Primer

Primer is the automated domain extension pipeline. It solves the problem of manually extending STRATT to new professional domains by creating a reproducible, gated 10-phase pipeline that generates 20-40 domain-specific units from a domain expert profile.

Problem & Vision

STRATT’s initial implementation covers 4 domains: dev, ops, docs, and shared. Manually extending STRATT to new domains (neuroscience, finance, law, nutrition, film, etc.) requires deep domain expertise and careful unit engineering.

Primer vision: Given a domain profile (30-60 minute questionnaire), automatically generate 20-40 production-ready units following the identical 10-phase pipeline, with human gates at every phase boundary.

Architecture

Core Components

Meta-domain: core (16 units)

Orchestrates the Primer pipeline itself
Contains the master chain: stratt://core/chain/domain-extension-pipeline@1.0.0

Meta-council: primer (3 agents)

ANALYST-01 — Phases 0-2 (discovery & mapping)
GENERATOR-02 — Phases 3-7 (unit generation & composition)
VALIDATOR-03 — Phases 8-9 (validation & RFC generation)

Pipeline: 10-phase chain with 9 internal gates

Each phase produces structured output (documented, versioned)
Each phase boundary has a gate requiring council approval
All phases respect SPEC-01 through SPEC-05

The 10 Phases

Phase 0: Domain Discovery

Agent: ANALYST-01 Input: Domain profile (questionnaire, research papers, workflow documentation) Output: Domain research summary

Extract key research areas and foundational papers
Identify interview subjects and domain experts
Map existing workflows and tooling
Document domain signal words and terminology

Gate: ANALYST-01 reviews discovery completeness

Phase 1: Workflow Mapping

Agent: ANALYST-01 Input: Domain research summary Output: Activity-to-unit-type mapping

Map 20-40 domain activities to STRATT unit types
Identify task boundaries (atomic workflows)
Identify chain composition points (sequences of tasks)
Identify role personas (domain experts)
Identify rules (compliance, quality gates, anti-patterns)
Identify supplies (templates, checklists, schemas)

Gate: ANALYST-01 reviews mapping accuracy

Phase 2: Tag Taxonomy Design

Agent: ANALYST-01 Input: Activity-to-unit-type mapping Output: Domain tag taxonomy

Design domain-specific tag vocabulary (signal words)
Define error classes and exception types
Define success criteria and quality gates
Map tags to trigger words for agent context detection

Gate: ANALYST-01 reviews taxonomy completeness

Phase 3: Rule Generation

Agent: GENERATOR-02 Input: Tag taxonomy Output: 5-10 domain-specific rules

Generate rules for compliance requirements
Generate rules for quality gates
Generate rules for anti-patterns (what NOT to do)
All rules follow SPEC-01 contract format
Example: stratt://neuro/rule/no-discussion-drafting@1.0.0 — prohibits writing Discussion section before Methods

Gate: GENERATOR-02 validates rules

Phase 4: Supply Generation

Agent: GENERATOR-02 Input: Tag taxonomy + rules Output: 4-8 domain templates and supplies

Generate schema templates (e.g., reagent audit schema)
Generate checklists (e.g., stats reporting checklist)
Generate style guides (e.g., journal style guide)
Generate prompt snippets for domain-specific roles
All supplies follow SPEC-01 contract format

Gate: GENERATOR-02 validates supplies

Phase 5: Task Generation

Agent: GENERATOR-02 Input: Workflow mapping + supplies Output: 10-30 domain-specific task units

Generate atomic workflow tasks (10-30 units)
Each task: single responsibility, typed contract, ~200 token prompt
All tasks follow SPEC-01 contract format
Example: stratt://neuro/task/extract-methods-from-paper@1.0.0 — isolate Methods section from PDF

Gate: GENERATOR-02 validates task quality

Phase 6: Chain Composition

Agent: GENERATOR-02 Input: Task units + rules Output: 3-6 domain-specific chains

Compose task units into higher-order workflows
Each chain: 5-15 steps, typed inputs/outputs
All chains follow SPEC-04 (gates embedded in chains)
Example: stratt://neuro/chain/manuscript-pipeline@1.0.0 — full editorial workflow

Gate: GENERATOR-02 validates chain logic

Phase 7: Role Definition

Agent: GENERATOR-02 Input: Domain profiles + supplies Output: 5-10 domain-specific roles

Define expert personas (author voices)
Map roles to chain steps (agent assignments)
Define role capabilities and expertise areas
All roles follow SPEC-01 contract format

Gate: GENERATOR-02 validates roles

Phase 8: Domain Validation

Agent: VALIDATOR-03 Input: All generated units (rules, supplies, tasks, chains, roles) Output: Validation report

Run stratt validate on all 20-40 units
Check fingerprints (SPEC-02 FM-01)
Check imports and DAG (SPEC-03 FM-03, FM-07)
Check contracts (SPEC-01 FM-06)
Check capabilities (SPEC-04 FM-09)
Fix any errors and regenerate units
All units must have valid fingerprints before proceeding

Gate: VALIDATOR-03 approves validation report (0 errors, all FM checks pass)

Phase 9: RFC Generation

Agent: VALIDATOR-03 Input: Validated units + domain research Output: RFC document + publication bundle

Generate RFC document (Request for Comments)
RFC summarizes all 20-40 units
RFC includes domain signal words for agent context detection
RFC includes integration examples (how to use the domain units)
RFC includes evolution roadmap (future extensions)
Archive RFC in domain history

Gate: VALIDATOR-03 approves RFC and publishes all units

Case Study: Neuro Domain (April 2026)

Status: Production (30 units, Dendrite council operational)

Questionnaire Input:

Domain: Biomedical neuroscience research
Research focus: Computational neuroscience, systems biology, confocal microscopy
Expert: Katy Hole, Francis Crick Institute
Timeline: 60 min questionnaire, 10 phase pipeline, 4 weeks to production

Generated Units (30 total):

Tasks (12):

extract-methods-from-paper
compare-protocols-across-studies
draft-figure-legend
format-references
draft-methods-section
draft-cover-letter
structure-reviewer-response
audit-reagent-usage
generate-analysis-script
summarise-paper
identify-confounds
structure-grant-section

Chains (4):

systematic-review (task orchestration for literature synthesis)
manuscript-pipeline (end-to-end editorial workflow)
grant-application (multi-phase funding proposal)
confocal-analysis (data processing + interpretation)

Rules (5):

no-discussion-drafting (disable Discussion before Methods complete)
cite-primary-sources (enforce primary literature)
stats-before-claims (statistics required before claims)
declare-ai-usage (mandatory AI disclosure)
no-data-fabrication (data integrity rules)

Supplies (4):

journal-style-guide (formatting templates)
stats-reporting-checklist (analysis validation)
PRISMA-flow-template (systematic review schema)
reagent-audit-schema (protocol tracking)

Roles (5):

as-lit-scout (literature discovery expert)
as-hostile-reviewer (critical peer reviewer persona)
as-methods-analyst (protocol & stats expert)
as-grant-strategist (funding proposal expert)
as-science-writer (clear communication expert)

Dendrite Council (6 agents)

The neuro domain has its own operational council with 6 specialized agents:

Agent	Focus	Responsibility
CORTEX-01	Executive function	Task planning & paper organization
RAPHE-02	Integration hub	Cross-study synthesis & evidence aggregation
FORNIX-03	Memory consolidation	Reagent/protocol tracking & audit
ARBOR-04	Knowledge tree	Taxonomy & reference management
PULVINAR-05	Sensory relay	Data input & experimental design
TECTUM-06	Reflex response	Rapid reviewer response & rebuttal drafting

New Capabilities

Primer added 2 new agent capabilities:

search — Domain literature search, protocol lookup
summarisation — Paper abstract generation, evidence synthesis

Integration with GRACE Specifications

SPEC-01 (Contracts): Each phase outputs units with typed contracts. Inputs and outputs validated per SPEC-01 schema.

SPEC-02 (Fingerprints): Phase 8 validates all generated units have correct Blake3 fingerprints (FM-01 check).

SPEC-03 (DAG): Phase 8 validates import graph has no cycles (FM-03 check), respects domain isolation rules.

SPEC-04 (Gates): Each phase boundary requires gate approval from Primer council. ANALYST-01 approves phases 0-2, GENERATOR-02 approves phases 3-7, VALIDATOR-03 approves phases 8-9.

SPEC-05 (Traces): Phase 9 generates execution trace summary, e.g., “extracted 47 papers → generated 30 units → 1 import error fixed → 25 units published”.

Future Domain Extensions

Primer template enables rapid extension to remaining 7 domains:

Domain	Focus	Target Units
Finance	Investment banking, portfolio analysis, risk management	25-35
Law	Legal research, contract analysis, litigation support	25-35
Nutrition	Dietary research, meal planning, clinical nutrition	20-30
Film	Screenwriting, production planning, post-production	20-30
Artist	Visual art theory, composition, critique	15-25
Neuro+	Expanded neuroscience (behavioral, clinical, neurotech)	30-40
Core/Shared	Foundational units used across all domains	10-20

Target state: All 11 domains populated with 250-350 total units, 9-11 operational councils, 50+ protected agents.

Execution Flow

1. Domain expert submits questionnaire
   ↓
2. ANALYST-01: Phase 0-2 (discovery, mapping, taxonomy)
   ↓ [Gate 1: ANALYST-01 approval]
3. GENERATOR-02: Phase 3-7 (rules, supplies, tasks, chains, roles)
   ↓ [Gate 2: GENERATOR-02 approval]
4. VALIDATOR-03: Phase 8 (CI validation, all FM checks)
   ↓ [Gate 3: VALIDATOR-03 approval]
5. VALIDATOR-03: Phase 9 (RFC generation, publishing)
   ↓ [Gate 4: Final VALIDATOR-03 approval]
6. Domain fully populated: 20-40 units live, council operational

Benefits

Reproducibility: Same pipeline for every domain
Auditability: Every gate produces audit records
Safety: All FM checks (SPEC-01 through SPEC-05) enforced before publication
Speed: 4 weeks from questionnaire to production (compared to 6 months manual)
Quality: Domain expertise codified into unit templates
Evolution: Domain units inherit future Primer improvements