The whole system is a controlled decision pipeline.
This map shows the full Behaviour Decision Intelligence ecosystem as a set of subsystem contracts: what each layer receives, what it transforms, what it emits, and which controls keep it auditable. It separates user-facing tools, corpus and live-signal inputs, retrieval and inference, review gates, delivery surfaces, feedback learning, and operations.
1024source candidates in the political strategy corpus
342,787indexed passages available to retrieval and source-pack
construction
5,523Bronze lakehouse artifacts tracked for provenance
Gatedlatest benchmark claims require attached gate status before
publication
End-to-end architectural map
Read left to right for the main decision path. The two lower rows are the platform's executive functioning — the control, operations, audit and learning rails that wrap every production path and keep it honest.
InputData/evidenceStore/indexProcessModel/inferenceControlOutputOperations
Hover, tap or focus any subsystem box to see its inputs, transform, outputs and controls.
Subsystem input/output contracts
Each subsystem has a narrow job. The ecosystem only works if outputs are explicit enough for the next subsystem to inspect, reject, repair, or promote.
| Subsystem | Primary inputs | Transform | Primary outputs | Controls and failure modes |
|---|---|---|---|---|
| 1. User and client surfaces | Client objective, actor, geography, constraints, available evidence, desired output, stakeholder questions. | Convert a rough request into a structured brief and make hidden assumptions visible before computation starts. | Confirmed specification, output contract, unresolved-question register, route hints. | Blocks vague briefs where possible. Failure mode: accepting a fashionable prompt instead of the real decision problem. |
| 2. Case contract and routing | Confirmed specification, jurisdiction, timeframe, evidence expectations, live-context toggle. | Classify question type, action stream, premise risks and evidence role requirements. | Case record, route contract, premise flags, evidence role gate, result artifact paths. | Question-type correctness checks. Failure mode: one generic schema for unlike problems. |
| 3. Source ingestion | Books, papers, official documents, Canberra bulletins, Australian Electoral Commission material, polling files, news and live feeds. | Canonicalise files, extract text, attach provenance, date, topic, system tags, and source-domain metadata. | Corpus records, chunk records, metadata manifests, ingestion logs. | No untagged corpus additions. Canberra PDFs are wrappers only; linked source documents are the ingest target. |
| 4. Lakehouse and evidence spine | Raw artifacts, extracted chunks, source-pack rows, eval scores, generated reports. | Separate raw Bronze artifacts, Silver normalised rows, and Gold evidence records for audit and reproducibility. | DuckDB/Parquet views, evidence records, source-pack composition summaries, eval lineage. | Source-pack composition is inspectable before promotion. Failure mode: outputs that cite sources but cannot prove where they came from. |
| 5. Embeddings and indexes | Tagged chunks, chunk text, metadata, new source additions. | Build vector embeddings, text search records, chunk metadata and rebuilt NumPy/vector indexes. | Dense retrieval index, text index, chunk lookup files, pending/progress state. | Corpus count, embedding count, index and tag counts must match. Failure mode: silent mismatch between corpus and retrieval. |
| 6. Live-signal layer | Global Database of Events, Language and Tone feeds, polling, media narratives, X/social signals, current Australian political context. | Interpret live information as timing, salience, actor and risk signals, not as primary doctrine. | Live-context bundles, leading/lagging indicators, tactical-window notes, entity/event summaries. | Live indicators are segregated from citeable doctrine. Failure mode: treating noisy live signals as evidence of durable mechanism. |
| 7. Retrieval router | Case contract, query decomposition, authority matrix, corpus indexes, route-specific requirements. | Run dense retrieval, keyword retrieval, hypothetical-document expansion, reciprocal-rank fusion, authority and geography weighting. | Candidate evidence pools, retrieval traces, authority-balanced source candidates. | Kill switches and ablations required. Failure mode: classifier overfit or retrieval that satisfies the wording rather than the decision need. |
| 8. Source-pack builder | Candidate pools, route contract, Australian floor, evidence role gates, source-quality rules. | Rerank, filter, reserve direct evidence, inject structured context separately, preserve retrieval trace. | Frozen source pack, reviewer trace, citation candidates, evidence role warning or hard fail. | Direct-evidence gates for specific routes. Failure mode: plausible citations that do not support the claim being made. |
| 9. Measurement and inference | Source pack, live indicators, scenario variables, known constraints, causal hypotheses. | Run causal checks, scenario modelling, uncertainty framing, influence-network or temporal analysis where justified. | Measurement objects, causal caveats, scenario expectations, falsifiers and pivot triggers. | Structured context is labelled as dataset context, not corpus citation. Failure mode: confusing a modelled signal with observed evidence. |
| 10. Synthesis engine | Source pack, measurement objects, case contract, live-context bundle, prior feedback, output schema. | Generate strategic options, recommendations, implementation sequence, owner/channel/timing suggestions, and caveats. | Decision packet, options, recommended path, evidence ledger, action instrumentation. | Advisor-quality gate rejects boilerplate. Failure mode: competent but ordinary advice that does not create advantage. |
| 11. Review and canary gates | Generated answer, source pack, schema, judge prompt, canary cases, previous baseline. | Check grounding, mechanism quality, Australian fit, actionability, calibration, refusal integrity, regressions. | Scores, review findings, repair instructions, go/no-go state, audit artifacts. | One change, one canary, one decision. Failure mode: bundling changes until score movement is uninterpretable. |
| 12. Delivery surfaces | Approved decision packet, generated artifacts, visibility classification, recipient/channel. | Render to protected pages, reports, email, Telegram updates, PDFs, search entries or client-facing tools. | Strategy report, review console, dashboard page, archive/search record, delivery notification. | Publication/auth gate applies before deploy. Failure mode: private capability or client data exposed through a new route. |
| 13. Feedback and learning | User selection, off-menu choices, implementation outcomes, observed response, quality ratings. | Measure the real-world impact of deployed tactics, then translate it into source credibility, tactic performance and preference updates. | Preference weights, source credibility deltas, tactic promotion/demotion/modification fed back to retrieval, the evidence spine and synthesis. | Learning is controlled and auditable. Failure mode: letting anecdotal feedback silently distort retrieval or strategy. |
| 14. Operations and supervision | Named jobs, cron schedules, systemd services, Albert Orchestration sessions, Alberta compute tasks, deploy/test commands. | Run durable supervision, checkpoints, logs, health checks, restarts, rollback and cross-box testing. | Job state, watchdog alerts, monitor logs, health checks, verified deployment state. | Long-running work uses supervisor. Failure mode: a job dies quietly and the visible product lies about freshness. |
Control planes
Governance and publication
Every protected surface is treated as closed by default. New pages must pass authentication, search/index checks, sensitivity classification and unauthenticated HTTP checks before being considered safe.
Evaluation discipline
Pipeline changes are interpretable only when the judge, excerpt strategy, prompt, config and source pack are pinned. Canary results are evidence only when they isolate one changed layer.
Operational supervision
Albert Orchestration handles interaction and orchestration on Albert. Alberta is the compute/test peer for GPU-heavy and cross-box validation. Long-running work requires an explicit supervisor and observable logs.
Feedback loops
Strategic learning loop
- A recommended tactic is selected, modified or rejected, then deployed through the delivery surfaces.
- The deployed tactic's real-world impact is measured against the action, audience, channel and evidence assumptions.
- Measured impact is translated into source-credibility, tactic and preference updates through controlled learning gates.
- Those updates feed back into the entire engine — retrieval weighting, the evidence spine and synthesis priors — while the audit trail stays visible.
System learning loop
- Canaries and reviews identify a specific defect.
- One layer is changed with a kill switch or ablation path.
- The same canary runs again under pinned conditions.
- The change is promoted, reverted or held as experimental based on the delta.
The conceptual shift is this: the platform is not a dashboard and not a single model call. It is a controlled transformer. The core product is the contract between inputs, evidence, inference, review, delivery and the measured-impact learning loop that flows back into it.
Glossary and boundaries
| Term | Meaning in this ecosystem | Boundary |
|---|---|---|
| Strategy Engine / Decision Engine workflow | The protected workflow: brief, evidence admission, synthesis, audit, delivery, measurement and controlled learning. | Signals and experimental capabilities are status-labelled before they influence advice. |
| Strategy Engine | The reasoning path that turns a confirmed problem into strategy, options, implementation sequence and evidence-backed decision packet. | Does not own all ingestion, publication or operations infrastructure. |
| Albert live-signal layer | The live/current-context layer for timing, salience, actors and campaign intelligence. | It localises doctrine; it is not primary citation doctrine by itself. |
| Global Database of Events, Language and Tone | External live event and media signal source used for current-context and salience analysis. | Useful as an indicator; not a substitute for source-grounded strategic evidence. |
| Australian Electoral Commission | Official electoral rules, disclosure and result context for Australian political work. | Legal/compliance gate and evidence source, not a strategy generator. |
| Alberta | The DGX compute peer used for validation, GPU retrieval/reranking work and cross-box testing. | It does not run Albert Orchestration as the primary control surface. |