How to Solve Anything

A structured root cause analysis (RCA) framework combining the 5 Ws and 5 Whys.
Probabilistic, formally proven, and domain-agnostic — engineering incidents, medicine, security, business, legal.

SOMETHING HAPPENED → TO SOMEONE → SOMEWHERE → AT SOME POINT → FOR SOME REASON

Not a template. An algorithm.
Bayesian probability at every node. Causal inference at every edge. LLM-assisted or fully manual.

📋 Framework   ·   🧮 The Math   ·   🔬 Proofs   ·   📊 Diagrams   ·   📖 Docs   ·   📚 Research

---

💡 The Core Insight

Every incident, outage, failure, or decision has the same anatomy.
The vocabulary changes. The causal structure never does.

The 5 Ws tell you what happened.
The 5 Whys tell you why it happened.
Together they tell you what to fix — and in what order.

HTSA replaces the blank postmortem doc and the sticky-note 5 Whys session
with a formal algorithm: DAG traversal + Bayesian updating + causal counterfactual tests.

---

🔍 The Four Layers

Layer 1 — Situation Map (5 Ws)

Establish the full picture before drilling into cause.

Question	What It Captures
Who	The actor, subject, or stakeholder involved
What	The event, problem, or incident
When	The timeline — before, during, and after
Where	The location, system, environment, or context
Why	The surface-level, immediately apparent reason

Layer 2 — Causal Chain (5 Whys)

Start at the surface Why. Ask why again.
Keep going until you hit something you can actually change.

Why (surface)
  └─► Why 1
        └─► Why 2
              └─► Why 3
                    └─► Why 4
                          └─► Why 5: ROOT CAUSE

Whys can and should branch. Real problems are rarely single-cause.

Layer 3 — Resolution

Map each root cause to a concrete change. Apply the counterfactual test:

"If this change had existed before the problem occurred,
would the problem still have happened?"

Each root cause is either fixed, mitigated, or accepted.

Layer 4 — Verification and Learning

Confirm the fix worked. Update your priors.
The framework compounds over time — but only if learning is explicit.

---

🌐 Works Everywhere

Domain	Who	What	When	Where	Why
⚙️ SRE / Engineering	System / Team	Outage, incident	Incident timeline	Service / Component	Alert or error
🏥 Medicine	Patient	Diagnosis	Onset	Body system	Presenting symptom
🔒 Security	Threat actor	Breach	Attack window	Vulnerability	Attack vector
📈 Business	Team / Process	Bottleneck	Quarter	Department	Stated reason
⚖️ Legal	Defendant	Act	Date	Jurisdiction	Motive
🧠 Personal	You	Decision	Moment	Context	Emotion

Each domain uses the same algorithm. The math is domain-agnostic.

---

🧮 The Math

The framework is an applied graph traversal algorithm for causal inference —
with probability weighting, entropy reduction, and Bayesian evidence updating at every node.

#	Concept	What It Answers
01	Graph Theory	What is the structure of an investigation?
02	Exponential Problem Space	Why do investigations feel overwhelming?
03	Causal Inference	How do you prove something caused something else?
04	Information Theory	How do you measure investigative progress?
05	Bayesian Reasoning	How do you weigh competing causes?
06	Search Algorithms	How do you move through the Why tree?
07	Cognitive Biases	What corrupts the investigation?
08	Evidence Evaluation	How do you know which evidence to trust?
09	Causation Theory	How do you classify and quantify actual causes?
10	Intervention Theory	How do you find the minimal set of fixes?

---

📏 Rules

Map before you drill. Complete the 5 Ws before starting the 5 Whys.

Evidence at every node. An assertion without evidence is a guess — tier your evidence.

Branch when reality branches. Real incidents have multiple root causes. Follow all of them.

5 is a heuristic, not a rule. Stop when you reach something you can actually change.

The counterfactual test closes the loop. If the fix had existed, would the incident still have happened? If yes, go deeper.

The framework is recursive. A root cause can become a new incident. Run it again.

---

⚙️ Engine

The framework is also available as a Python library (v2.0.0) with built-in LLM integration.
Works with any provider — OpenAI, Anthropic, Groq, Mistral, Ollama, or any OpenAI-compatible endpoint.

cd engine && uv run python -c "from htsa_engine import Investigation; print('ready')"

Auto-investigate with any LLM — one call, all 4 layers:

from htsa_engine.llm import LLMAdvisor

advisor = LLMAdvisor("https://api.openai.com/v1", api_key="sk-...", model="gpt-4o")
inv = advisor.run("API returning 500 errors since 2:47 AM, EU region only")

print(inv.root_causes)
inv.save("investigation.json")

Or drive it manually — full control over every decision:

from htsa_engine import Investigation, Evidence, EvidenceTier, EvidenceDirection

inv = Investigation(title="API 500 errors", pruning_threshold=0.05)
inv.set_situation(who_affected="Users", what="500 errors", when_during="2:47 AM", where="EU-west", why_surface="Load spike")
inv.complete_situation()
origin = inv.start_causal_chain("Server errors under load")
branch = inv.add_hypothesis(origin, "Memory leak", probability=0.6)
# ... add evidence, mark root cause, resolve, verify
inv.save("investigation.json")

v2 — Causation analysis — quantify and prioritize root causes:

# HP2015 + NESS three-stage counterfactual test
result = inv.run_hp2015_test(branch, origin)
print(result.is_root_cause, result.w_partition)

# Probability of Necessity and Sufficiency
pns = inv.compute_pns(branch, pn=0.8, ps=0.7)
print(pns.causation_type)  # "single_root_cause" | "and_node" | "or_node"

# Find the smallest set of fixes that achieves 90% coverage
intervention = inv.compute_minimal_intervention_set(theta=0.90)
print(intervention.minimal_set, intervention.coverage)

# Evidence budget — how many Tier-1 evidence items are needed?
budget = inv.evidence_budget(branch, alternative_posteriors={"other_node": 0.3})
print(budget.n_required, budget.is_indistinguishable)

Full engine documentation →

---

---

🔍 How HTSA Differs

	HTSA	PyRCA / BARO	DoWhy	Postmortem templates
Approach	Structured algorithm	ML / metrics	Statistical	Blank form
Input	Any problem	Prometheus metrics	Data frames	Text
Causal proof	HP2015 + NESS + PNS	Correlation-based	do-calculus	None
Works without data	Yes	No	No	Yes
Cross-domain	Yes	AIOps only	Research only	Yes
LLM integration	Built-in	No	No	No

---

MIT Licensed · Contributions welcome