25 production-ready feature issues spanning immediate GPU compute revenue through long-horizon quantum operating systems — with automated TRL, benchmark, and documentation updates on every merge.
Polls CoreWeave, Nebius, Lambda Labs, Vast.ai, Together AI every 15 min. Normalises to $/GPU-hour and $/1M tokens. Publishes to NATS compute.price.updated. Alerts when arbitrage exceeds 20% threshold.
Scores each GPU class against WorkloadProfile (batch_size, seq_len, model_params_b, latency_slo_ms). Routes to lowest-cost GPU satisfying the SLO. Catalogue covers A100 through GB200 and MI300X.
Aggregates spot-price volatility, earnings backlog disclosures, and waitlist signals. Produces weekly CapacityForecastVector per provider × GPU class. Redis TTL 7 days. Companion to contract lock-in advisor.
Scores HW monoculture, provider concentration, leverage risk (debt/revenue), contract lock-in duration, and migration cost estimate. Emits AlertHuman when two or more dimensions breach thresholds simultaneously.
Combines CapacityForecastAgent output with spot/reserved price delta. Computes NPV across 3m, 6m, 12m terms. Recommends Lock In Now / Wait 30 Days / Stay Spot with confidence score and monthly cash-flow table.
Live GPU slot inventory. JobSpec (model, batch, latency SLO, cost target). Ranks slots by cost × SLO satisfaction. Re-dispatches on spot preemption within 30s. Tracks history in TimescaleDB for FinOps.
Normalises CoreWeave $/GPU-hour, Together AI $/1M tokens, Replicate $/second into a single comparable unit. Derives actual cost/token from measured throughput × GPU-hour price for raw providers. 24h Redis TTL on price sheets.
30-day utilisation history from TimescaleDB. Base + burst profile model. Stochastic spot interruption model per provider × GPU class. Outputs OptimalBlend with expected monthly cost and interruption risk score.
Extends TopologyValidatorAgent with continuous monitoring mode during active training jobs. Tracks symbol errors, link flaps, buffer overruns via ibstat/perfquery. Feeds anomaly data to LogIntelligenceAgent for cross-agent correlation.
Egress cost matrix per provider-pair. Estimates data transfer volume (model weights + checkpoints + dataset) for each dispatch decision. Adds egress cost to cross-provider scheduler. Recommends data locality strategy.
Stores SLA terms per provider. Continuous synthetic latency probes every 5 min. 30-day rolling compliance window. SlaBreach NATS event with credit entitlement calculation. Monthly PDF SLA report per provider.
DCGM metrics ingestion: GPU_UTIL, SM_ACTIVE, MEM_COPY_UTIL. Sustained underutilisation detection: <40% for >2h on reserved capacity = waste alert. UtilisationAlert to NATS with waste cost estimate per day.
Aggregates price signals, capacity heat map, vendor risk alerts, fleet exposure %, and quantum backend status into a single GET /v1/market-intelligence endpoint + web dashboard. 15-min Redis cache. This is the product face that converts trial to paid.
QPU Monitoring & Scheduling
Full ORAL lifecycle agent polling backend availability APIs. Composite quality score: Q = fidelity × (1/queue_wait_norm) × availability. Publishes QpuStatusObservation to NATS quantum.qpu.status. AgentRegistry key: "qpu-availability".
Models QPU warm-up overhead from calibration history and queue depth. Predicts estimated_start_latency_s, cold_start_penalty_shots, time_to_peak_fidelity_s. Integrates dispatch penalty into CrossProviderSchedulerAgent scoring.
Runs standard QV circuits weekly on each backend. CLOPS via IBM-standard parametric benchmarking protocol. Alerts on QV halving or CLOPS −20%. TimescaleDB benchmark history. Integrates with QpuAvailabilityAgent de-ranking.
Quantum Error Correction (QEC) — Priority
Accepts QuantumCircuitSpec. Applies surface code overhead formula: physical_qubits = d² × logical_qubits where distance d = ⌈log(p_th/p_phys) / log(p_th/p_L)⌉. Estimates runtime overhead from syndrome extraction cycles. Checks QPU qubit count feasibility. Recommends Feasible / DegradedMode / Infeasible. Supports Surface Code, Steane [[7,1,3]], Bacon-Shor, and Flag Fault-Tolerance for NISQ era.
Wire: QpuAvailabilityAgent → physical qubit count check · QuantumNoiseCharacterisationAgent → measured p_phys (not vendor-claimed) · VariationalQuantumTrainingAgent → QEC-adjusted circuit depth for VQE.
Noise, Algorithms & Variational
Runs RB circuits (Clifford sequences, decay fit F(m) = A·p^m + B). EPG: r = (1-p)(d-1)/d. Per-qubit per-gate error model. Identifies noise hotspots (2Q error > 2× average). Feeds into QEC threshold calculation.
Manages VQE/QAOA optimisation as full ORAL lifecycle. Detects barren plateaus (Var[∂E/∂θ_i] < ε for all i). Switches optimiser: COBYLA → SPSA when gradient noise high. Backend switching on fidelity drop via QpuAvailabilityAgent.
Routes QUBO problems (scheduling, portfolio, NAS, feature selection) to D-Wave Leap, QAOA on gate QPU, or SA fallback. Routes by variable_count vs qubit limit. Tracks quantum advantage ratio vs classical SA baseline.
Classifies WorkloadGraph nodes: QUBO-amenable, VQE-amenable, Quantum-ML-amenable, Classical-only. Estimates quantum speedup and feasibility. Generates HybridDispatchPlan splitting job between GPU and QPU stages.
Transpiles OpenQASM 3.0 to native gate set (CX/RZ/SX for IBM; MS/GPI/GPI2 for IonQ). Reports native_gate_count, circuit_depth, estimated_cost_usd. Budget policy blocks submission if cost exceeds threshold.
Quantum Networks, Cryptography & OS
Discovers QKD nodes, fiber lengths, entanglement swap nodes. Validates QBER < 11% (BB84 security bound). Checks QPU network reachability. Emits QuantumNetworkReport to NATS quantum.network.topology.
Integrates with ETSI-standard QKD KMS. Monitors QBER, key rate, inventory level. Recommends QKD keys when rate > demand; falls back to PQC when inventory low. Publishes quantum.qkd.key_status.
Scans TLS endpoints for cipher suites. Classifies: PQC-Ready / Hybrid-TLS / Classical-Only / No-TLS. Mosca risk: time_to_migrate + data_shelf_life > time_to_CRQC → immediate action. Maps FIPS 203 ML-KEM-768 migration path.
Reads quantum SDK manifest. Checks against compatibility matrix (sdk, version_range, backend, breaking_from). Covers IBM Runtime v1 vs v2 API, Qiskit 1.0 breaking changes, AWS Braket SDK, PennyLane device API v0.38.
Validates QVM config: noise model matches backend calibration, basis gate set consistent, qubit topology (coupling map) matches target. Bell pair test protocol. Memory requirement calculator: 2^n × 16 bytes statevector.
Hybrid TLS: ECDH-P256 + ML-KEM-768 key exchange. ML-DSA-44 for certificate signing. Graceful fallback to classical for legacy clients. Axum middleware layer. Enables DeployIQ to pass vendor security assessments at regulated enterprises.
Each feature advances through TRL 1–9 automatically as GitHub issues are closed and PRs are merged. The CI workflow updates docs/benchmarks/trl.html on every qualifying merge and redeploys to GitHub Pages.
| TRL | Gate Condition | Features at This Level | Progress |
|---|---|---|---|
| TRL 1 | Issue filed, concept defined | All 25 new features (#1324–#1354) · 14 existing agents at higher TRL |
|
| TRL 2 | Design doc merged to docs/ |
State vector layout defined · API route spec written · Acceptance criteria reviewed |
|
| TRL 3 | PR opened, unit tests present | Agent struct + ORAL lifecycle stub · State vector populated · Unit tests >70% coverage |
|
| TRL 4 | PR merged, CI green, all unit tests pass | cargo test passes · No compile warnings · Agent registered in AgentRegistry::standard() |
|
| TRL 5 | Integration tests pass with mock infrastructure | NATS publish verified · API route e2e tested · Heuristic fallback confirmed · Semaphore respected |
|
| TRL 6 | Staging deployment, live infrastructure tested | Feature tested against real CoreWeave/Nebius API or QPU backend · DRP voting verified end-to-end |
|
| TRL 7 | Design partner pilot, feedback incorporated | Feature shipped to ≥1 design partner · Feedback addressed · Performance benchmarks published |
|
| TRL 8 | Production deployment, SLA met for 30 days | Feature in continuous production use · p99 latency within SLO · Zero P0 incidents in 30-day window |
|
| TRL 9 | Multiple production customers, benchmark published | Feature deployed at ≥3 paying customers · Benchmark report updated · Revenue attributed in FinOps report |
| Agent | Type | TRL | Status |
|---|---|---|---|
| CanaryVerificationAgent (CVA) | ai-infra | TRL 5 | Unit + integration tests pass · staging pending |
| LogIntelligenceAgent (LIA) | ai-infra | TRL 5 | Threshold tests against LiaConfig::default() ✓ |
| ConfigDriftAgent (CDA) | ai-infra | TRL 5 | serde_json::Map fix · CI green |
| TopologyReasoningAgent (TRA) | network | TRL 5 | Config fix · 32/32 tests pass |
| RollbackOrchestratorAgent (ROA) | ai-infra | TRL 5 | AtomicBool double-rollback guard ✓ |
| PolicyComplianceAgent (PCA) | compliance | TRL 5 | Functional audit passed |
| MigrationVerificationAgent (MVA) | db | TRL 5 | instrument field fix · clean build |
| PreflightAgent (PFA) | gpu | TRL 4 | Semaphore bulkhead ✓ · staging pending |
| CanaryAiAgent (CVA-AI) | ai-infra | TRL 4 | CI green |
| TrainingWatchdogAgent (TWD) | ml | TRL 4 | Semaphore ✓ |
| LemonNodeDetectorAgent (LND) | infra | TRL 4 | Semaphore ✓ |
| FinOpsAgent | cost | TRL 4 | act() fixed · semaphore added |
| TopologyValidatorAgent | network | TRL 4 | AgentType fix + act() + semaphore ✓ |
| CloudParityAgent | multi-cloud | TRL 4 | AgentType fix + semaphore ✓ |
| PipelineGateAgent | ci-cd | TRL 4 | StateVector dims 0–5, semaphore ✓, registry ✓ |
Every feature in the roadmap advances automatically through TRL levels as code is merged. No manual documentation updates required. The pipeline reads issue labels and closed state from the GitHub API, updates the TRL table and benchmark report, and redeploys to GitHub Pages.
Automation Triggers
feature closed → feature advances to TRL 4 (if linked PR merged) or TRL 2 (if closed without PR)feature, computes their TRL level from their state and linked PR status, and patches docs/benchmarks/trl.html with the updated table. Commits and pushes, then the benchmark-deploy workflow picks up the file change and deploys to Pages.fetch-depth: 0GET /repos/{owner}/{repo}/issues?labels=feature&state=all&per_page=100docs/scripts/update-trl.sh to patch the TRL table HTML in-placedocs/scripts/update-benchmark.sh to add feature to benchmark report with current TRL"chore(trl): auto-advance #{issue_number} to TRL {level} [skip ci]"benchmark-deploy.yml → Pages redeployed within 2 minsed to find the feature row in trl.html by issue anchor and updates the TRL badge class, progress bar width, and status text. Idempotent — safe to run multiple times.ISSUE_NUM, NEW_TRL, STATUS_TEXT from environment variablesgrep -n "issues/${ISSUE_NUM}" docs/benchmarks/trl.htmltrl-{old} class with trl-{new}$((NEW_TRL * 11))%STATUS_TEXTlast_updated meta tag to current UTC timestampdocs/benchmarks/benchmark-latest.html. Each feature shows: issue number, title, TRL level, implementation status, linked PR, and performance metrics (once available from CI). Weekly cron refreshes market data rows (GPU prices, capacity signals) from public APIs.Document Lineage — What Auto-Updates