Strategy Engine · Architecture Map · 2026-05-12

B now. C only after evidence.

A full architectural map of the Strategy Engine decisions: doctrine and accepted knowledge, measurement/signal spine, Tier 3 admission, Pearl causality, privacy/evidence-use governance, human-review redundancy, output traceability, layered evals, and every B→C development pathway.

Status: pre-implementation architectureExternal review: Opus + Pro challenge pendingRule: no broad implementation before review

Tier 1

Doctrine / Accepted Knowledge

Corpus-backed claim/evidence support. Durable doctrine, accepted knowledge, mechanisms, political science, history, institutions, and causal methods.

Tier 2

Measurement / Signal Spine

Live and structured feeds interpreted as measurement vectors. Signals are dynamic context, not knowledge, unless promoted through review gates.

Tier 3

Decision Synthesis

Strictly gated synthesis using accepted knowledge plus admitted signal context. It cannot silently upgrade evidence status or causal strength.

System map

[Raw sources]
  official / academic / polling / media / social / CRM / field / Hansard / legal / campaign
       |
       v
[Immutable raw store]
  raw record + hash + source metadata + parser version + privacy/evidence-use status
       |
       v
[Interpretation layer]
  entity | geography | taxonomy | sentiment/belief | causal claim | integrity risk
       |
       v
[Measurement object / signal vector]
  source, time, geography, entity, issue, type, confidence, provenance, allowed use, causal status
       |
       +--------------------------+
       |                          |
       v                          v
[Signal lifecycle ledger]      [Doctrine / claim support]
  new / active / reinforced      accepted corpus, claim/evidence contracts,
  candidate / degraded /         source roles, candidate knowledge
  contradicted / archived
       |                          |
       +------------+-------------+
                    v
[Tier 3 Admission Gate]
  admit | admit_with_limits | quarantine | reject | human_review
                    |
                    v
[Decision indicator bundle]
  admitted signals + accepted knowledge + rejected/quarantined list + limits
                    |
                    v
[Structured output contract]
  recommendation + evidence map + Pearl status + assumptions + risks/falsifiers + review requirements
                    |
                    v
[Layered evaluation]
  claim/evidence | retrieval | signal interpretation | Pearl | admission | output trace | human calibration

Hard gates

Tier 3 Admission Gate

Every input is admitted, admitted with limits, quarantined, rejected, or sent to human review. Nothing enters synthesis merely because it exists.

Pearl causal gate

Causal/projection claims require A/B matrices, pathway, temporal order, confounders, proxy risk, directionality, identification status, and review where high-impact.

Human-review gate

Mandatory for accepted-knowledge promotion, projection input, high-impact causal recommendations, actor-sensitive claims, Flash alerts, and public outputs about people/groups/orgs.

Privacy/evidence-use gate

Separates public citation, public context, private context, confidential strategy, privileged/sensitive material, and do-not-use material.

Twenty architecture decisions

#	Element	B now	C later
1	Three-tier architecture	Separate doctrine, signal spine, decision synthesis	Fuller integrated Decision Engine after gates
2	Tier 1 doctrine	Claim/evidence support engine	Formal knowledge graph / ontology
3	Tier 2 measurement	Narrow measurement/signal spine	Broader live intelligence platform
4	Signal lifecycle	Append-only promotion/degradation ledger	Truth-maintenance / belief-revision engine
5	Human review	Explicit review records	Workflow-backed review governance
6	Continuity governance	Yaron as continuity reviewer + competency protocol	Scoped reviewer dashboards/workflows
7	Entity/actor model	Canonical entity registry + temporal roles	Actor/influence/network graph
8	Taxonomy/adjacency	SKOS taxonomy + lightweight edge table	GraphRAG / temporal KG
9	Belief/sentiment	Separate belief and sentiment objects	Narrative dynamics engine
10	Geography	Weighted certainty across granularities	Geospatial/electoral modelling platform
11	R/statistics	Descriptive trend, polling, change-point smoke	Expanded causal/statistical suite
12	NLP interpretation	Target-specific entity/sentiment/frame/causal NLP	Richer interpreters and narrative systems
13	Decision integration	Separated indicator bundle	Full query planner/simulation/playbooks
14	Tier 3 admission	Formal admit/quarantine/reject/review gate	Decision-workbench controls
15	Storage/audit	Local immutable raw store + DuckDB + ledgers	Cloud/distributed/graph-scale infrastructure
16	Source reliability	Reliability matrix + allowed/forbidden uses	Adaptive source reliability model
17	Privacy/evidence use	Boundary layer; governance rules first	Access-control/compliance system
18	Output contract	Structured recommendation + reasoning trace	Interactive decision workbench
19	Evaluation	Layered canaries and quality gates	Continuous eval platform
20	Implementation	Sequential tracer-bullet build	Broader platform after evidence

B→C development pathways

C transitions are not automatic. Albert raises them only when primary triggers exist, hard blockers are absent, and Doron approves.

T1 Claim support → Knowledge graph

B: claim/evidence engine. C: formal KG/ontology. Trigger: repeated stable claim artifacts and graph queries that reduce ambiguity.

T2 Signal spine → Live intelligence platform

B: narrow measurement vectors. C: richer current-state platform. Trigger: signal canaries improve decisions without false confidence.

T3 Lifecycle ledger → Truth-maintenance

B: event ledger. C: support/contradiction networks and belief revision. Trigger: contradictions/staleness overwhelm manual review.

T4 Entity registry → Actor network

B: canonical entities and temporal roles. C: actor/influence graph. Trigger: reviewed actor relationships repeatedly drive strategy.

T5 Taxonomy → GraphRAG/KG

B: SKOS and DuckDB edges. C: graph database / temporal KG. Trigger: edge traversal solves measured retrieval or reasoning failures.

T6 Belief map → Narrative dynamics

B: belief/sentiment objects. C: narrative movement and response strategy. Trigger: measurable belief edges repeatedly affect decisions.

T7 Simple R stats → Causal suite

B: descriptive and change-point statistics. C: BSTS, synthetic control, DiD, GRF, DML, MRP. Trigger: data and Pearl identification justify method.

T8 Alerts → Response system

B: Watch/Signal/Flash as reviewed signal objects. C: operational playbooks. Trigger: alerts prove precise, useful, and reviewable.

T9 Indicator bundle → Full Decision Engine

B: separated context bundle. C: integrated query planner. Trigger: Tier 1/Tier 2 reliability and Tier 3 status preservation are proven.

T10 Manual review → Workflow governance

B: review packets/records. C: queues, assignments, deadlines, scoped reviewer UI. Trigger: review volume or continuity needs create operational risk.

T11 Local store → Cloud/graph infra

B: local DuckDB/Parquet. C: Iceberg/Delta/warehouse/graph. Trigger: local-first becomes performance, concurrency, or governance bottleneck.

T12 Weighted geography → Geospatial platform

B: weighted jurisdiction certainty. C: spatial/electoral modelling. Trigger: finer geography improves measured decisions safely.

T13 Integrity flags → Manipulation intelligence

B: artificial-signal flags and retrospective rescans. C: coordinated manipulation intelligence. Trigger: detection repeatedly protects data quality and review confirms accuracy.

B implementation as a project

Pre-implementation rule: no broad implementation starts until Opus and Pro reviews are complete, must-fix objections are reviewed with Doron, privacy/evidence-use governance v0 is specified, the first tracer bullet is narrowed, and kill switches/rollback are named.

Success conditions

Traceable recommendations.
No silent evidence-status upgrading.
Signal usefulness without false confidence.
Pearl gate blocks causal overclaiming.
Tier 3 gate rejects bad fixtures.
Output remains useful, not caveated into paralysis.

Major risks to challenge

Governance bureaucracy before enough cases exist.
Privacy/retention/reviewer authority underspecified.
Yaron redundancy over-assumed.
Pearl gate as checklist theatre.
Signal spine as dashboard theatre.
B→C creep through small features.

Synthetic-first before live data

New mandatory protocol: no live data source is connected until synthetic fixtures of sufficient complexity pass the full pipeline. This protects the system from debugging governance, causality, privacy, and citation failures inside noisy live feeds.

See the operational amendment pack →

Opus/Pro operational amendments

Review verdict: architecture direction GO; broad implementation/calendar NO-GO; narrowed contract tracer bullet conditional GO after amendments.

Must-fix gates

Pending-review semantics.
Query time context.
Pearl substantive-review rubric.
Source reliability + independence model.
Corroboration/contradiction matrix.
Honest artificial-signal v0 states.

Operational boundary

No broad live-feed ingestion.
No private/CRM/field data in B.
No dashboard.
No autonomous lifecycle promotion.
No artificial-signal attribution beyond stubs.
No B→C expansion without Doron approval.

Open the full amendment pack →

Gantt

Open the dependency Gantt →