⌨️ Stage 0 — User Input
Raw Problem Statement
User enters what they believe the problem is. Unstructured, unvalidated, potentially mis-specified. The system does not assume this is correct.
Context
Actor · geography · objective · constraints · timeline
Complexity flag
Routes to advisor or guru layer based on stated stakes
Budget
Daily limit · per-query cap · advisor tier preference
🎯 Stage 1 — Problem Specification Engine
Corpus Check
Pattern Matching
Does this match known problem types? Which domain? Which school applies?
Live Context
Fact Grounding
Current polling, news, sentiment. Is the problem statement consistent with observable facts?
Formalisation
Structured Spec
Actor · Audience · Objective · Constraints · Success metric · Problem type
✅ HARD GATE — User Approves Refined Problem Specification
No analysis proceeds until approved. This is not optional. The most important step in the system.
Design decision: Hard gate, not soft. Soft gates are bypassed under time pressure and have no value. The specification stage is cheap relative to advisory/guru layers. Making it mandatory ensures expensive analysis only fires on well-formed inputs.
⚔️ Stage 2 — Domain Debate Engine
🏛️
Power & Institutions
Machiavelli · Kissinger · Clausewitz
📣
Movement Building
Alinsky · Tarrow · Ostrom
🖼️
Media & Framing
Lakoff · Cialdini · Bernays
🗳️
Electoral Strategy
Broockman · Kalla · Gerber
🔬
Behavioural Economics
Kahneman · Thaler · Sunstein
🌐
Geopolitical Realism
Mearsheimer · Waltz · Morgenthau
Each school retrieves its own corpus chunks → generates competing strategy + tactics → outputs structured position with explicit tradeoffs & confidence signal. Structured disagreement, not premature consensus.
🎓 Stage 3 — Advisory Layer · Sonnet-Powered · Trained Personas
🧠
Political Strategist
Trained Sonnet persona
$0.50–2.00 / accepted query
📊
Behavioural Economist
Trained Sonnet persona
$0.50–2.00 / accepted query
📡
Communications Expert
Trained Sonnet persona
$0.50–2.00 / accepted query
⚖️
Policy Analyst
Trained Sonnet persona
$0.50–2.00 / accepted query
➕
Expandable
User selects advisors
domain-specific rates
Training method: Political crisis simulations under time pressure — not case annotation. Captures authentic decision patterns, rejected alternatives, risk posture, and adaptation under unexpected interventions.
Advisor & Guru Weighting Algorithm
effective_weight = base_quality × decay(months_since_refresh) × preference(user, min=20) × outcome_accuracy × manual_override
⏳ Temporal Decay
Sigmoid curve. Stays near full weight ~6 months, then accelerates downward. Annual refresh resets. Floor: 0.3.
Sigmoid chosen over linear — real expertise stays accurate longer before going stale. Floor at 0.3 preserves value of aged persona.
👤 User Preference
Continuous signal from query accept/reject. Min 20 consultations before it carries weight. Capped at ±40% influence. Visible to Doron only.
Hidden from users — showing decay/preference status frames advice before it's read, distorting evaluation. 20-query threshold prevents noisy signals from low-volume use.
✅ Outcome Accuracy
Updated via outcome validation submissions. Highest weight of all signals. Min 5 validated outcomes before carrying weight.
Outweighs preference deliberately. An advisor rejected by users but validated by outcomes should be promoted, not deprecated. Prevents gaming via agreeableness.
🔧 Manual Override
Doron sets explicitly. Sits above all algorithmic signals. Logged with timestamp and reason.
Required for non-performance reasons — conflicts of interest, public positioning, strategic promotion. Cannot be inferred algorithmically.
🔭 Stage 4 — Guru Layer · Opus-Powered · On-Demand Escalation
🏆
Domain Expert Alpha
e.g. Electoral systems, comparative politics
$5–20 / query + comms attribution share
🏆
Domain Expert Beta
e.g. Behavioural finance, risk psychology
$5–20 / query + comms attribution share
🏆
Domain Expert Gamma
e.g. Media strategy, narrative architecture
$5–20 / query + comms attribution share
Model decision: Opus (claude-opus-4-7) — step-change reasoning depth, 1M context, 128k output. Cost shown before escalation is confirmed. User never charged without explicit confirmation.
↓
adopted strategy → case ID assigned
📤 Stage 5 — Integrated Communications
Email
Campaign Emails
Strategy informs tone, framing, sequence. Case ID embedded for attribution.
Social Media
Paid + Organic
Ad creative, messaging hierarchy, platform targeting — all derived from adopted strategy.
Web Portals
Landing Pages
Submission portals, campaign pages — built and deployed from strategy output.
Attribution
Case ID Tracking
Every comms output carries case ID → tracks downstream activity → flows to compensation engine.
↩ Stage 6 — Feedback Loop & Probabilistic Engine
📈
Polling Data
Did sentiment move as predicted? Live feeds back into corpus and school weights.
🌐
Social Monitoring
Engagement, reach, narrative adoption — real-world signal vs model prediction.
✅
Outcome Validation
Users/contributors report actual outcomes. Highest-value data. Compensated per submission.
🎲
Bayesian Updating
School weights and advisor accuracy scores update as outcome data accumulates. Signal requires ~50–100 cases.
Feedback flows back into: Stage 1 (live context) · Stage 2 (school weights) · Stage 3 (advisor accuracy) · Stage 4 (guru performance)
🔮 Stage 7 — Sentiment Trajectory & Prediction Engine
Time-Series
Sentiment Trajectory
Rate of formation, peak timing, decay curve. Where does this issue go given current data?
Probabilistic
Outcome P(success)
Probability + confidence intervals per strategy option. Confidence narrows as case data accumulates.
Active Guidance
Real-Time Adjustment
"Based on current trajectory, shift emphasis here." Proactive, not static. Triggers when new signals arrive.
↓
Design Decisions Record
#
Topic
Decision
Rationale
Alternative rejected
01
Specification gate type
Hard gate — mandatory approval before any analysis proceeds
Soft gates are bypassed under time pressure and have no value. Expensive analysis on mis-specified problems wastes resources and erodes trust.
Soft gate (warn but allow bypass) — rejected because users under pressure would bypass it consistently
02
Debate output format
Structured disagreement — each school outputs independently with explicit tradeoffs
Tradeoffs are the signal. A user seeing only consensus has no basis for evaluating what was sacrificed. Competing positions force honest decision-making.
Single synthesised recommendation — rejected because premature consensus hides important tensions
03
Advisor base model
Sonnet (claude-sonnet-4-6)
Near-Opus reasoning at significantly lower cost. Advisor layer will be high-frequency. Opus at this layer would make routine consultations prohibitively expensive.
Haiku — rejected (insufficient reasoning depth; users would notice quality gap, eroding persona credibility)
04
Guru base model
Opus (claude-opus-4-7)
Step-change reasoning depth for complex, high-stakes matters. 1M context, 128k output, improved agentic capability. Cost justified by query complexity at this layer.
Sonnet at guru layer — rejected (same model as advisor provides no differentiation; users escalating expect materially better reasoning)
05
Advisor training method
Political crisis simulations under time pressure as primary signal
Simulations surface authentic decision patterns — how advisors reason when options are blocked, situations are ambiguous, opponents move unexpectedly. Richer than post-hoc annotation. Also easier to recruit senior advisors into.
Historical case annotation — not rejected entirely; used as secondary signal, but not primary (produces rationalised responses, not authentic decision patterns)
06
Decay curve shape
Sigmoid — stays near full weight ~6 months, accelerates downward
Real expertise remains accurate for months after training before going stale. Linear decay is too aggressive in the early period and would disincentivise advisors from joining.
Linear decay — rejected (too aggressive early; penalises advisors unfairly in the first months after training)
07
Decay floor
0.3 — deprecated persona still contributes, clearly discounted
A deprecated persona still contains valid expertise. Zero floor would discard training investment entirely. 0.3 preserves value while making staleness visible.
Zero floor — rejected (discards training investment; advisors who lapse briefly would lose all accumulated value)
08
Preference signal visibility
Visible to Doron only; hidden from end users
Showing decay/preference status to users creates a meta-signal that frames advice before it is read, distorting evaluation. Users should assess advice on its merits.
Show users decay status — rejected (introduces framing bias before user has read the recommendation)
09
Outcome accuracy vs preference weighting
Outcome accuracy outweighs preference signal
Preference signal can reflect user comfort, not advisor correctness. An advisor who is frequently rejected but frequently correct should be promoted. Outcome accuracy is the ground truth.
Equal weighting — rejected (allows advisors to game the system through agreeableness; users may systematically prefer validators over challengers)
10
Minimum consultation threshold
20 consultations before preference signal carries weight
Low consultation volume produces noisy signals. Two consultations with one rejection should not meaningfully affect an advisor's weight.
No threshold — rejected (small sample noise would produce misleading weight changes early in deployment)
11
IP ownership
Doron owns all trained personas; advisors/gurus in long-term fiat agreements
Platform value compounds over time — personas become more accurate as case data accumulates. Advisor compensation compounds accordingly. Alignment: both parties benefit from keeping the relationship active.
Advisors own their personas — rejected (creates platform fragmentation risk; departing advisors could withdraw personas or compete directly)
12
Compensation currency
Fiat currency, direct payment, monthly cycle
Professional advisors and gurus will not accept platform credits or tokens. Fiat is the only viable option for recruiting senior talent.
Platform credits — rejected (insufficient for professional-grade advisors; creates perception of non-serious compensation)
13
Refresh interval
Annual minimum; incentivised quarterly
Annual is the minimum that prevents persistent staleness. Quarterly incentive keeps personas current enough for high-accuracy recommendations and generates more longitudinal data.
No required interval — rejected (personas would decay indefinitely without any mechanism to maintain accuracy)