Implementation Changelog
Cross-reference between the GIM Architecture Spec and what’s been built. Updated April 12, 2026. Coverage: ~70% of spec implemented, with significant beyond-spec additions.Phase Status
| Phase | Name | Spec Status | Implementation |
|---|---|---|---|
| P0 | Foundation | 7/7 | Proto codegen, shared packages, DB migrations, 12 capability dimensions |
| P1 | Runtime Tier 0 | 9/10 | Gate 0, classifier, security profiles, canaries, invariants. Dashboard UI remaining |
| P2 | Endpoint Distribution | 9/9 | 18MB binary, install script, Docker, Tailscale, fleet, offline buffer, auto-update |
| P3 | Behavioral Tier 1 | 11/12 + redesign | Fingerprint, Gates 1-3, flow sketches, group envelopes, shadow mode, threat detections |
| P4 | Multi-Proxy Sync | 9/10 | Serialization, Redis, NATS, BI Service, drift detection. LLM classification deferred |
| P5 | GNN Tier 2 | 10/10 | Complete in v1.0.0 — Memgraph, VGAE, GNNScorer, signature distillation |
| P6 | Compliance | 0/7 | Not started |
| P7 | Adaptive | 0/7 | Not started |
Implemented — Matching Spec
| Component | Spec | Actual | Status |
|---|---|---|---|
| 3-Tier Architecture | Tier 0/1 local, Tier 2 remote | Proxy (Go) + BI Service (Python) | Exact match |
| AgentFingerprint | ~3.1KB probabilistic struct | ~6KB with flow sketches (v2) | Extended |
| Capability Distribution | [12]float32 | [12]float32 | Exact |
| Tool Frequency | CMS 4x256 uint16 | CMS 4x256 uint16 | Exact |
| Hierarchical Bloom | 3 levels (domain/server/tool) | 3 Bloom filters | Exact |
| Temporal Profile | 24h EWMA + interval | HourlyActivity[24] + IntervalEWMA | Exact |
| Risk Baseline | Welford’s algorithm | WelfordStats (mean, m2, min, max) | Exact |
| Sequence Model | Markov chain | 32-slot hash-based MarkovChain | Exact |
| Cardinality | HyperLogLog | 3x HLL (tools, servers, IPs) | Exact |
| Gate 1 | Envelope membership | Bloom + CMS + JSD, 114ns | Exact |
| Gate 2 | 6 deviation signals | All 6 implemented, 433ns | Exact |
| Gate 3 | Corroboration + session | Flow divergence + structural evidence | Redesigned |
| Confidence Bands | KNOWN_SAFE / UNCERTAIN / ANOMALOUS | 3-band classification | Exact |
| Delta Flush | 30s to Redis | DeltaFlusher, 30s periodic | Exact |
| Session Export | On session end | 60s periodic + session end | Extended |
| Cold Start Bootstrap | From Redis, <5s | BootstrapFromRedis, ~230ms for 10K | Exceeds |
| BI Service | Single Tier 2 authority | Python, NATS consumer, 780K events/sec | Matches |
| Correction Flow | Baselines + corrections via NATS | CorrectionReceiver with ApplyBaseline/ApplyCorrection | Matches |
| Shadow Mode | Observe-only before enforcement | BehavioralConfig with shadow_mode toggle | Exact |
Implemented — Diverged from Spec
| Spec Said | What Was Built | Why |
|---|---|---|
| Session buffer with 4 hardcoded relationship types | Flow matrix accumulator ([12][12]uint16) | Hardcoded patterns required code changes for new attacks. Flow matrices detect structural divergence without enumerating specific capability pairs |
| Gate 0 in behavioral pipeline (~200ns) | Gate 0 logic lives in gateway/intercept layers | Deny lists and rate limiting are pre-behavioral concerns. The behavioral pipeline starts at Gate 1 |
| Redis Stack server-side merge (BF.ADD, CMS.INCRBY) | Go-side GET → unmarshal → MergeFingerprints → SET | Simpler, avoids Lua script complexity. Convergent structures (CMS, Bloom, HLL) tolerate last-write-wins |
| Fingerprint deltas via dedicated NATS stream | Deltas go directly to Redis via DeltaFlusher | More efficient when the merge target is Redis — no intermediate NATS hop needed |
| ~3.1KB per fingerprint | ~6KB per fingerprint | Flow sketches (FlowMatrix, ResourceFlowSketch, DepthProfile, TemporalFlow) added ~2.9KB. Still fits 21K agents in 128MB |
Implemented — Beyond Spec
These features were NOT in the original GIM Architecture doc but were added during implementation:| Feature | What It Does | Why It Matters |
|---|---|---|
| Baseline Floors | Minimum thresholds that EWMA learning can never lower | Anti-poisoning: prevents slow envelope corruption |
| Threat Signatures | 5 built-in structural FlowMatrix shapes matched via JSD | Detection at Tier 0 without cloud — exfiltration, credential relay, staging, recon, cover tracks |
| Signature Scoping | AgentTypes + MinDepth fields per signature | Different agents have different normal flows — sub-agent depth matters |
| TightenFloors Protocol | BI Service can only tighten floors, never relax | Prevents compromised BI channel from weakening local defenses |
| Signature Validation | MaxJSD, Weight, FlowShape validated on Add/LoadFromJSON | Prevents crafted signatures from blinding detection |
| Flow Matrix EWMA | [12][12]float32 capability transition tracking | Structural behavioral shape beyond individual action statistics |
| Resource Flow Sketch | CMS64 tracking resource boundary crossings | Detects data moving in unexpected directions |
| Depth Profile | [12][8]float32 capability × depth tracking | Detects capabilities at unusual nesting depths |
| Temporal Flow | Per-transition timing EWMA + variance | Detects timing anomalies on specific flow types (not just globally) |
| Noise Budget Counters | Atomic band distribution tracking with health status | ”healthy” / “noisy” / “blind” classification for calibration |
| Session Reaper | TTL-based eviction of stale sessions (30min idle) | Prevents unbounded memory growth in sessions sync.Map |
| GroupID Filtering | computeGroupEnvelope filters by agent type | Correct group baselines instead of merging all agents |
| PeriodicWorker | Extracted shared ticker/goroutine lifecycle | DRY pattern across DeltaFlusher, SessionExporter, DriftDetector |
| NATS Overflow Buffer | Disk-backed JSONL overflow when NATS backpressures | Prevents event loss during network outages |
| Agent Eviction (BI) | Stale agent stats evicted after 1hr idle | Prevents unbounded memory growth in BI Service |
| Calibrator Statistics | False negative/positive/agreement rate tracking | Observability for shadow mode calibration |
Not Yet Implemented (Spec’d)
| Spec Item | Phase | Dependency | Status |
|---|---|---|---|
| GNN Structural Analysis | P5 | Memgraph + model training | Not started |
| Graph Autoencoder | P5 | Memgraph | Not started |
| Memgraph Deployment | P5 | Infrastructure | Not started |
| LLM Fallback (Tier 2) | P5 | BI Service + quint-graph RAG | Infrastructure exists, not wired |
| 5 Structural Invariant Detectors | P5 | GNN | Not started |
| Predictive Detection | P5 | GNN | Not started |
| Contrastive Refinement | P5 | GNN + labeled data | Not started |
| Threat Feed Integration (MITRE ATT&CK) | P5 | External feeds | Not started |
| Compliance Frameworks (SOC2, NIST, ISO) | P6 | Schema design | Not started |
| Compliance Dashboard | P6 | API + frontend | Not started |
| Audit Report Export | P6 | Compliance data | Not started |
| Federated Intelligence | P7 | Global BI infrastructure | Not started |
| Threat Simulation Engine | P7 | GNN + LLM | Not started |
| Dynamic Confidence Thresholds | P7 | Calibration data | Not started |
| Active Learning Loop | P7 | Analyst feedback | Not started |
| LLM Classification Pipeline | P4 | Infrastructure design | Deferred (QDEV-133) |
Performance vs Spec Targets
| Metric | Spec Target | Actual | Verdict |
|---|---|---|---|
| Tier 0+1 Latency | <1-5ms | ~2-5μs (1000x faster) | Exceeds |
| Tier 2 Latency | 10-50ms async | BI Service via NATS (non-blocking) | Meets |
| Fingerprint Update | O(1) | 212ns, all O(1) | Meets |
| Gate 1 Fast Path | ~100ns | 114ns, 0 allocs | Meets |
| Gate 2 Signals | ~500ns | 433ns, 8 allocs | Meets |
| Gate 3 Corroboration | ~1μs | 293ns (no-evidence), 1.2μs (with flow JSD) | Meets |
| Cold Start | <5s for 10K agents | ~230ms | Exceeds (20x) |
| Delta Flush | 30s interval | 30s (exact) | Meets |
| Memory per Agent | ~3.1KB → ~6KB | 5,990 bytes serialized | Meets revised |
| Cache Budget | Not specified | 128MB default (~21K agents) | Reasonable |
| Noise Budget | 95% Band 1, <0.5% Band 3 | Scale test: 85.6% Band 1, 2% Band 3 | Close (needs tuning) |
| False Positive Rate | Not specified | 0.04% (flow-based Gate 3) | Strong |
| BI Service Throughput | Not specified | 780K events/sec (batch) | Strong |
| BI Service p99 | Not specified | 0.002ms | Strong |
Known Issues
| Issue | Severity | Status |
|---|---|---|
| AgentFingerprint data race (no per-instance mutex) | Medium | Tracked — concurrent Score + ForEach can race on fingerprint fields |
| Noise budget not hitting 95% KNOWN_SAFE target | Low | Needs threshold tuning in production with real traffic |
| Health tests fail on Python 3.14 (aiohttp compat) | Low | aiohttp test client needs update |
| Threat signatures: only recon (QT-SIG-004) fires reliably in tests | Low | Other signatures need more diverse test scenarios |
Codebase Statistics
| Component | Files | LOC | Tests |
|---|---|---|---|
| Proxy behavioral package (Go) | 31 source, 28 test | 18,534 | 321 |
| BI Service (Python) | 10 source, 9 test | ~2,558 | 122 |
| API handlers + store | 15 handler, 10 store | ~3,000 | ~50 |
| Pipeline writer | 1 | ~460 | ~20 |
| Migrations | 12 | 1,115 | — |
| Mintlify docs | 13 pages | ~3,000 | — |
| Total | ~130 files | ~28,700 | ~513 |