stemedb/arena-roadmap.md

Arena Roadmap: The Simulation

Goal: Incrementally evolve the simulator from Spine validation to a full Agent-Based Modeling environment. Philosophy: Make it run. Then add. Verify at every step. Alignment: Tracks main roadmap.md phases; exercises features as they land.

---

Current State (Baseline)

The simulator (stemedb-sim) currently validates Phase 1: The Spine:

Component	Status	What It Proves
WAL Durability	✓ Works	Writes persist
rkyv Serialization	✓ Works	Roundtrip correctness
Ed25519 Signatures	✓ Works	Sign on write, verify on read
Ingestor Pipeline	✓ Works	WAL → KV async flow
Agent Identity	✓ Works	Keypair generation

Run command: cargo run --bin stemedb-sim

What's NOW tested (Arena 1-3):

• ✅ Queries via QueryEngine
• ✅ Lens resolution (Recency, VoteAwareConsensus)
• ✅ Lifecycle filtering
• ✅ Voting & consensus
• ✅ Query audit trail
• ✅ Materialized Views (Arena 3)
• ✅ Fast-path MV reads
• ✅ MV freshness under load

What's NOT yet tested:

• ❌ Concurrent agents (Arena 6)
• ❌ TrustRank (Arena 5)
• ❌ Time-travel queries (Arena 7)

---

Arena Phases

Arena 0: Make It Verifiable ✅ COMPLETE

Goal: Add assertions for what the simulation proves. Currently it prints logs; we need programmatic success/failure.

Why first: Without assertions, we don't know if later changes break things.

• 0.1 Return Result from main(): Change from print-and-exit to structured outcome.

  • Define SimulationResult { assertions_written: u64, assertions_verified: u64, errors: Vec<SimulationError> }.
  • Library function run_simulation(config) -> Result<SimulationResult, SimulationSetupError>.
  • Print summary at end, exit 0 on success, exit 1 on failure, exit 2 on setup error.

• 0.2 Integration Test Wrapper: Make the sim runnable as a test.

  • Add crates/stemedb-sim/tests/smoke.rs with 6 integration tests.
  • Assert on SimulationResult fields.
  • Run in CI via cargo test -p stemedb-sim.

Exit Criteria: cargo test -p stemedb-sim passes ✅

---

Arena 1: Query Path (Exercises Phase 2 Features) ✅ COMPLETE

Goal: Extend simulation to read via QueryEngine, not direct KV access.

Depends on: Phase 2 complete (Query Engine, Lenses) Aligns with: roadmap.md Phase 2 "The Lattice"

• 1.1 Add Query Engine to Simulation

  • Import stemedb-query crate.
  • After ingestion wait, query each assertion via QueryEngine::execute().
  • Verify result matches what was written.

• 1.2 Exercise Recency Lens

  • Write 2 assertions for same subject+predicate with different timestamps.
  • Query with lens=Recency.
  • Verify most recent wins.

• 1.3 Exercise Lifecycle Filtering

  • Write one Proposed and one Approved assertion for same fact.
  • Query with lifecycle=Approved.
  • Verify only Approved returned.
  • Use Case Alignment: This is the JWT signing algorithm bug from Agile Agent Team.

• 1.4 Query Audit Verification

  • Set X-Agent-Id header (or equivalent in direct API call).
  • After queries, call AuditStore::get_audits_for_agent().
  • Verify audit trail exists with correct contributing assertions.
  • Use Case Alignment: "What did the deployment agent query?"

Exit Criteria: Simulation writes, ingests, queries, and verifies all three scenarios. ✅

---

Arena 2: Voting & Consensus (Exercises Phase 2 VoteStore) ✅ COMPLETE

Goal: Simulate agents voting on assertions and resolving via VoteAwareConsensusLens.

Depends on: Arena 1 complete, Phase 2 VoteStore Aligns with: roadmap.md Phase 2 "The Ballot Box"

• 2.1 Add Vote Creation to Agents

  • Agent::vote(assertion_hash, weight) method.
  • Votes stored directly via VoteStore (bypasses WAL for now - see note below).

• 2.2 Conflicting Assertions with Votes

  • Scientist_Alpha asserts "Protein_X binds Receptor_Y" (confidence 0.8).
  • Scientist_Beta asserts "Protein_X binds Receptor_Z" (confidence 0.8).
  • Alpha votes for own assertion (weight 1.0).
  • Beta votes for own assertion (weight 1.0).
  • Third agent (Believer) votes for Alpha's assertion.
  • Query with lens=VoteAwareConsensus.
  • Verify Alpha's assertion wins (2 votes vs 1).

• 2.3 Troll Vote Resistance

  • Troll creates low-confidence assertion contradicting consensus.
  • Troll votes for own assertion.
  • Verify high-vote assertions still win.

Exit Criteria: Vote-based consensus correctly resolves conflicts. ✅

---

Arena 2.5: Hardening (Critical Gap Remediation) ✅ COMPLETE

Goal: Fix critical gaps discovered during Arena 0-2 review before adding more features.

Depends on: Arena 2 complete Blocks: Arena 3+ (don't add features on a shaky foundation) Rationale: Gap analysis revealed 58/100 production readiness score with 3 critical blockers.

• 2.5.1 Fix Vote Cache Race Condition (P0 - CRITICAL)

  • VoteStore put_vote() uses fetch_and_add_u64 + compare_and_swap_f32 (vote_store.rs:182-189)
  • Two concurrent calls can lose updates (final count = N instead of N+1)
  • Solution: Add atomic increment or compare-and-swap operation
  • Add test: 100 concurrent put_vote() calls, verify final count
  • File: crates/stemedb-storage/src/vote_store.rs:181-206

• 2.5.2 Add API Integration Tests (P0 - CRITICAL)

  • Create crates/stemedb-api/tests/http_integration.rs
  • Test POST /assertions - create assertion via HTTP
  • Test POST /votes - submit vote via HTTP
  • Test GET /query - query with lens parameter
  • Test error responses (400 Bad Request, 500 Internal Error)
  • Test rate limiting via QuotaStore middleware
  • Gap: Entire HTTP layer is currently untested

• 2.5.3 Add Crash Recovery Test (P0 - CRITICAL)

  • Write assertions to WAL
  • Kill IngestWorker mid-step (simulate crash)
  • Restart IngestWorker with same WAL + KV store
  • Verify: cursor resumes correctly, no duplicate ingestion (worker.rs:1518)
  • Verify: all pre-crash data is recoverable
  • Validates: Durability claims in architecture.md

• 2.5.4 Add Input Validation (P1 - HIGH)

  • Max subject length: 1024 characters
  • Max predicate length: 1024 characters
  • Confidence range: 0.0 to 1.0, reject NaN/Inf
  • Vote weight: non-negative, reject NaN/Inf
  • Timestamp: reject values > current time + 1 hour (clock skew protection)
  • Add validation in IngestWorker::validate_assertion() and validate_vote()
  • File: crates/stemedb-ingest/src/worker.rs

• 2.5.5 Replace Sleep Timers with Ingestion Sync (P1 - HIGH)

  • wait_until_ingested() cursor-based polling replaces all sleeps
  • Add: wait_for_ingestion(store, expected_count, timeout) helper
  • Poll store until expected assertions exist or timeout
  • Replace all hardcoded sleeps in simulation
  • Benefit: Faster tests, deterministic behavior

• 2.5.6 Fix Defensive Error Handling (P2 - MEDIUM)

  • Input validation propagates errors properly
  • Change to propagate error or skip candidate with warning
  • worker.rs:161-173: Ambiguous EOF handling treats all I/O errors as "no data"
  • Distinguish true EOF from transient errors

Exit Criteria:

• Vote cache is atomic (concurrent test passes)
• API layer has integration tests (POST/GET work via HTTP)
• Crash recovery is verified (no data loss on restart)
• Input validation rejects malformed data
• No hardcoded sleep timers in simulation
• Production readiness score: 75+ (up from 58)

---

Arena 3: Materialized Views (Exercises Phase 2 Materializer) ✅ COMPLETE

Goal: Verify fast-path MV reads work under simulation load.

Depends on: Arena 2 complete, Phase 2 Materializer Aligns with: roadmap.md Phase 2 "Materializer"

• 3.1 Materializer Integration

  • Spin up Materializer alongside Ingestor.
  • Wire Notify between IngestWorker and Materializer.
  • After ingestion, verify MV keys exist in store.

• 3.2 Fast-Path Verification

  • Query via QueryEngine with subject+predicate.
  • Log whether fast-path or slow-path was used (add debug output).
  • Verify MV winner matches slow-path result.

• 3.3 MV Freshness Under Load

  • Write 10 assertions in rapid succession.
  • Wait for materialization.
  • Verify MV reflects latest state.
  • Aligns with: Phase 2.5 "MV Staleness Detection"

Exit Criteria: Fast-path queries return correct results under load. ✅

---

Arena 4: Agent Personas (First Strategy Differentiation) ✅ COMPLETE

Goal: Agents behave differently based on persona. No longer uniform.

Depends on: Arena 3 complete Aligns with: Vision document "The Players"

• 4.1 AgentStrategy Trait

  • AgentStrategy trait with decide_action(), base_confidence(), name().
  • AgentAction enum: Assert, Vote, Query, Skip.
  • WorldState, StrategyMetrics, AgentSpec, StrategyType.
  • GroundTruth hardcoded dataset (5 known-true facts).
  • File: crates/stemedb-sim/src/strategy.rs

• 4.2 Scientist Strategy

  • High base confidence (0.9).
  • Even ticks: assert next ground truth fact.
  • Odd ticks: vote for truth-aligned assertions, or assert if none found.

• 4.3 Troll Strategy

  • Low base confidence (0.4).
  • Contradicts existing assertions (NOT_value, negate, flip, FAKE_ref).
  • Skips on empty world (tick 0).

• 4.4 Believer Strategy

  • Medium base confidence (0.65).
  • Votes for highest-confidence assertion (weight 0.7).
  • Pure amplifier - never creates assertions, only votes.

• 4.5 Strategy-Driven Tick Loop

  • SimulationConfig.agents: Vec<AgentSpec> replaces agent_count.
  • Each tick: strategy decides action, executed by agent.
  • Per-strategy metrics tracked and logged.
  • Arena 4 verification: differentiated behavior check.

Exit Criteria: Different agent types produce different behaviors in logs. ✅

---

Arena 5: TrustRank Integration (Exercises Phase 4 Foundation)

Goal: Reputation updates based on agent behavior.

Depends on: Arena 4 complete, TrustRank implemented Aligns with: roadmap.md Phase 4 "TrustRank Engine"

• 5.1 Initialize TrustRank for Agents

  • Each agent starts with base TrustRank (e.g., 0.5).
  • Store in TrustRankStore at simulation start.

• 5.2 Reputation Adjustment After Votes

  • When an assertion gains votes, increase author's TrustRank.
  • When an assertion is contradicted by consensus, decrease author's TrustRank.
  • Use TrustRankStore::record_outcome().

• 5.3 TrustAwareAuthorityLens Verification

  • Two assertions from different agents, same confidence.
  • Agent with higher TrustRank should win via TrustAwareAuthorityLens.
  • Use Case Alignment: "Expert vs. junior weighting" from Agile Agent Team.

• 5.4 Troll Reputation Decay

  • After 100 ticks, verify Troll's TrustRank has decreased.
  • Verify Scientist's TrustRank has increased.
  • Success Criteria: "Trust clusters form naturally without hardcoded rules."

Exit Criteria: TrustRank diverges based on behavior; Troll reputation tanks.

---

Arena 6: Concurrent Agents (Performance Validation)

Goal: Move from sequential to parallel agent execution.

Depends on: Arena 5 complete Aligns with: Vision "1000 concurrent agents without locking"

• 6.1 Tokio Task Per Agent

  • Wrap each agent's tick in tokio::spawn().
  • Use Arc<Mutex<Journal>> for WAL access (already in place).
  • Run 10 agents concurrently.

• 6.2 Scale to 100 Agents

  • Parameterize agent count.
  • Run with 100 agents for 50 ticks.
  • Verify no deadlocks, no data corruption.

• 6.3 Contention Metrics

  • Add timing around WAL lock acquisition.
  • Log P50/P99 latencies.
  • Identify bottlenecks if any.

• 6.4 Target: 1000 Agents

  • Run with 1000 agents (stretch goal).
  • May require connection pooling or batching.
  • Document findings.

Exit Criteria: 100 agents run concurrently without errors.

---

Arena 7: Time-Travel & Epochs (Exercises Phase 3 Features)

Goal: Validate temporal queries and epoch supersession.

Depends on: Arena 6 complete, Phase 3 Time-Travel + EpochAwareLens Aligns with: roadmap.md Phase 3 "Time-Travel Engine", Phase 2.5 "EpochAwareLens"

• 7.1 Time-Travel Query Verification

  • At tick 50, record timestamp T1.
  • At tick 100, write a new assertion superseding an old one.
  • Query with as_of=T1.
  • Verify result reflects tick-50 state, not tick-100 state.
  • Use Case Alignment: "What was the state of knowledge at 9pm?"

• 7.2 Epoch Creation and Supersession

  • Create epoch "v1" at tick 0.
  • Create epoch "v2" superseding "v1" at tick 50.
  • Assertions referencing "v1" should be filtered by EpochAwareLens.
  • Use Case Alignment: "Security team migrates from RS256 to ES256."

• 7.3 Epoch Cascade Verification

  • Chain: v3 supersedes v2 supersedes v1.
  • Query with EpochAwareLens.
  • Only v3 assertions visible.

Exit Criteria: Historical queries and epoch filtering work correctly.

---

Arena 8: Skeptic & Conflict (Exercises Phase 3 Lenses)

Goal: Surface disagreement, measure consensus.

Depends on: Arena 7 complete, Phase 3 Skeptic Lens + Conflict Score Aligns with: roadmap.md Phase 3C "Skeptic Lens", Phase 3A.2 "Conflict Score"

• 8.1 High-Conflict Scenario

  • 3 Scientist agents assert conflicting values for same fact.
  • Each votes for own assertion.
  • Query with lens=Skeptic.
  • Verify conflict_score is high (> 0.5).

• 8.2 Low-Conflict Scenario

  • 3 Scientists assert same value (agreement).
  • Query with lens=Skeptic.
  • Verify conflict_score is low (< 0.2).

• 8.3 Skeptic Surfaces Outlier

  • Consensus is A, one dissenter says B.
  • Skeptic lens returns B (the controversial position).
  • Use Case Alignment: Financial Due Diligence "disagreement is the information."

Exit Criteria: Conflict score accurately reflects disagreement.

---

Arena 9: Full Gameplay Loop (The Vision)

Goal: Run the complete vision scenario end-to-end.

Depends on: Arena 8 complete, all Phase 3 features Aligns with: simulation-vision.md "The Gameplay Loop"

• 9.1 Ground Truth Injection

  • Load ground truth from YAML config.
  • Scientists read ground truth, assert facts.

• 9.2 The 5-Tick Scenario

  • Tick 1: Scientist asserts "Protein_X binds Receptor_Y".
  • Tick 2: Troll forks with "Protein_X binds Nothing".
  • Tick 3: Believer queries, votes for Scientist.
  • Tick 4: TrustRank updates (Scientist up, Troll down).
  • Tick 5: Verify consensus via lens.

• 9.3 Extended Run (1000 Ticks)

  • Run full scenario for 1000 ticks.
  • Track metrics:
    • truth_convergence: % of facts matching ground truth.
    • reputation_distribution: Scientist vs Troll ranks.
    • fork_depth_max: Deepest contradiction chain.

• 9.4 Success Criteria Verification

  • ✓ Truth survives: High-reputation assertions outlive spam.
  • ✓ Lenses work: Consensus lens filters Troll noise.
  • ✓ Performance: 1000 ticks complete in < 30 seconds.
  • ✓ Emergence: Trust clusters form naturally.

Exit Criteria: All 4 success criteria from vision document pass.

---

Alignment with Main Roadmap

Arena Phase	Exercises Roadmap Phase	Key Features Validated
Arena 0 ✅	-	Test infrastructure
Arena 1 ✅	Phase 2	QueryEngine, Lenses, Lifecycle, Query Audit
Arena 2 ✅	Phase 2	VoteStore, VoteAwareConsensusLens
Arena 2.5 ✅	- (Hardening)	Race conditions, API tests, crash recovery, input validation
Arena 3 ✅	Phase 2	Materializer, Fast-Path MV, MV Freshness
Arena 4 ✅	-	Agent personas: Scientist, Troll, Believer (simulator-only)
Arena 5	Phase 4	TrustRank, TrustAwareAuthorityLens
Arena 6	Phase 4	Concurrency, Performance
Arena 7	Phase 2.5 + Phase 3	Time-Travel, Epochs, EpochAwareLens
Arena 8	Phase 3	Skeptic Lens, Conflict Score
Arena 9	All	Full integration

---

Alignment with Use Cases

Use Case	Arena Phase That Validates It
Agile Agent Team
- Lifecycle filtering	Arena 1.3
- Query audit trail	Arena 1.4
- Time-travel debugging	Arena 7.1
- Expert weighting	Arena 5.3
- Persistent learning	Arena 5.4 (TrustRank)
Financial Due Diligence
- Conflict detection	Arena 8.1, 8.3
- Time-travel	Arena 7.1
- Epoch cascades	Arena 7.2, 7.3
Consumer Health
- Source-class hierarchy	Phase 3 dependency (not in Arena yet)
- Layered consensus	Phase 3 dependency

---

Development Cadence

Week	Focus	Deliverable
1	Arena 0	CI-runnable simulation ✅
2	Arena 1	Query path verified ✅
3	Arena 2	Voting verified ✅
4	Arena 2.5	Hardening: race fix, API tests, crash recovery
5	Arena 3	Materializer + MVs verified
6	Arena 4	Agent personas differentiated
7-8	Arena 5-6	TrustRank + concurrency
9-10	Arena 7-8	Time-travel + Skeptic
11-12	Arena 9	Full gameplay loop

---

Metrics to Track

Once Arena 6+ is complete, export these to logs (and eventually Prometheus):

Metric	Description	Success Target
truth_convergence	% of facts matching ground truth	> 95%
troll_reputation	Troll agent TrustRank at end	< 0.2
scientist_reputation	Scientist agent TrustRank at end	> 0.8
fork_depth_max	Deepest contradiction chain	< 10
p99_write_latency_ms	Write path latency	< 10ms
p99_query_latency_ms	Query path latency	< 50ms
concurrent_agents	Max concurrent agents without errors	1000

---

Non-Goals (Kept Simple)

These are explicitly out of scope for the Arena:

• Prometheus/Grafana integration - Logs suffice for Phase 3.
• YAML scenario config - Hardcoded scenarios are fine until Arena 9.
• Full chaos injection (network partitions, node kills) - Basic crash recovery in 2.5; advanced chaos deferred to Phase 4+.
• External agent frameworks (ADK-Go) - Simulator uses Rust agents.

Note: HTTP API testing was previously a non-goal but is now addressed in Arena 2.5.2 due to critical gap discovery.

---

Next Step

Arena 0-4 and Arena 2.5 are complete. Proceed to Arena 5: TrustRank Integration.

# Verify Arena 0 + 1 + 2 + 2.5 + 3 + 4 still work:
cargo test -p stemedb-sim

# Binary also works (shows persona differentiation):
RUST_LOG=info cargo run --bin stemedb-sim