Master Strategist Evaluation
New
Premium tier that adds named human expert perspectives on top of
the engine's output. Strategists go through 10 complex profiling
scenarios that capture their cognitive style, doctrine affinities,
risk posture, and decision-making patterns. At query time, each
profile is injected into the model to simulate how that expert
would evaluate the options.
Built
10 profiling scenarios (3 tiers), deliberation capture,
cognitive profile synthesis, multi-perspective evaluation,
attribution + compensation ledger.
Next step
First master strategist enrollment. Profiling sessions ready
to run.
AXIOM — Autonomous Political Engagement System
Live
Dual Instance
Full-stack autonomous Twitter/X campaign engine running dual instances across the Australian political spectrum. Each instance generates authentic, values-driven commentary, engages curated AU intellectuals, scans trending political conversations, and posts at optimised irregular intervals — entirely without human intervention. Simultaneously a communications platform and a longitudinal research instrument for comparative political messaging analysis.
Three engagement layers
Dual instances: One centre-left account (engages Denniss, Joshi, Loewenstein, Triggs, O'Shea et al) and one centre-right account (engages Berg, Roskam, Sloan, Switzer, Patrick et al). Both run identically — same architecture, separate data stores, independently scheduled.
Content pipeline: Scans AU political RSS feeds per-instance (left: Guardian, ABC, Crikey, Michael West; right: AFR, The Australian, Spectator, IPA). Generates substantive 2–3 sentence commentary via local LLM. Posts 2–6×/day at irregular intervals (6am–1am Sydney).
Thinker engagement: Monitors 10 curated AU intellectuals per instance. Generates substantive replies — extend, question, bridge, challenge modes — max 3/day per account.
Topic scanner: Searches #auspol + 7 topic streams per account. Comments on posts gaining traction, max 5/day.
Engagement layer: Likes (up to 80/day), follows (up to 15/day), and quote-retweets with AI-generated supportive commentary (up to 5/day). 3×/day per account. Hard content filter — extremism, hate, and racism blocked at generation before any post.
Research-grade data collection
Per-tweet: Emotional tone, sentiment, style register, sophistication level (1–4), emotion/reason ratio, rhetorical device, content type, hierarchical topic taxonomy (9 domains → categories → subcategories), full message metrics, time-series engagement snapshots, velocity checkpoints (T+1h, T+6h, T+24h, T+72h).
Audience panel: Longitudinal panel — geography, AU state/city, gender (imputed), age range (imputed), political lean, profession, issue interests, repeat interaction tracking. GDPR-excluded. Aggregates only.
Network diffusion: Directed propagation graph — hop number, actor influence tier (macro/mid/micro/nano), node role (broadcaster/influencer/peer/listener). Interactive D3.js force-directed visualisation.
AXIOM Panel Weighting Architecture Live
Every interactor in the audience panel carries a composite panel_weight that corrects for bot activity, posting volume bias, and account maturity before any aggregate analysis. This ensures high-volume accounts and inauthentic actors do not skew sentiment, engagement, or audience distribution metrics.
Three-factor composite weight
panel_weight = authenticity × activity × maturity
Authenticity (bot probability): Heuristic scoring on tweet velocity, follower/following ratio, mass-following patterns, bio keywords, username digit patterns, account age. Score 0.0–1.0; human_weight = 1 − score. Obvious bots weight 0.0, established accounts with rich bios weight 1.0.
Activity (volume category): Tweets/day bucketed as dormant (<0.1) → light → moderate (baseline 1.0) → active → high_volume → extreme (>50/day, weight 0.15). Prevents a single power-user from dominating aggregates.
Maturity: Account age bucketed as new (<90 days, weight 0.6) → established → veteran → legacy (>5 years, weight 1.0). Newer accounts carry weaker signal.
Six additional panel fields
Engagement depth: lurker / one-time / recurring / advocate — with depth_weight for advocacy-intensity weighting.
Network tier: nano / micro / mid / macro / mega by follower count — reach_weight for audience-size weighting.
Account status: active / cooling / churned based on days since last engagement — enables retention and churn tracking over time.
Temporal pattern: Hour-of-day and day-of-week distributions with peak hour/day — feeds scheduling optimisation.
Content affinity: Which topic tags reliably trigger engagement per user, updated longitudinally — enables topic-level audience segmentation.
Interaction sentiment: supportive / neutral / critical per observation — distinguishes real supporters from passive or adversarial engagers.
Social Messaging — Signal & WhatsApp Live
Two-way Signal (signal-cli, Albert-DGX device). Two-way WhatsApp (Pixel ADB). MacroDroid notification routing (5 platforms). Autonomous reply via qwen2.5:14b.
Strength
Full send/receive on Signal and WhatsApp. Notifications from Instagram, LinkedIn, and Twitter routed through MacroDroid to Albert in real time.
Current limit
SMS send via MacroDroid macro not yet wired. LinkedIn profile photo upload pending.
Political Strategy Intelligence Assistant Live
Four-stage engine: Stage 1 problem spec gate → Stage 2 corpus retrieval (441 books, 225,872 chunks) → Stage 3 iterative strategy refinement loop → Stage 4 tactics decomposition (platform-executable vs user-led). GDELT live context enabled by default.
Strength
Full pipeline from problem spec to accepted strategy to concrete tactics, with user refinement loop and preference-ordered execution plan.
Current limit
Tactics execution fully wired: 16 capabilities, real handlers for deploy, email, forms, research, monitoring, and all document drafts. Master strategist enrollment pending first enrollee.
Political Strategy Source Accumulation
Rebuilt around ingestibility, not store ownership. Books,
journals, reports, speeches, and postmortems are now treated as
different source classes.
Strength
Clear doctrine-gap-driven sourcing policy and efficient
sourcing workflow now exist.
Current limit
No live per-source operational ledger yet.
Attio CRM
API-first operational CRM for political, donor, advocacy, and
event structures.
Strength
Schema is live and tailored to the Australian political
context. Tester enrollments auto-sync.
Current limit
Views, pipelines, and live data import still need further
work.
Canary Marker + Email
Leak-attributable broadcasts and programmatic email handling are
operational.
Strength
Tested end-to-end with encrypted maps.
Current limit
Operational/legal use still requires judgment.
Websites + Intake Infrastructure
Cloudflare Pages and Workers provide autonomous website and form
deployment.
Strength
Operational deployment pipeline with analytics and KV-backed
forms.
Current limit
Domain purchase still has a manual CAPTCHA step.
Memory + Knowledge
Persistent local semantic memory, wiki compilation, and long-term
continuity.
Strength
Zero API-cost memory with semantic recall.
Current limit
Quality still depends on disciplined capture.
Observability + Recovery
Grafana, Prometheus, Loki, watchdogs, and backup protocols now run
with much better discipline.
Strength
Current-work state and cost semantics are far more truthful
than before.
Current limit
More domain-specific panels can still be added.
AZHA Antisemitism Classifier
Two-stage AI pipeline for Royal Commission submission analysis. Stage 1: IHRA definitional gate (GPT-5.4 primary, Sonnet fallback). Stage 2: seven-dimension social and moral theory scoring — identity economics, moral utility, moral disengagement, dehumanisation, norm cascades, signalling theory. Evidence collection portals live for community and individual submitters.
Strength
Gold-standard calibration set seeded. Community portal (submit.professionalopinions.com.au) and individual portal (evidence.professionalopinions.com.au) live. Files stored in Cloudflare R2, metadata in KV. Royal Commission strategy at royalcommission.professionalopinions.com.au.
Current limit
4 gold-standard examples so far; needs ~100 for paper-quality accuracy. Classifier runs locally — API pipeline ready to activate.
Payments + Donation Flow
Stripe fully deployed live end-to-end. Webhook worker live at
azha.professionalopinions.com.au/api/stripe-webhook.
Strength
Products, payment links, webhook handler, and Attio CRM
integration all operational. STRIPE_WEBHOOK_SECRET verified.
Current limit
No customer-facing payment dashboard or refund flow yet.
GDELT Live Context
Real-time geopolitical event signals injected into strategy
synthesis. Enabled by default, fault-tolerant.
Strength
Keyword extraction + geography-aware queries for focused
results. Non-fatal fallback if GDELT unavailable.
Current limit
Phase 1 — coverage depends on query quality and media
source availability.
Tester Enrollment Pipeline
End-to-end: web form → Cloudflare Worker → KV →
enrollment sync → users.jsonl + preference profiles →
Attio CRM upsert.
Strength
Deterministic user attribution via SHA-256 email hash.
Self-rating capture with sliding scales.
Current limit
Preference profile feedback loop not yet wired to model
calibration.
Political Strategy Simulation Engine
Multimodal crisis simulation for adviser training. Albert facilitates completely — 7 characters with distinct voices across Telegram, email, and SMS. Timed decisions, adversarial information, compounding consequences. Participants use the decision engine as their primary tool. Triggered via /startsim command; participants receive SMS from Marcus Webb, complete enrollment profile, then enter the live scenario.
Strength
Live at sim.professionalopinions.com.au. Enrollment at join.professionalopinions.com.au. Sprint (30min), deep (2hr), slow-burn (multi-day). Post-simulation debrief with decision scoring, info quality reveals, corpus lessons.
Current limit
One scenario built (The 36-Hour Window — election crisis). SMS delivery activates when Pixel is configured. Ministerial scandal scenario in build queue.
Core Function Protection
SHA-256 manifest checksums on 13 protected files, script
permission auto-repair, gateway health monitoring, update
governance protocol.
Strength
Automated integrity checks, watchdog every 20 minutes,
documented rollback procedures.
Current limit
Rollback is manual — no automated snapshot-and-restore
yet.
Print & Physical Output
Print-ready PDFs, decks, and reports can be generated for
real-world use.
Strength
Operational report/deck rendering paths exist.
Current limit
Printer-device control is still not explicitly
integrated/documented.