AGNN Co-Learning Capability Ground Truth

Feeder doc — this is GPT base iteration — build-sequence detail; the canonical spine reconciles it (reading-order #3). Start at the family front door: delivery/AGNN-INDEX-AND-ALIGNMENT-2026-07-04.md. The canonical capability map + backlog is delivery/AGNN-CO-LEARNING-OBJECTIVES-CAPABILITIES-FLOOR-2026-07-04.md — if this doc disagrees with it on a code-truth claim, the canonical map wins.

Date: 2026-07-04
Purpose: Base document for translating the GNN Game-Players / autonomous build showcase pattern into Conduir capabilities.
Scope: Reconciles the target co-learning model with the Conduir base capability model, local contracts, service code, Guardian design material, AGNN research notes, and the Confluence showcase source page.

1. Ground Truth

The useful capability is not "a GNN that plays games." The reusable capability is an arena factory:

state graph -> competing player policies -> isolated trial -> mechanical settlement
            -> OutcomeEvent / ObjectiveScore -> governed promotion or rejection

For Conduir, this means the game-player pattern should become a governed co-learning substrate used by Guardian, self-evolution, routing, memory, blueprint simulation, cost optimisation, HITL reward mining, and meta-review.

The system should not promote neural novelty as the product. It should promote repeatable arenas with:

explicit players,
explicit state graphs,
isolated side-lane execution,
mechanical settlement,
canonical telemetry,
strong baselines,
operator-controlled promotion.

2. Source Material Reviewed

Confluence Target Source

The Confluence page GNN Game-Players - Autonomous Build Showcase (Tic-Tac-Toe -> Chess -> Poker) was accessed on 2026-07-04 through the Confluence Cloud API using Vault/Jira-provided credentials. No credential material is copied into this document.

The page demonstrates one autonomous build session progressing through tic-tac-toe, chess, Kuhn poker, and Leduc hold'em. The important transfer pattern is:

graph-state representation for each domain,
player policy expressed as a small message-passing model,
learning method chosen to match the game structure,
exit criteria settled by oracles or mechanical game evidence,
all failed training attempts preserved,
metrics validated against known fixed points before being used as gates,
swarm findings split into confirmed/folded vs rejected/refuted,
formal gates pre-registered where sample-shopping risk existed,
caveats recorded beside results instead of hidden.

The page is therefore treated as evidence for the method: iterate, measure against ground truth, adversarially review, fold confirmed findings, re-verify, and keep the audit trail.

Operator-Provided Capability Mapping

The excerpt maps AGNN / co-learning features to:

CAP-5 self-evolution and missing contradictor/refuter field,
CAP-4 routing,
CAP-8 blueprints,
CAP-2 memory,
CAP-1 team protocols,
CAP-6 exploration,
CAP-7 twin,
CAP-3 block library,
Guardian,
CAP-9 substrate,
cross-cutting causal fabric, HITL, meta-review, and cost game.

Local Ground-Truth Documents

delivery/CON-4464-ADVERSARIAL-GNN-RESEARCH-2026-07-04.md
delivery/NEURAL-NETWORK-VALUE-AVENUES-SPEC.md
delivery/GUARDIAN-WARGAME-BLUEPRINT-2026-07-03.md
delivery/GUARDIAN-FABRIC-DECISION-CHARTER-2026-07-02.md
delivery/GUARDIAN-FABRIC-EPIC-PHASE-1-2026-07-02.md
delivery/GUARDIAN-FABRIC-PLATFORM-DESIGN-2026-07-02.md
delivery/GUARDIAN-INFRA-PROD-ROADMAP-2026-07-02.md
delivery/CAP-6-PORT-PLAN.md
delivery/CAP-8-PORT-PLAN.md
delivery/CAP-10-DELTA-AUDIT-2026-05-18.md
CAPABILITIES.md

Local Code Surfaces Reviewed

contracts/python/move_proposal_v0.py
contracts/schemas/move_proposal.v0.schema.json
contracts/python/side_lane_envelope_v0.py
contracts/python/objective_score_v1.py
contracts/python/outcome_event_v1.py
services/evolution/src/evolution/swarm/red_team.py
services/evolution/src/evolution/swarm/meta_swarm.py
services/evolution/src/evolution/swarm/cost.py
services/evolution/src/evolution/swarm/promotion_gate.py
services/evolution/src/evolution/swarm/board.py
services/evolution/src/evolution/swarm/panel.py
services/evolution/src/evolution/inbox/store.py
services/symposium/src/symposium/experiment_store.py
services/symposium/src/symposium/bandit_manager.py
services/symposium/src/symposium/smart_router.py
services/symposium/src/symposium/memory/grafosa_memory_layer.py
services/api-bff/src/types/contracts/move-proposal.ts
services/api-bff/src/types/contracts/objective-score.ts
services/api-bff/src/services/hitlQueue.service.ts
services/api-bff/src/services/feedback.service.ts

3. Terminology

Term	Meaning in this document
AGNN	Broad adversarial graph / neural / game-player pattern. Not every arena needs a trained GNN in v0.
Arena	A bounded contest with players, state, action space, settlement, telemetry, and promotion rules.
Player	A policy that proposes, attacks, routes, defends, prunes, critiques, or simulates. May be scripted, LLM-based, bandit-based, or neural.
Referee	Substrate, Guardian, cert-fabric, quality-fabric, or operator authority that settles outcomes. The referee is not a contestant.
Settlement signal	Mechanical evidence used to score an arena. Examples: test pass, sandbox fitness, attack/detection rate, operator accept/reject, token cost, route score.
Side lane	Isolated experiment lane wrapped by `SideLaneEnvelope`; cannot contaminate main-lane learning until promotion criteria pass.
Promotion	Governed transition from shadow/side-lane evidence to an allowed product behavior.

4. Non-Negotiable Design Rules

1. Mechanical settlement over LLM self-grading. An arena can use LLMs as players, but its reward must be settled by code, tests, sandbox results, telemetry, operator decisions, or signed security evidence.
2. CAP-9/Substrate is the referee, not a player. ActorContext, OutcomeEvent, tenant isolation, audit chain, path maturity, and policy gates must remain outside self-mutation.
3. Guardian GatePolicy is immutable-by-swarm. Red/blue panels may learn, but signed Guardian policy and production enforcement rules are operator/substrate authority.
4. Side-lane isolation is mandatory. New players and learned policies start as shadow or side-lane experiments before product use.
5. Strong scripted baselines are required. The research sweep notes that heuristics can beat MARL in cyber gyms; neural agents must beat practical baselines, not just weak model opponents.
6. Adaptive attacks are required for graph defenses. Homophily pruning and trimmed-mean aggregation are not sufficient as standalone claims.
7. Metric gates need fixed-point validation. A known-good strategy must score good and a known-bad strategy must score bad before any metric can gate promotion.
8. Pre-register formal gates when sampling can be gamed. A single fixed-seed or fixed-window run should be declared before the evaluation when pass/fail could be sample-shopped.
9. Failures are evidence. Failed arena attempts, rejected reviewer findings, and corrected ruler bugs must stay in the audit trail because they are part of the learning substrate.
10. Rules can referee; learned players should not redefine legality. Legality masks, sandbox boundaries, and Guardian policy are allowed as external constraints, but they are not proof that the learned player owns the rule.
11. No capability claim without a settlement signal. Every claimed co-learning capability needs a named metric and a replayable evidence source.

5. Current Code-Truth Snapshot

Surface	Current truth	Capability implication
`MoveProposalV0`	Exists with proposer, evidence, benefit, rollback, tier, and status fields. No explicit `refuter`, `contradictor`, or adversarial review population field.	CAP-5 proposer-vs-refuter is the clearest missing contract-side opportunity. Start with sidecar `RefutationRecordV0` before changing frozen proposal shape.
Evolution Red Team	`red_team.py` implements Red-Queen challenger fitness and enable gate, explicitly shadow/default-off.	CAP-5 self-evolution has proof substrate but should remain flag-off until shadow difficulty and downstream regression evidence are stable.
Meta-swarm	`meta_swarm.py` emits advisory proposals and explicitly "applies NOTHING."	Meta-review is a real but inert loop. It can become a shadow arena by scoring reviewer proposals against later ground truth.
Cost-aware selection	`cost.py` has cost utility, default `cost_utility_weight = 0.0`.	Cost game exists as a dormant lever. Turn it on first in shadow with `ObjectiveScore.cost_efficiency` and quality guardrails.
ExperimentStore	`experiment_store.py` records task intent, verification, cost, latency, tokens, quality, model, provider, and per-stage outcomes.	CAP-6 is the natural arena factory backend. It can host many co-learning trials without inventing a new storage service.
SideLaneEnvelope	Ported contract exists. CAP-6 port plan still calls out missing parity/hardening items.	Use as the isolation envelope for every arena. Do not introduce a parallel experiment wrapper unless this contract cannot carry the evidence.
ObjectiveScoreV1	Ported production reward contract with real/simulated and main/sidelane concepts.	Use as the canonical score carrier for routing, simulation, cost, and arena promotion.
OutcomeEvent	Canonical learning/audit bus exists across repo docs and contracts.	Arena outputs should become OutcomeEvents, not private logs only.
HITL queue / feedback	API-BFF has human approval/rejection and feedback surfaces.	HITL reward mining is high signal-per-dollar because operator decisions already exist in product flow.
GraFOSa memory layer	Graph memory nodes/edges, temporal validity, and traversal surfaces exist.	CAP-2 memory keep/prune game can be framed over existing memory graph structures.

6. Capability Build Matrix

Product Capabilities

Capability	Verdict	Buildable capability	Settlement signal	Current status	Recommended next move
CAP-6 exploration	Highest leverage	Arena Factory: shared arena lifecycle using ExperimentStore, SideLaneEnvelope, ObjectiveScore, OutcomeEvent.	experiment creation to adoption time; side-lane pass rate; graduation rate	Substrate partially present; not fully lane-wrapped.	Build the arena contract pack first.
CAP-5 self-evolution	Helps now	Red-Queen challenger: coevolving adversarial tests against swarm proposals.	hidden tests, sandbox fitness, delayed regression	Red-team primitive exists, shadow/default-off.	Keep shadow; add promotion criteria and regression attribution.
CAP-5 contradictor	Missed opportunity	Proposer-vs-refuter proposal review with adversarial refutation records.	proposal acceptance vs later regression; refuter hit/miss	Verified contract gap: `MoveProposalV0` has proposer but no refuter.	Add `RefutationRecordV0` sidecar referencing `proposal_id`.
CAP-4 routing	Helps after substrate	Worst-case task adversary hardens router/bandit choices.	ObjectiveScoreV1, route pass rate, latency, cost	Hypothesis; depends on SUB-9.14/path maturity and score wiring.	Start with offline replay of historical tasks before online routing.
CAP-8 blueprints	Reshape after simulator	Ideator-vs-simulator-vs-critic tournament for blueprints that arrive pre-attacked.	what-if simulation outcome; simulated vs real delta	Simulator/contracts still being ported; hard prerequisite.	Do not build tournament until `BlueprintSimulationResultV0`/simulator path is real.
CAP-2 memory	Reshape	Consolidator-vs-forgetter learns keep/prune decisions over memory graph.	downstream retrieval hit, contradiction rate, retrieval poison avoidance	Graph memory surfaces exist; neural keep/prune is not wired.	Start with rule baseline and shadow labels from retrieval outcomes.
CAP-1 team protocols	Reshape	Proposer-vs-refuter handoff protocol arena for debate/team structures.	downstream test/schema pass; handoff edge success	SwarmHandoffEdge exists as substrate; verdict quality alone is not enough.	Score team protocols by downstream artifacts, not by debate style.
CAP-7 twin	Weak/contested	Twin generator vs discriminator on held-out user/operator actions.	held-out action accuracy; user correction rate	Could be single-agent imitation, not true co-learning.	Keep as deferred unless operator-visible correction loop is present.
CAP-3 block library	Reshape, deferred	Block adoption marketplace where templates compete by measured reuse efficacy.	template efficacy, adoption, rollback rate	Off critical path; useful after CAP-6/CAP-8 mature.	Defer until arena and blueprint telemetry are stable.
Guardian	Helps already	Security red/blue wargame with signed mechanical settlement.	attack rate, detection rate, false positive rate, signed GatePolicy decision	Correct security instance; policy mutation is no-go.	Productionize Guardian as referee plane, not learner-owned policy.
CAP-9 substrate	Referee	No player role. Provides identity, events, isolation, score, audit, replay.	substrate conformance	Foundation, not peer product.	Protect from swarm mutation; close substrate seams before strong claims.

Cross-Cutting Arenas

Arena	Buildable capability	Settlement signal	Status	Recommendation
Cross-service causal fabric	Learn from outcome streams crossing service boundaries.	mission-level OutcomeEvent and replayed causal chain	Biggest miss; requires clean OutcomeEvent coverage.	Treat as second-wave high leverage after arena contracts exist.
HITL boundary	Mine operator approve/reject/correct as reward.	operator decision, correction text, later regression	Highest signal-per-dollar.	Build early; it reuses product events instead of synthetic labels.
Meta-review	Reviewers co-evolve from hit/miss history.	reviewer hit/miss vs ground truth	Code exists but advisory only.	Start in shadow; never let reviewer self-grade.
Cost game	Budget adversary vs quality defender.	token_cost_usd plus quality/objective score	Dormant code path; weight defaults to 0.0.	Turn on as a shadow lane with quality floors.

7. Recommended Base Architecture

New Shared Concept: Arena Contract Pack

Do not start with a new service. Start with a shared contract and adapters around existing stores.

Minimum V0 concepts:

Concept	Purpose
`ArenaSpecV0`	Names the arena, capability owner, player roles, state graph, action space, allowed tools, isolation mode, settlement signal, and promotion gate.
`ArenaRunV0`	Records a single trial, inputs, players, side-lane envelope, timestamps, and artifact refs.
`ArenaSettlementV0`	Records mechanical settlement, ObjectiveScore reference, OutcomeEvent reference, and promotion recommendation.
`RefutationRecordV0`	Sidecar to `MoveProposalV0` for CAP-5 contradictor/refuter evidence without mutating frozen proposal v0.
`HITLRewardEventV0`	Normalized operator approve/reject/correct signal for routing, proposals, blueprints, and twin suggestions.

These can initially be documented/contracts-first and mapped to existing stores:

ExperimentStore for arena run records,
SideLaneEnvelope for isolation,
OutcomeEvent for evidence emission,
ObjectiveScoreV1 for scoring,
MoveProposalV0 for self-evolution proposals,
Guardian signed verdicts for security settlement.

Arena Lifecycle

1. Register ArenaSpecV0
2. Select player policies and baseline policy
3. Run isolated side-lane trial
4. Collect mechanical settlement
5. Emit OutcomeEvent and ObjectiveScore
6. Compare against scripted baseline and previous champion
7. Promote, continue shadow, or reject
8. Record replay/audit evidence

Promotion Requirements

An arena can affect product behavior only when all are true:

side-lane isolation has been preserved,
settlement is mechanical and replayable,
baseline comparison is included,
cost and latency are reported,
tenant/ActorContext requirements are satisfied,
operator or policy gate signs off where required,
rollback path exists for any behavior-changing proposal.

8. First Build Sequence

Phase A: Groundwork, No Product Claim

1. Author ArenaSpecV0, ArenaRunV0, ArenaSettlementV0, and RefutationRecordV0.
2. Add adapters from arena run records to ExperimentStore.
3. Define the canonical OutcomeEvent types or payload conventions for arena start/settle/promote/reject.
4. Add a read-only dashboard/table that shows arena runs, players, settlement, score, and status.
5. Keep all players in shadow/side-lane mode.

Phase B: Three Early MVP Arenas

1. CAP-5 Proposer-vs-Refuter. Refuter produces RefutationRecordV0; settlement is later regression, test failure, or operator rejection.
2. HITL Reward Miner. Operator approve/reject/correct signals become reward data for proposals, routing, and blueprint suggestions.
3. Cost Game. Cost-aware utility runs in shadow with cost_utility_weight > 0 only for score reporting, not behavior change.

Phase C: Product-Adjacent Arenas

1. Routing adversary. Generate worst-case historical tasks and evaluate router choices offline with ObjectiveScoreV1.
2. Meta-review. Score reviewers by later confirmed/rejected findings; use results to route future review work.
3. Memory keep/prune. Use retrieval success, contradiction avoidance, and poison quarantine outcomes to train keep/prune decisions.

Phase D: Higher-Leverage Arenas

1. Blueprint tournament. Requires real simulator contracts and what-if evidence.
2. Cross-service causal fabric. Requires reliable mission-level OutcomeEvent replay.
3. Block marketplace. Requires mature telemetry for template reuse, rollback, adoption, and efficacy.

9. Capability Backlog

ID	Work item	Owner capability	Dependency	Acceptance signal
AGNN-01	Arena contract pack: `ArenaSpecV0`, `ArenaRunV0`, `ArenaSettlementV0`	CAP-6/CAP-9	SideLaneEnvelope, OutcomeEvent, ObjectiveScore	Contract parity tests and one sample arena run record.
AGNN-02	`RefutationRecordV0` sidecar for `MoveProposalV0`	CAP-5	MoveProposalV0	Proposal can carry external refuter evidence without changing frozen proposal v0.
AGNN-03	Proposer-vs-refuter shadow arena	CAP-5	AGNN-01/02	Refuter hit/miss measured against later regression/operator rejection.
AGNN-04	HITL reward adapter	Cross-cutting	HITL queue / feedback events	Operator approve/reject/correct normalized into reward events.
AGNN-05	Cost game shadow lane	CAP-10/CAP-6	ObjectiveScore cost fields	Cost-adjusted score reported without changing routing/promotion.
AGNN-06	Meta-review scoring loop	Cross-cutting	meta_swarm advisory output	Reviewer proposals scored against later ground truth.
AGNN-07	Routing adversary offline replay	CAP-4	SUB-9.14, ObjectiveScoreV1	Worst-case task set changes router score distribution measurably.
AGNN-08	Memory keep/prune arena	CAP-2	GraFOSa memory outcomes	Keep/prune choices improve retrieval hit without poison/contradiction increase.
AGNN-09	Blueprint tournament prereq closure	CAP-8	Blueprint simulation contracts	Simulator can emit source/confidence/what-if results before tournament work.
AGNN-10	Guardian policy immutability proof	Guardian/CAP-9	Guardian GatePolicy	Test proves swarm cannot mutate signed production policy.
AGNN-11	Cross-service causal graph spike	Cross-cutting	OutcomeEvent coverage	Mission-level causal chain reconstructs route, cost, artifact, and outcome.
AGNN-12	Strong baseline suite	All arenas	Arena contract pack	Every neural/LLM player is compared against scripted baseline.

10. Explicit Claim Boundaries

Claimable Now

The arena pattern is the correct Conduir abstraction for translating game-player co-learning into product capability.
CAP-6 is the natural factory because ExperimentStore, SideLaneEnvelope, ObjectiveScore, and OutcomeEvent already align with side-lane experimentation.
CAP-5 has a concrete Red-Queen primitive, but it is intentionally shadow/default-off.
The missing contradictor/refuter field is a real CAP-5 opportunity.
Meta-review and cost-aware selection have code surfaces but are not product-active learning loops.
HITL reward mining is likely the fastest useful signal source because operator decisions already exist.

Not Claimable Yet

Full autonomous self-evolution.
GNN-robust security without adaptive attack evaluation.
Blueprint co-learning without a real simulator path.
Routing co-learning without objective score/path-maturity evidence.
Memory keep/prune learning without downstream retrieval and poison/contradiction metrics.
Guardian policy mutation by swarm. This remains forbidden.

11. Decision Recommendation

Build the Arena Contract Pack first, then ship three shadow arenas:

1. proposer-vs-refuter,
2. HITL reward miner,
3. cost game.

This creates the reusable ground for every later AGNN/co-learning capability while staying honest about current code truth. It also avoids the highest-risk mistake: trying to make the neural player the authority. In Conduir, neural or LLM players may compete; substrate, Guardian, operator policy, and mechanical settlement decide.

AGNN Co-Learning Capability Ground Truth