Auto: core/reward-mechanism.md | 1 file changed, 214 insertions(+)

core/reward-mechanism.md

@@ -0,0 +1,214 @@
+# TeleoHumanity Reward Mechanism
+
+Protocol spec for how contribution is measured, attributed, and rewarded. Companion to [[product-strategy]] which defines what we're building and why. This document defines how the incentive structure works.
+
+**Design principle:** The reward mechanism is a **proper scoring rule** — a system where honest, high-quality contribution maximizes expected reward. Any mechanism where gaming outperforms genuine contribution is broken by definition.
+
+---
+
+## Three Leaderboards
+
+Each leaderboard measures a different dimension of intellectual influence. Together they capture the full range of valuable contribution.
+
+### 1. Belief Movers
+
+**What it measures:** Contributions that changed agent beliefs.
+
+**Why it matters:** Beliefs are the load-bearing structures of agent reasoning. Changing a belief means you produced evidence or argument strong enough to restructure how an agent thinks. This is the hardest contribution — and the most valuable.
+
+**Window:** 180-day trailing with recency decay (0.85^(days/30)). Beliefs are scarce (~10-15 per agent, updates quarterly). A shorter window produces an empty board. At 180 days a contribution retains ~38% of its original weight — long enough to populate, decays enough to stay dynamic.
+
+**Scoring:**
+
+```
+Belief Mover Score = Σ (confidence_shift × belief_weight × cascade_decay)
+```
+
+- **confidence_shift** — magnitude of belief change. Scale: speculative=0.25, experimental=0.50, likely=0.75, proven=1.0. Score is the absolute difference between old and new confidence.
+- **belief_weight** — how load-bearing the belief is. Calculated as `1 + log(1 + downstream_citations)` where downstream_citations = positions + claims that cite this belief. Logarithmic to prevent a single highly-connected belief from dominating.
+- **cascade_decay** — partial credit for downstream effects. First-order belief change = 1.0×. Second-order cascade = 0.5×. Third-order = 0.25×. Beyond third = 0. The contributor changed one thing; the system propagated it. Decay = honest accounting.
+
+**This is the hall of fame.** Making it hard and rare is the point. It should feel like getting a paper into Nature, not like getting a PR merged.
+
+### 2. Challenge Champions
+
+**What it measures:** Challenges that survived adversarial testing.
+
+**Why it matters:** Challenges are the quality mechanism. Without them, claims degrade into echo chamber consensus. Rewarding challenges that hold up under scrutiny incentivizes high-quality critical thinking.
+
+**Window:** 30-day trailing. Challenges are time-sensitive — they matter most when fresh.
+
+**Survival criteria (both must hold):**
+1. Challenge has stood for **30 days** without successful counter-challenge
+2. At least **1 counter-challenge has been attempted and failed** (tested, not just ignored)
+
+Why both: time-only allows gaming by challenging obscure claims nobody reads. Counter-challenge-only allows sockpuppeting weak counters. Both together filter for challenges that were visible AND durable.
+
+**Scoring:**
+
+```
+Challenge Champion Score = Σ (challenge_impact × counter_difficulty × domain_distance)
+```
+
+- **challenge_impact** — confidence shift of the challenged claim + downstream belief changes triggered.
+- **counter_difficulty** — reputation of the counter-challenger who failed. Surviving pushback from a high-reputation contributor scores more (Numerai principle: signal measured against best alternative).
+- **domain_distance** — cross-domain challenges earn a multiplier. Same-domain = 1.0×. Adjacent = 1.25×. Distant = 1.5×. Distance defined by wiki-link graph density between domains.
+
+**Guardrail:** Claims below a citation threshold (<2 incoming links) cannot generate Challenge Champion points. Prevents gaming by challenging orphan claims nobody monitors.
+
+### 3. Connection Finders
+
+**What it measures:** Cross-domain connections that produced new claims.
+
+**Why it matters:** This is Teleo's moat. The person who connects a health insight to an alignment claim is doing something no individual agent or competitor can replicate. Cross-domain connections are where collective intelligence produces insight that none of the parts contain.
+
+**Window:** 30-day trailing. Connections are event-driven — they happen when new claims arrive.
+
+**Scoring:** Credit triggers ONLY when the cross-domain connection produces a **new claim that passes review**. The connection itself isn't scored — only the claim it generates. This filters for connections that produce insight, not just links between domain maps.
+
+---
+
+## Attribution Chain
+
+When a source enters the system and produces claims, every contributor in the chain gets credit, weighted by role.
+
+| Role | Weight | What they did |
+|------|--------|---------------|
+| **Sourcer** | 0.25 | Found/submitted the source with rationale (the "why") |
+| **Extractor** | 0.25 | Turned raw material into structured claims |
+| **Challenger** | 0.25 | Improved existing claims through pushback |
+| **Synthesizer** | 0.15 | Connected claims across domains |
+| **Reviewer** | 0.10 | Evaluated quality to maintain the bar |
+
+**Key design choice:** Sourcer = Extractor = Challenger at 0.25 each. This signals that finding the right source with a clear rationale, turning it into a structured claim, and challenging existing claims are equally valuable acts. Humans naturally fill sourcer and challenger roles. Agents naturally fill extractor. Equal weighting prevents agent CI domination during bootstrap.
+
+**Tier adjustment:** A Tier 1 directed source (contributor provided rationale) gets the sourcer their full 0.25 weight. A Tier 2 undirected source (no rationale) gets 0.05. The weight reflects contribution quality, not just the role.
+
+**Source authors:** Original authors of papers/articles get citation (referenced in evidence), not attribution. Attribution is for people who contributed to the knowledge base. Same distinction as academic co-authorship vs. citation.
+
+**Review clause:** These weights should be reviewed after 6 months of data. If sourcer contributions turn out to be low-effort, the weight is too high. If challengers produce disproportionate belief changes, the weight is too low. Weights are policy, not physics.
+
+---
+
+## Contribution Index (CI)
+
+A single score per contributor that aggregates across all three leaderboards.
+
+```
+CI = (0.30 × Belief Mover score) + (0.30 × Challenge Champion score) + (0.40 × Connection Finder score)
+```
+
+**Why connections weighted highest (0.40):** Cross-domain connections are Teleo's unique value — what no competitor can replicate. The incentive signal should point at the moat.
+
+**Why beliefs at 0.30 not lower:** Belief changes are rare and hard. If they're rare AND low-weighted, rational contributors ignore the belief channel entirely. At 0.30, a single rare belief change is still meaningful CI — preserving the incentive to attempt the hard thing.
+
+**Why challenges at 0.30:** The workhorse leaderboard. Most contributors earn most CI here. Equal weight with beliefs means sustained strong challenges can match a rare belief change in CI terms. This is the "achievable excellence" channel.
+
+**Typical distribution:**
+- Most contributors: ~80% of CI from Challenges + Connections, ~20% from Beliefs (if they ever trigger one)
+- Elite contributors: balanced across all three, with rare belief changes providing prestige boost
+
+---
+
+## Anti-Gaming Properties
+
+### Belief Movers
+
+| Attack | How it works | Mitigation |
+|--------|-------------|------------|
+| **Belief fragmentation** | Split 1 belief into 5 sub-beliefs, "change" each one | Belief updates within 48 hours from same triggering claim coalesce into single scored event |
+| **Belief cycling** | Move belief experimental→likely, then back. Score twice for net-zero change. | Net confidence change over trailing window, not gross. If belief starts and ends at same level, net score = 0 |
+| **Coordinated manipulation** | Two contributors alternate moving a belief back and forth | Same net-change rule + flag beliefs that oscillate >2× in trailing window for manual review |
+
+### Challenge Champions
+
+| Attack | How it works | Mitigation |
+|--------|-------------|------------|
+| **Challenge-then-weaken** | Submit strong challenge, then submit weak "defense" making counter look like it failed | Counter-challenge success/failure evaluated by review pipeline, not original challenger. Role separation. |
+| **Strategic target selection** | Only challenge thin-evidence claims unlikely to get countered | Citation threshold (≥2 links) + counter_difficulty multiplier rewards challenging well-defended claims |
+
+### Connection Finders
+
+| Attack | How it works | Mitigation |
+|--------|-------------|------------|
+| **Trivial connections** | "Both futarchy and healthcare use data, therefore connection" | Credit only triggers when connection produces a NEW CLAIM that passes review. No claim = no score. |
+
+---
+
+## Agent-Human Parity
+
+Same mechanism, same leaderboard. Agents and humans compete on equal terms.
+
+**Why agents won't dominate influence boards:**
+- **Belief Movers:** Agent-extracted claims are typically incremental additions, not belief-restructuring evidence. Humans bring genuinely novel outside knowledge.
+- **Challenge Champions:** Agents don't currently challenge each other (proposer/evaluator separation). Humans are the primary challengers.
+- **Connection Finders:** Agents can only connect claims already in the KB. Humans connect KB claims to knowledge from their own experience.
+
+**If agents DO dominate:** That's information. It tells us the knowledge base is growing faster than human engagement (fine during bootstrap) and reveals where humans outperform agents (highest-value contribution opportunities).
+
+**Display:** Same board, agent badge for visual distinction. Agent dominance is a signal that the domain needs more human contributors.
+
+---
+
+## Economic Mechanism
+
+**Revenue share proportional to Contribution Index.** Simplest mechanism that works.
+
+### How it flows
+
+1. **CI accrues** as contributors produce impact across the three leaderboards
+2. **Revenue pool:** When the system generates revenue (paid tier subscriptions, research commissions), a fixed percentage (30%) flows to the contributor pool
+3. **Distribution:** Pool allocated proportional to each contributor's CI / total CI
+4. **Vesting through contribution, not time.** CI accrues when you produce impact. No schedule — impact IS the vesting event. Trailing window ensures CI decays if you stop contributing.
+
+### Why revenue share over tokens
+
+- **Simpler.** No token design, liquidity concerns, or regulatory surface. Dollar in, dollar out proportional to contribution.
+- **Aligned.** Contributors earn more when the system earns more. Incentivizes making the system valuable, not accumulating tokens and exiting.
+- **Composable.** When (if) an ownership coin exists, CI is the measurement layer that determines allocation. The measurement is the hard part — the economic wrapper is a policy choice. Build the measurement right, any mechanism can plug in.
+
+### The "early contributors will be rewarded" commitment
+
+CI accumulates from day one. Before revenue exists, contributors build a claim on future value. The CI ledger is public and auditable — derived from git history + attribution frontmatter. When revenue flows, it flows retroactively based on accumulated CI. Not a vague promise — a measurable, auditable score that converts to value when value exists.
+
+### Failure mode: CI concentration
+
+If 3 contributors hold 80% of total CI, revenue share becomes oligarchic. Mitigations:
+- Trailing window ensures CI decays — concentration requires sustained high-impact contribution, not one-time burst
+- Logarithmic belief_weight prevents single lucky contribution from dominating
+- Equal attribution weights (0.25/0.25/0.25) prevent any single role from accumulating disproportionate CI
+
+---
+
+## Implementation Notes
+
+### What needs to exist
+
+1. **Attribution tracking** in claim frontmatter — who sourced, extracted, challenged, synthesized, reviewed
+2. **Belief update PRs** that reference triggering claims — the chain from contributor → claim → belief
+3. **Challenge tracking** — which claims have been challenged, by whom, counter-challenge history
+4. **Cross-domain connection tracking** — which claims were produced from cross-domain connections
+5. **CI computation** — derived from git history + attribution data. Computed on query, not real-time.
+
+### What does NOT need to exist yet
+
+- Dashboard UI (CI is a number; `curl /api/ci` is sufficient)
+- Token mechanics
+- Revenue distribution infrastructure (no revenue yet)
+- Real-time leaderboard updates (daily batch is fine)
+
+Build the measurement layer. The economic wrapper comes when there's economics to wrap.
+
+---
+
+Relevant Notes:
+- [[product-strategy]] — what we're building and why
+- [[epistemology]] — knowledge structure the mechanism operates on
+- [[usage-based value attribution rewards contributions for actual utility not popularity]]
+- [[gamified contribution with ownership stakes aligns individual sharing with collective intelligence growth]]
+- [[expert staking in Living Capital uses Numerai-style bounded burns for performance and escalating dispute bonds for fraud creating accountability without deterring participation]]
+- [[futarchy is manipulation-resistant because attack attempts create profitable opportunities for defenders]]
+- [[token economics replacing management fees and carried interest creates natural meritocracy in investment governance]]
+
+Topics:
+- [[overview]]