prioritization
ranked plan · typed rationale
07 dry-lab readiness
five-vector: synthesis · coverage · power · trust · contradiction
08 handoff recommendation
include · defer · exclude · escalate

06—Assay-preparation handoff

Drafts only.
Review-gated.

The runtime drafts assay requests; it does not submit them. It prepares review batches and operator-reviewed provider packets; the operator review gate is the only path to external action.

step 01

Composed from adjudication digests, cycle priority signal, operational constraints.

step 02

Review batch is a typed object — candidates, rationale, panel, uncertainty.

step 03

Provider packet draft — operator-reviewed.

step 04

Handoff trace — replayable, evidence-linked, policy-versioned.

step 05

Operator review gate — signed receipt advances; absence holds.

step 06

Provider-side execution lies outside the runtime — never auto-submitted.

07—Verifiable cognition

Cognition,
through the Ledger.

Capability invocations, memory deltas, contradiction reads, and review-gate events are summarized into typed records on the Protean Ledger. A reviewer starts from the chain, walks the event-derived lineage, reproduces the Digest, and then checks any supplemental replay artifact when one exists.

Replay validates integrity, not biological activity. A passing verification confirms that Base mainnet events, indexer state, Digest, lineage, and any supplemental artifact agree. It does not confirm that the candidate works.

08—Operational tiers

Four tiers.
One escalation path.

Capability tiers give cognition actions explicit, policy-visible autonomy budgets. A capability never escalates itself; the kernel refuses at schedule time.

tier 0 · observation

Read-only. Cognition trace writes.

tier 1 · private artifact write

Typed scope under the private vault.

tier 2 · public artifact write

Subject to the publication guard at every site.

tier 3 · review-gated escalation

Approval receipt required before action.

09—Implementation status

What is
shipped,
partial, reserved.

Each row maps to a paper section. The runtime publishes its own implementation status alongside the paper — including the things that aren’t built yet.

Implementation status

What is, what is not — said out loud.

paper §13 · table 1

Subsystem	Status	Boundary
Protean Ledger on Base mainnet	implemented	UUPS proxy 0xE3c261F3…94cf5f0 · protean.ledger.v1 · publication attestation schema · 17 RecordTypes · 20 RelationTypes.
Public indexer + state digest	implemented	Open-source index_ledger_from_genesis.py · reproducible sha256 at /ledger/api/v1/indexer/digest.
Typed lineage edges	implemented	20 RelationTypes as first-class on-chain state · candidate, family, variant, and publication lineage.
Published candidate attestations	implemented	Full sequence + sequence hash + salted commitment + operator attribution · salt private.
Evidence lake	implemented	Deduplicated + content-addressed · trust-tier metadata.
Deterministic validators	implemented	Four-enzyme protease panel · structural + motif gates · claim QA.
Scoring with readiness gates	implemented	Evidence, failure distance, biophysical gates, advisory model signals · weight-bound verified.
Contradiction artifacts	partial	Typed substrate output · kernel consumes at adjudication · back-edge reserved.
Sequence-space artifacts	implemented	Embedding-derived clusters · motif families · unexplored regions.
Disclosure state machine	implemented	PrivateCommitmentOnly · RedactedPublic · CounselReviewed · PatentPending · Public · Retracted.
Redaction guard	implemented	Fails closed at every public-write site · CI leak scanner second pass.
Protean Ledger write path	implemented	Galen proposal · operator approval · Bankr broadcast · explicit RBAC.
ScientificCapability schema	implemented	Typed inputs · outputs · replay · provenance · tier · governance pinned.
Capability scheduling	implemented	DAG verified before execution · type-mismatch refused at schedule time.
Candidate adjudication	implemented	Eight-stage path · five-vector dry-lab readiness · advisory recommendation.
Contradiction routing	implemented	Influences priority, not scoring · re-derivable from substrate state.
Verifiable cognition trace	implemented	Capability invocations + memory deltas + reviews summarized into Ledger records and supplemental artifacts.
Operational capability tiers	implemented	Four tiers · declared in schema · enforced at schedule time.
Assay-preparation handoff	review-only	Drafts-only · 15 remaining lifecycle states reserved as substrate types.
Per-step DAG enforcement	flag-gated	Verified before execution · per-step enforcement inside the executor pending.
OpenClaw orchestration	implemented	Imported skills · scheduler, telemetry, review surfaces · capability-aware.
Approval-receipt verification	partial	HMAC infrastructure present · idempotent-action consumer not yet wired.
External provider escalation	partial	Local providers allowed unconditionally · remote escalation review-gated.
Cohort assay batches + result ingest	reserved	Substrate types declared · executor + reviewed-learning export pending.

10—Reproducibility

Every claim,
re-checkable.

A reviewer can read every record from the Protean Ledger, walk every lineage edge as typed on-chain state, and run the open-source indexer against any Base RPC from genesis. The state digest the indexer returns must equal the digest at /ledger/api/v1/indexer/digest.

Schema

protean.ledger.v1

Content digest

contract keccak256 over ABI-encoded fields

Chain

Base mainnet (8453)

Ledger proxy

0xE3c261F3…94cf5f0

Public indexer

scripts/index_ledger_from_genesis.py

Verification digest

/ledger/api/v1/indexer/digest

Open the live provenance graph

Citation

How to cite.

@misc{protean_galen_v2_2026,
  title  = {Protean and Galen: A Deterministic Scientific Runtime Substrate
            and a Bounded Scientific Cognition Kernel on the Protean Ledger},
  author = {Protean Labs},
  year   = {2026},
  url    = {https://www.protean.sh/paper},
  note   = {Computational rankings are research prioritization signals,
            not biological proof.}
}

Boundary

Computational rankings are research prioritization signals. They do not prove biological activity, therapeutic effect, safety, efficacy, patentability, clinical readiness, or experimental validation. Wet-lab validation and human scientific review remain the authoritative downstream layer.

Protean & Galen.Substrate and kernel.

Chain first,mirror second.

Substrate &kernel.