AICF SLA — Latency, Quality, Redundancy, Penalty Schedule

Source: docs/aicf/SLA.md — this page mirrors the repository documentation.

This document defines the Service Level Agreement (SLA) for AI/Quantum providers participating in the AI Compute Fund (AICF). It specifies measured metrics, evaluation windows, target thresholds, penalties, and recovery rules. The evaluator lives in:

• Code: aicf/sla/{metrics.py,evaluator.py,slash_engine.py}
• Policy knobs (network-defined): aicf/config.py, aicf/policy/example.yaml
• Registry/penalties integration: aicf/registry/{heartbeat.py,penalties.py}

All measurements are recorded with monotonic timestamps, aggregated per epoch (configurable block/time length), and persisted for auditability.

---

1) Scope

Applies to all ACTIVE providers (AI and/or Quantum capability). A provider must:

1. Accept assigned jobs (leases).
2. Return verified results/proofs within latency targets.
3. Maintain quality & availability above thresholds.
4. Respect redundancy policies where multiple results are required.

---

2) Metrics (per provider)

2.1 Latency

• Definition: t_complete - t_assign, in seconds, measured server-side on the queue.
• Stats: p50, p90, p95, max.
• Out-of-band caps: absolute timeout job_ttl_s.
• Exclusions: jobs canceled by dispatcher before lease start are ignored.

2.2 Availability

• Definition: completed_jobs / assigned_jobs (within epoch), excluding dispatcher cancels.
• Windowed: exponential decay with half-life H_avail (e.g., 24h).

2.3 Quality (AI)

• QoS Score (0..1): deterministic function of normalized metrics:
  • correctness proxies (checksum, output schema),
  • redundancy agreement (see 2.5),
  • throughput stability (jitter bounds),
  • optional evaluator-specific checks (toxicity guard, etc.).
• Traps Ratio: fraction of known trap prompts answered within acceptable bounds.
  • Threshold: traps_ratio >= traps_min (e.g., 0.985).

2.4 Quality (Quantum)

• Trap Families Pass Rate: pass / total traps (Clifford, T-depth suites).
• Attest Continuity: attestation hash unchanged or re-verified on rotation.
• Shot QoS: variance and success probability within policy bounds.

2.5 Redundancy Agreement (k-of-n)

When policy requests redundancy (e.g., k=2 of n=3):

• Agreement Rate: fraction of jobs where at least k results match within tolerance.
• Tolerance: hash-equality for exact outputs, or metric band (e.g., MSE ≤ ε) for ML.

2.6 Validity

• Proof Validity Rate: valid_proofs / submitted_proofs (post-verification in proofs/*).
• Any invalid proof is treated as a critical fault.

---

3) Evaluation Windows

• Epoch: accounting bucket (e.g., 1 hour or N blocks). Used for payouts and slashing.
• Rolling Windows: EWMA with half-lives H_lat, H_qos, H_avail (policy parameters).
• Grace Period: new providers have grace epochs G (reduced penalties; still measured).

---

4) Targets & Thresholds (defaults; network may override)

# Example policy (see aicf/policy/example.yaml)
sla:
  job_ttl_s: 60
  latency:
    p95_ai_s: 10
    p95_quantum_s: 30
    H_lat_hours: 24
  availability:
    min: 0.98
    H_avail_hours: 24
  ai:
    traps_min: 0.985
    qos_min: 0.95
  quantum:
    traps_min: 0.98
    shot_qos_min: 0.95
  redundancy:
    enabled: true
    k: 2
    n: 3
    agree_min: 0.97
  weights:     # composite score weights
    w_latency: 0.35
    w_avail:   0.25
    w_quality: 0.25
    w_agree:   0.15

---

5) Composite SLA Score

For a provider in an epoch (or EWMA window), compute:

latency_ok  = clamp01( target_p95 / max(eps, p95_obs) )          # higher is better
avail_ok    = clamp01( (avail_obs - min_avail) / (1 - min_avail) )
quality_ok  = clamp01( min(qos_obs / qos_min, traps_obs / traps_min) )
agree_ok    = clamp01( agree_obs / agree_min )

S = w_latency*latency_ok + w_avail*avail_ok + w_quality*quality_ok + w_agree*agree_ok

Where clamp01(x)=min(max(x,0),1). Networks may swap in piecewise penalties for tails.

---

6) Fault Classes & Penalties

Class Trigger (examples) Immediate Action Epoch Penalty (stake/weight) Minor S < 0.9 once; p95 over target by ≤2×; avail in [0.95, 0.98) reduce assignment weight -20% for next epoch Demerit 1; auto-clear after 2 clean epochs Major S < 0.8, avail <0.95, traps below min once, redundancy agree <0.95 weight -50%, temporary jail (1 epoch) Demerit 2, Slash 0.25% of bonded stake Critical Invalid proof, repeated timeouts (≥3 in epoch), forged attestation, lease drop Immediate jail, revoke leases, notify ops Slash 2–5% of bonded stake; cooldown C epochs

Notes:

• Slashing magnitudes and cooldown C are policy-configurable.
• Repeated majors in rolling window may escalate to critical.
• During grace (G epochs), replace slashes with extended jailing unless security-critical.

---

7) Recovery & Appeals

• Auto-Recovery: after a clean epoch (S ≥ 0.95 and no majors), weights gradually restore (+20%/epoch).
• Appeal Window: provider may submit logs and signed metrics within A epochs for re-evaluation.
• Audit Trail: all penalties recorded with reason codes (aicf/sla/slash_engine.py) and hashes of evidence.

---

8) Measurement & Telemetry

Heartbeat payload (aicf/registry/heartbeat.py):

{
  "provider_id": "provider:abc123",
  "attest_hash": "0x…",
  "metrics": {
    "latency_p95_s": 8.7,
    "availability": 0.991,
    "ai": { "qos": 0.964, "traps_ratio": 0.988 },
    "quantum": { "traps_ratio": null, "shot_qos": null },
    "redundancy_agree": 0.975
  },
  "load": { "leases_active": 3, "leases_capacity": 8 },
  "timestamp": 1713202001,
  "signature": "0x<sig>"
}

Dispatcher confirms job outcomes (success/timeout/cancel) and emits per-job records used by the evaluator.

---

9) Redundancy Rules

• When redundancy is enabled (k-of-n):
• All n results are requested in parallel (or staggered).
• Earliest k passing agreement gate settlement; late arrivals still counted for SLA.
• Disagreeing providers are penalized by quality/agree components.

---

10) Example Outcomes

• AI provider, good state: p95=7s, avail=0.995, qos=0.97, traps=0.99, agree=0.98 → S≈0.97 (healthy).
• Quantum provider, slow: p95=45s vs 30s target → latency_ok≈0.67; others OK → S≈0.86 → Minor.
• Invalid proof: immediate Critical → jail, slash, cooldown.

---

11) Penalty Encoding (on-chain/off-chain)

• Reason Codes: SLA_MINOR, SLA_MAJOR, INVALID_PROOF, DROPPED_LEASE, ATTEST_FORGED.
• Record: {provider_id, epoch, reason, delta_stake, cooldown_epochs, evidence_hash}.
• Integration: penalties propagate to aicf/treasury/state.py and registry status.

---

12) Tuning Guidance

• Start with conservative targets; tighten as network stabilizes.
• Prefer weight reductions before slashes unless safety impacted.
• Keep trap suites rotating; publish public seeds post-epoch for audit.
• Separate policy for burst epochs (surge demand) to avoid unfair slashing.

---

13) References

• docs/aicf/OVERVIEW.md, docs/aicf/JOB_API.md
• docs/economics/SLASHING.md
• proofs/* (attest & proof verification)
• aicf/metrics.py (Prometheus: latency histograms, availability gauges)