Diagnosis
- RADIUS, 802.1X, EAP-TLS, PEAP, TACACS, and CoA analysis
- Component health for endpoint, NAS, AAA, directory, DNS, DHCP, NTP, and policy paths
- Named, machine-readable cause families — never just "it's flaky"
TestPulse v1.4 · Self-Learning Diagnostics
The AI learns to explain failed authentication.
TestPulse turns noisy 802.1X, RADIUS, EAP-TLS, and TACACS evidence into one run story — what failed, where it failed on the wire, what proves it, and what to do next. And it gets sharper with every run: a closed reinforcement loop trains it against real, labeled faults until it earns the right to diagnose each cause family on its own.
Learns & graduatesA self-grading loop trains the engine against held-out, labeled faults — it ships a family only after it proves mastery.
Wire-evidence gatedNo verdict without proof: diagnoses are gated on real packet evidence, never on guesses or injected answers.
Forensic OTel traceThe whole AAA path as a distributed trace — the single root-cause hop is flagged with the exact log line and the fix.
Watch it work — four live walkthroughs
From "it's flaky" to the proven root cause — live.
Real devices, real frames on the wire — nothing simulated. Four demos, one story: Demo 1 — 802.1X/RADIUS base-DN root cause with OTel forensic trace. Demo 2 — TACACS+ transport_degradation fault injection → AI verdict. Demo 3 — the full Teaching Cell: Curriculum Trainer, SME Orchestrator, Human Sign-Off, Chain of Custody, ML/RAG Assurance → sealed LoRA adapter on an Acer GN100 running offline. Demo 4 — the Diagnoser Chat copilot: grounded, live Bedrock inference, and a human-approval guardrail on every agentic action.
1 802.1X / RADIUS · directory base-DN root cause · trains the AI · 2:13
2 TACACS+ · transport_degradation · fault-inject → OTel · 1:53
3 The Teaching Cell · Curriculum Trainer · SME seal · DeepSeek-R1 LoRA → Acer GN100 · 4:26
4 Diagnoser Chat · grounded copilot · live Bedrock inference · human-approval guardrail · 1:34
Built for AAA, NAC, and evidence-heavy support workflows.
TestPulse is not generic chat over logs. It is a diagnostics platform with shared evidence contracts, run history, PCAP intelligence, anomaly scoring, and workflow closure — across three surfaces that share one engine.
Forensics
- PCAP digest, pure-Python frame counts, and Wireshark drill-down
- OTel forensic trace: the AAA path as spans with the root-cause hop flagged
- Artifact maps that keep raw proof attached to every finding
Prognosis
- Noise baseline and RADIUS load correlation
- Anomaly vectors, drift detection, and ranked cause families
- Time-series pass-rate trend and NOC dashboard signals
The differentiator
Diagnostics that train themselves — and earn the right to diagnose.
Most "AI for logs" tools guess once and never improve. TestPulse runs a closed reinforcement loop: it injects a real, labeled fault on the wire, diagnoses it blind, auto-grades the verdict against the held-out truth, and reinforces only when it is right. A cause family is not "supported" until the engine proves mastery across six graduation gates.
Injectlabeled fault on real wire
→
Diagnose blindno label leakage
→
Auto-gradevs held-out truth
→
Reinforcegold-anchor the win
→
Graduatefamily proven
Chain of custody on every verdict
Each diagnosis carries a three-layer provenance trail, so a conclusion is defensible — not a black box.
- L1 — kernel witness: the packet capture is the ground truth the AI reasons from.
- L2 — AI reasoning journal: the evidence-to-cause reasoning is recorded, signed, and replayable.
- L3 — human signature: a reviewer vouches the run, sealing the chain for training credit.
Six graduation gates per cause family
A family only ships to diagnosis once the engine clears all six — proof, not vibes.
✓ Breadth ≥3 families
✓ Mastery ≥3 correct
✓ Volume ≥12 runs
✓ Accuracy ≥85%
✓ Stability last-8 ≥85%
✓ Integrity sealed chains
See the failure as a trace, not a log dump.
TestPulse synthesizes the effective AAA path — endpoint → DHCP/DNS → NAS → RADIUS/TACACS → policy → directory → enforcement — into an OpenTelemetry trace. Every hop is a span; the one that broke is the only error span, expanded to show the exact evidence line and the fix direction. Operators stop scrolling logs and start reading a story.
- One span per AAA hop, with timing against per-stage budgets
- Root-cause span carries the proof events and the recommended fix
- Consistent with the diagnosis verdict — the trace and the cause family agree
Faults derived from the network's real intent.
Fault scenarios are driven from the NetBox source of truth, so the lab reproduces drift the way it actually happens in production — a base-DN that moved, a key that changed, a transport that degraded. The injected ground truth is held out as an answer key for grading only; the AI never sees it while it diagnoses, which keeps the learning honest.
- Config-drift faults sourced from the system of record, not hand-written stubs
- Held-out labels score the AI — they are never fed into its reasoning
- Evidence-gated grading: no "false correct," no "false miss"
Feature tiers for pilots, teams, and enterprise labs.
The repo model recognizes core, standard, premium, and enterprise. For the public BMI2 page, packaging is simplified into Core, Premium, and Enterprise.
Evidence Starter
Best for a single team proving the workflow with public AAA evidence.
- Run ingestion and single-run diagnosis with named cause families
- RADIUS / public AAA baseline pack
- Evidence bundle, timeline, component health, artifact map, and service metrics
- CLI, API, MCP, and React Run Viewer access
AI Diagnostics
Best for teams that need learning, comparison, and operational analytics.
- Self-learning reinforcement loop and graduated cause-family coverage
- OTel forensic trace and evidence-gated multi-family diagnosis
- Anomaly engine, RAG / vector retrieval, and cited evidence explanations
- Noise baseline, RADIUS load analysis, prognostics, and trend analytics
Scale-Out Assurance
Best for multi-site labs, customer-safe exports, and proprietary packs.
- Multi-tenant isolation and durable SQLite/Postgres service stores
- Source-of-truth fault validation and full chain-of-custody sealing
- Real SSE streaming, NOC scheduler, retention/audit policy, and support workflows
- Enterprise-only PCAP capture control and proprietary pack entitlements
Where the product is now
The v1.3 production scale-out shipped — and v1.4 adds the self-learning diagnostics layer.
- v1.3 delivered scale-out: tenancy, durable stores, scheduler, SSE, prognosis, error boundaries, and a WCAG baseline.
- v1.4 adds the reinforcement / graduation loop, the OTel forensic trace, and source-of-truth fault validation.
- Diagnosis now spans multiple trained cause families — transport, directory, and NAS/relay among them — each evidence-gated.
- Three surfaces (MCP, REST, React) stay aligned on one evidence contract.
What changed for the landing page
This page leads with the differentiator that competitors don't have: an AI that trains itself to diagnose AAA failures from real wire evidence and graduates per cause family.
- Hero and product preview now show the forensic trace and the chain of custody.
- A dedicated "how it learns" section explains the loop and the six gates.
- Premium and Enterprise tiers now name the learning loop, forensic trace, and SoT validation.
Make this the public TestPulse page on BMI2.com.
Use this as the new product page at https://www.bmi2.com/testpulse/. On GoDaddy cPanel, upload it as /public_html/testpulse/index.html. On GoDaddy Websites + Marketing, rebuild the sections and paste this copy. Questions or a demo — reach out below.