Imported from ashfordeOU/aero-agent-skills (skills/systems-engineering-safety/arp4761a/event-tree-analysis/SKILL.md). Install upstream with npx skills add ashfordeOU/aero-agent-skills --skill event-tree-analysis. Copyright stays with the author (Apache-2.0).

ARP4761A Event-Tree Analysis (systems-engineering-safety/arp4761a/event-tree-analysis)

Use when the task is forward event-tree analysis of an aircraft system safety case per ARP4761A: starting from one initiating event, enumerate every binary branch path through an ordered list of mitigating functions, roll up each end-state outcome frequency, sum the frequency of the failure end state, and screen the dominant sequences against the per-flight-hour severity targets. This leaf is the forward dual of fault tree analysis: FTA works backward from a top event through gates to basic events, while this leaf works forward from an initiator through successive mitigating-function branches. It consumes the FHA severity class of each end state (rated by functional-hazard-assessment, never re-derived here) and feeds end-state frequencies into the SSA close-out work. Pairs with systems-engineering-safety/arp4761a/fta-fmea for the backward side and with systems-engineering-safety/arp4761a/particular-risk-analysis for single-event forward risk that does not branch.

Domain quick reference

• An event tree starts from one initiating event with a per-flight-hour frequency q0 and lets each ordered mitigating function (for example fire detection, extinguisher discharge, pilot action) either succeed with probability p or fail with probability 1 - p.
• Full binary expansion: N branch nodes give 2N end-state paths. Enumeration is the ascending binary mask with node i mapped to bit i, so the all-failure path (mask 0) comes first and the all-success path (mask 2N - 1) comes last. The cap is BRANCH_NODES_MAX = 12 nodes = 4096 paths; 13 nodes (8192 paths) raise ValueError.
• Path probability: the product of the branch probabilities along the path, p for a success branch and 1 - p for a failure branch. The probabilities of the full expansion sum to exactly 1.0.
• End-state frequency: frequency = q0 x path probability for every end-state path; the frequencies of the full expansion sum to exactly the initiator frequency q0 (the expansion partitions the initiator).
• Failure end state: under the series-barriers reading, the path where no mitigating function contained the initiator (every outcome False). Its frequency is the top-function failure frequency, the undetected and uncontained end state that the mitigating chain must push below its severity target.
• Dominant sequence: an end-state sequence whose frequency strictly exceeds the severity target of its FHA class (frequency > target; equality is NOT dominant). Ratio = frequency / target.
• ARP4761A-class per-flight-hour severity targets, magnitude only and mirrored from the FHA probability-target mapping: CATASTROPHIC 1e-9, HAZARDOUS 1e-7, MAJOR 1e-5, MINOR 1e-3. The class of each end state is an analyst input from the FHA and is never derived in the module: screen each end state against the target of its own class.
• Deterministic binary enumeration only: no Monte Carlo, no time integration, no analyst weighting inside the rollup.

Workflow

1. Fix the initiating event: state the initiator and its per-flight-hour frequency q0 from the prior analysis (for example an FHA or preliminary assessment). q0 must be non-negative; outcome_frequencies rejects a negative initiator frequency with ValueError.
2. List the ordered mitigating functions: build the branch node list of (name, p_success) tuples in the order the functions act; the failure probability of each function is 1 - p_success. At most BRANCH_NODES_MAX = 12 nodes keep the expansion at 4096 paths; the boundary probabilities 0.0 and 1.0 are legal.
3. Enumerate the binary branch tree with build_paths: the full 2N expansion from the all-failure path (mask 0, first) to the all-success path (mask 2N - 1, last). Every end-state path carries its readable sequence string (name:S or name:F per node in node order), its bool path (True = the function succeeds) and the branch-probability product along the path.
4. Roll up and rank the end-state frequencies with outcome_frequencies(q0, nodes): frequency = q0 x path probability for every end-state path, ranked descending by frequency with ties broken by enumeration order (stable sort).
5. Sum the failure end-state frequency with is_failure_end_state and top_function_failure_frequency: the all-failure path is the failure end state where no mitigating function contained the initiator; the function returns its sequence and the sum of the frequencies of the paths reaching it.
6. Screen the ranked sequences with dominant_sequences: pass the per-flight-hour severity target of each end state's FHA class (CATASTROPHIC 1e-9, HAZARDOUS 1e-7, MAJOR 1e-5, MINOR 1e-3, the magnitude targets only). A frequency strictly above its class target flags the sequence dominant with ratio = frequency / target; an end state sitting exactly on its target is not dominant.
7. Judge the mitigating chain: any dominant sequence, or a top-function failure frequency above its class target, means the chain alone does not meet the safety objective for the initiator. Add an independent barrier or lower the initiator frequency and re-run steps 2 to 6.
8. Confirm deterministic behavior with the contract test: python3 scripts/test_event_tree_analysis.py.

Worked example

Cargo-compartment fire. Initiating event: cargo fire ignition at initiator_frequency q0 = 3e-5 per flight hour. Three mitigating functions in series, branch nodes [("detect", 0.95), ("extinguish", 0.90), ("pilot", 0.80)]: fire detection, extinguisher discharge, pilot action. Failure probabilities are 0.05, 0.10 and 0.20. Real module outputs from the smoke run (python3, stdlib, deterministic):

Path probabilities in enumeration order (module sum exactly 1.0): detect:F extinguish:F pilot:F 0.001, detect:S extinguish:F pilot:F 0.019, detect:F extinguish:S pilot:F 0.009, detect:S extinguish:S pilot:F 0.171, detect:F extinguish:F pilot:S 0.004, detect:S extinguish:F pilot:S 0.076, detect:F extinguish:S pilot:S 0.036, detect:S extinguish:S pilot:S 0.684. Each raw float sits within 1e-15 of the products of the decimal inputs, so the displayed values are the module outputs rounded.

outcome_frequencies(3e-5, nodes) ranked descending (module frequency sum exactly 3e-05 = the initiator):

• detect:S extinguish:S pilot:S probability 0.684, frequency 2.052e-5
• detect:S extinguish:S pilot:F probability 0.171, frequency 5.13e-6
• detect:S extinguish:F pilot:S probability 0.076, frequency 2.28e-6
• detect:F extinguish:S pilot:S probability 0.036, frequency 1.08e-6
• detect:S extinguish:F pilot:F probability 0.019, frequency 5.7e-7
• detect:F extinguish:S pilot:F probability 0.009, frequency 2.7e-7
• detect:F extinguish:F pilot:S probability 0.004, frequency 1.2e-7
• detect:F extinguish:F pilot:F probability 0.001, frequency 3e-8

top_function_failure_frequency returns {"sequences": ["detect:F extinguish:F pilot:F"], "frequency": 3e-8}: the undetected, unsuppressed fire end state sits at 3e-8 per flight hour, which exceeds the catastrophic target 1e-9 by a factor of 30 (module ratio 30.000000000000014, within 1e-9 of 30), so this chain alone does not meet the safety objective and the design needs an additional independent barrier or a lower initiator frequency.

Example FHA class input for the screening (analyst rating by consequence belongs to the FHA, not the module): extinguish success rates the end state minor; extinguish failure with detect success rates major; extinguish failure with detect failure rates hazardous when the pilot action succeeds and catastrophic when it fails. Screening each end state against the target of its own class with dominant_sequences (real module verdicts):

• detect:F extinguish:F pilot:F vs CATASTROPHIC 1e-9: 3e-8 > 1e-9, DOMINANT, ratio 30.0 (mitigation required).
• detect:F extinguish:F pilot:S vs HAZARDOUS 1e-7: 1.2e-7 > 1e-7, DOMINANT, ratio 1.2 (module 1.200000000000001, thin margin, needs justification or improved detection and suppression).
• detect:S extinguish:F pilot:S and detect:S extinguish:F pilot:F vs MAJOR 1e-5: 2.28e-6 and 5.7e-7, not dominant.
• the four extinguish-success end states vs MINOR 1e-3: largest is the all-success chain at 2.052e-5, not dominant.

Global screening caution from the same run: screening the whole ranked list against MAJOR alone flags only the all-success end state (2.052e-5

1e-5, module ratio 2.052), which is not the dangerous sequence; each end state must be screened against the target of its own FHA-rated class.

Verification

• build_paths over the fire nodes returns 8 paths in the documented binary-mask order and the path probabilities sum to exactly 1.0.
• outcome_frequencies(3e-5, fire nodes) returns the ranked list above; the frequencies sum to exactly 3e-05 and each entry carries exactly the keys sequence, path, probability, frequency.
• A single node at p = 0.5 and initiator 1e-5 splits evenly into g:F 5e-6 then g:S 5e-6 (ties keep enumeration order).
• Boundary probabilities are legal: p = 1.0 puts the whole initiator on the success end state, p = 0.0 on the failure end state.
• The 12-node all p = 0.5 tree enumerates 4096 paths with all-success probability exactly 2**-12 = 0.000244140625 and max frequency 2**-12.
• The failure end-state frequency 3e-8 matches the initiator x (0.05 x 0.10 x 0.20) identity within 1e-15; the dominant ratios 30.0 and 1.2 hold within 1e-9.
• ValueError rejection of non-physical inputs: negative initiator frequency, empty node list, 13 branch nodes, branch probability 1.5 or -0.1, and severity target 0 or negative all raise ValueError.
• Run the contract test offline: python3 scripts/test_event_tree_analysis.py (33 tests, deterministic, all pass in under a second).

Related leaves

• systems-engineering-safety/arp4761a/fta-fmea: the backward dual, top-event probability and cut sets from AND/OR gate structures over basic events.
• systems-engineering-safety/arp4761a/functional-hazard-assessment: the FHA severity classes that each end state consumes as input.
• systems-engineering-safety/arp4761a/ssa-closure: closes the SSA over the assessed conditions using the per-flight-hour severity targets.
• systems-engineering-safety/arp4761a/particular-risk-analysis: single-event forward risk combination without branch trees.
• systems-engineering-safety/arp4761a/markov-analysis: time-domain state probabilities from transition rates, the neighbor this leaf does not overlap.

Pitfalls

• Screening the whole ranked list against one target: in the worked example a single MAJOR screen flags only the all-success end state (2.052e-5 > 1e-5), which is not the dangerous sequence, while the catastrophic failure end state at 3e-8 sails under that screen. Each end state must be screened against the target of its own FHA class.
• Reading equality with the target as dominant: the comparison is strict, frequency > target, so an end state sitting exactly on its severity target is not flagged and still needs a margin argument.
• Re-ordering or dropping a mitigating function after the rollup: the failure end state is the all-failure path of the exact node list in force; every barrier added multiplies the failure end-state frequency by its own 1 - p and every re-order changes the sequence strings, so re-run the enumeration after any chain change.
• Misreading the mask order: node i maps to bit i of an ascending binary mask, so mask 0 (all failure) is the FIRST path and mask 2**N - 1 (all success) is the LAST; a reversed reading swaps the failure and success end states.
• Letting the module invent severity classes: the class of each end state is an analyst input from the FHA (extinguish failure with detect failure rates hazardous or catastrophic by consequence); the module only screens against the target it is handed.
• Reaching for this leaf outside its scope: backward fault-tree gates, cut sets and top-event logic from basic events, lognormal uncertainty bands, and Markov time evolution belong to the sibling leaves listed above, not to the forward branch rollup.

Contract test

The contract test scripts/test_event_tree_analysis.py (stdlib unittest, offline, deterministic) exercises the whole workflow: the binary branch-tree enumeration and probability table of step 3, the ranked end-state frequency rollup and exact partition identities of step 4, the failure end-state frequency sum of step 5, the dominant-sequence screening with ratios and the strict-equality rule of step 6, every ValueError rejection of the validation list, the 12-node expansion cap, and determinism. Run it from the repo root:

python3 skills/systems-engineering-safety/arp4761a/event-tree-analysis/scripts/test_event_tree_analysis.py

Behavior contract (gate 3)

The event-tree logic is exercised by the gate 3 contract test: scripts/test_event_tree_analysis.py against scripts/event_tree_analysis_logic.py (stdlib unittest, offline). Run:

python3 scripts/test_event_tree_analysis.py

from the leaf directory, or the full path from the repo root as shown in the Contract test section. Exit 0 with 33 tests passing satisfies the behavior contract gate.

Compliance

• Standards referenced, not reproduced: ARP4761A is proprietary SAE guidance (reference-only per standards-map.yaml); this leaf states engineering methodology in its own words, summary-only, and the severity targets appear by magnitude only as module constants.
• The event-tree method is the forward dual of fault tree analysis as practiced in system safety assessments; no verbatim standard text is included.
• compliance: STANDARDS-REF, gated: false.