diff --git a/docs/specs/complex-synthetic-benchmark-design.md b/docs/specs/complex-synthetic-benchmark-design.md
new file mode 100644
index 0000000..8be6057
--- /dev/null
+++ b/docs/specs/complex-synthetic-benchmark-design.md
@@ -0,0 +1,281 @@
+# Complex Synthetic Benchmark Design
+
+## Purpose
+
+This document designs a future local-only synthetic IAMScope benchmark/demo fixture named
+`complex_shared_uncertainty_iam_benchmark_001`.
+
+The purpose is to make the public demo more representative without adding live
+AWS evidence. The future fixture should test whether IAMScope can locally and
+deterministically separate validated, blocked, precondition-only, and
+inconclusive paths in a more complex synthetic IAM graph, while also grouping
+shared uncertainty classes that affect multiple path-shaped candidates.
+
+This is a design/spec slice only. It does not add fixture JSON, tests, runner
+integration, reasoner changes, live AWS behavior, benchmark semantics, or new
+evidence claims.
+
+## Reviewer-Facing Claim Boundary
+
+The future fixture may support one narrow reviewer-facing story:
+
+IAMScope reduces overconfident path interpretation by separating validated,
+blocked, precondition-only, and inconclusive paths, and by grouping shared
+uncertainty.
+
+The design is not broad IAMScope correctness. It is not exploitability proof.
+It is not production readiness. It is not live AWS validation.
+
+## Synthetic Graph Overview
+
+The fixture should model a synthetic multi-principal AWS IAM graph with
+overlapping PassRole and AssumeRole patterns. It should use only the synthetic
+account `000000000000`.
+
+Suggested future fixture path:
+
+`tests/fixtures/demo/complex_shared_uncertainty_iam_benchmark/`
+
+The graph should include:
+
+- multiple source principals that can reach different services or roles;
+- multiple target roles with different trust and constraint states;
+- service principals for Lambda and ECS;
+- synthetic resources that make path-shaped rows concrete without implying
+  live AWS reachability;
+- repeated ambiguity sources so several inconclusive findings share the same
+  missing evidence class.
+
+## Principals
+
+The future fixture should include five source principals:
+
+- `AnalystUser`
+- `CICDRole`
+- `DeveloperRole`
+- `ReadOnlyAuditRole`
+- `BreakGlassCandidateRole`
+
+Each principal should have deterministic synthetic identifiers under account
+`000000000000`. No real account IDs or real IAM ARNs should be committed.
+
+## Roles and Resources
+
+The future fixture should include six target roles:
+
+- `LambdaExecutionAdminLikeRole`
+- `LambdaExecutionScopedRole`
+- `ECSExecutionAdminLikeRole`
+- `CrossAccountOpsRole`
+- `BoundaryConstrainedRole`
+- `UnknownResourceScopeRole`
+
+The future fixture should include two service principals:
+
+- `lambda.amazonaws.com`
+- `ecs-tasks.amazonaws.com`
+
+Optional synthetic resources may include:
+
+- one Lambda function resource pattern;
+- one ECS task or service pattern;
+- one Secrets Manager secret pattern;
+- one S3 bucket pattern.
+
+Resource identifiers should be synthetic and deterministic. Any role named
+`AdminLikeRole` should be treated as a synthetic label for fixture structure,
+not as downstream authorization proof or exploitability proof.
+
+## Edge Categories
+
+The fixture should cover these edge categories:
+
+- PassRole-to-Lambda candidate edges.
+- PassRole-to-ECS candidate edges.
+- AssumeRole chain edges.
+- Cross-account trust-shaped edges using only synthetic account values.
+- Permission boundary blocked path evidence.
+- SCP blocked path evidence scoped only to this synthetic fixture.
+- Identity-Deny suppressed path evidence scoped only to this synthetic fixture.
+- Missing `iam:PassRole` precondition evidence.
+- Missing target trust precondition evidence.
+- Shared unknown resource-scope evidence.
+
+The fixture should not add new reasoners or benchmark semantics. If an expected
+path cannot be represented with existing local IAMScope behavior, the next
+implementation slice should document that gap in the oracle instead of faking a
+finding.
+
+## Expected Naive Candidate Categories
+
+The future fixture should include roughly 35 to 60 deterministic naive
+path-shaped candidate rows. A naive candidate should be any source ->
+action/precondition -> target row emitted by the fixture before evaluating
+blockers, preconditions, or uncertainty.
+
+Naive candidate categories should include:
+
+- Lambda `CreateFunction` plus `iam:PassRole` shaped rows.
+- ECS task-definition/run plus `iam:PassRole` shaped rows.
+- direct AssumeRole shaped rows.
+- chained AssumeRole shaped rows.
+- cross-account trust-shaped rows using synthetic account identifiers.
+- rows blocked by permission boundary evidence.
+- rows blocked by SCP-like evidence scoped only to this fixture.
+- rows suppressed by selected identity-Deny evidence scoped only to this
+  fixture.
+- rows missing `iam:PassRole` evidence.
+- rows missing target trust evidence.
+- rows with unknown target resource scope.
+
+The naive list is not IAMScope output and should not be treated as evidence of
+reachability.
+
+## Expected IAMScope Verdict Categories
+
+The future fixture should produce roughly 15 to 25 IAMScope finding/verdict
+rows. The exact count should be frozen only in the implementation slice, after
+the fixture oracle is derived from existing local behavior.
+
+Intended lower bounds:
+
+- at least 4 validated rows;
+- at least 4 blocked rows;
+- at least 3 precondition-only rows;
+- at least 6 inconclusive rows;
+- at least 3 shared uncertainty classes.
+
+If the implementation slice chooses exact counts, those counts should be
+documented as fixture-oracle counts and pinned by local tests.
+
+Expected verdict categories:
+
+- `validated`: all modeled evidence required by the selected existing reasoner
+  is present under the synthetic fixture rules.
+- `blocked`: a selected blocker, such as permission boundary, SCP-like fixture
+  evidence, or identity-Deny fixture evidence, suppresses the path.
+- `precondition_only`: the row represents a missing prerequisite such as absent
+  `iam:PassRole` or absent target trust and should not be promoted to validated.
+- `inconclusive`: the row has path-shaped structure but depends on unresolved
+  resource-scope, trust-condition, boundary, session-policy, or SCP-like
+  context.
+
+## Shared Uncertainty Classes
+
+The fixture must include these shared uncertainty classes:
+
+- `shared_passrole_target_resource_scope_unknown`
+- `shared_cross_account_trust_condition_unknown`
+- `shared_boundary_or_session_policy_context_missing`
+
+Optional additional class:
+
+- `shared_scp_scope_unknown`
+
+The primary public lesson should be that several inconclusive paths can share
+one missing evidence source. A reviewer should resolve the shared uncertainty
+before treating the affected rows as independent validated risks.
+
+## Exact Files To Build Next
+
+The next implementation slice should add proposed fixture files under:
+
+`tests/fixtures/demo/complex_shared_uncertainty_iam_benchmark/`
+
+Proposed files:
+
+- `scenario.json`
+- `binding_metadata.json`
+- `findings.json`
+- `naive_candidates.json`
+- `expected_uncertainty_groups.json`
+- `README.md`
+
+The fixture files should be synthetic, deterministic, sanitized, and local-only.
+Generated `/tmp` outputs should not be committed.
+
+## Local Tests To Build Next
+
+The next implementation slice should add focused local tests that verify:
+
+- fixture files exist;
+- only synthetic account `000000000000` appears;
+- no raw live IAM ARNs appear;
+- no generated outputs, Terraform state/cache/provider artifacts, plan files,
+  outputs JSON, or raw live artifacts are committed;
+- naive candidate count is within the chosen design range or equals a frozen
+  oracle count;
+- finding/verdict counts match the frozen fixture oracle;
+- each naive candidate maps to one IAMScope finding or one documented
+  non-finding reason;
+- each inconclusive finding referenced by an uncertainty group exists in
+  `findings.json`;
+- shared uncertainty classes include
+  `shared_passrole_target_resource_scope_unknown`,
+  `shared_cross_account_trust_condition_unknown`, and
+  `shared_boundary_or_session_policy_context_missing`;
+- no output claims live AWS validation, exploitability proof, production
+  readiness, broad correctness, composite benchmark score, or pass/fail
+  benchmark label.
+
+## Demo Runner Integration Plan
+
+After the fixture and oracle tests exist, a later slice may wire the fixture
+into the local public demo review runner.
+
+The integration should remain local-only and should:
+
+- read the committed synthetic fixture;
+- write generated outputs under `/tmp` or a caller-provided path outside the
+  repository;
+- summarize naive candidate counts, verdict counts, and shared uncertainty
+  classes;
+- preserve the existing public claim boundary;
+- avoid composite scoring or pass/fail benchmark labels;
+- avoid live AWS, Terraform, AWS CLI, STS, Lambda API, and `iam:PassRole` calls.
+
+Runner integration should not be part of the next implementation slice unless a
+future review explicitly narrows it to a small, report-only change.
+
+## Evidence Boundaries
+
+This design is for a local synthetic benchmark/demo fixture only.
+
+The future fixture should not be described as live AWS evidence. It should not
+claim replay equivalence unless a separate local replay-equivalence test proves
+that equivalence using existing IAMScope machinery without reasoner or benchmark
+semantic changes.
+
+The fixture may claim only that, for this synthetic deterministic graph,
+IAMScope's local outputs separate path-shaped candidates into validated,
+blocked, precondition-only, and inconclusive categories under the fixture oracle.
+
+## Non-Claims
+
+This design and the future fixture do not claim:
+
+- broad IAMScope correctness;
+- broad PassRole correctness;
+- generic Deny correctness;
+- resource-policy Deny support;
+- SCP Deny support beyond selected synthetic fixture behavior;
+- exploitability proof;
+- downstream authorization proof;
+- Lambda invocation behavior;
+- production readiness;
+- correctness for real AWS environments;
+- correctness for other principals, roles, accounts, regions, conditions,
+  permission boundaries, SCPs, resource policies, or findings;
+- composite benchmark score;
+- pass/fail benchmark label.
+
+Short-form review boundaries:
+
+- not broad IAMScope correctness;
+- not exploitability proof;
+- no composite benchmark score;
+- no pass/fail benchmark label.
+
+## Exact Next Implementation Slice
+
+Recommended next slice: implement complex synthetic benchmark fixture and oracle tests.