Code-review baseline: current public main as inspected on 2026-10-13; see the dated review record for the pinned revision and validation limits. This map covers 97 direct Python files in experiments/, interop/, sim/, and randomness_lab/ (excluding interop/tests/). It is source/declaration navigation, not an independent line-by-line security review or a release qualification.
Read this before interpreting results: sim/ and randomness_lab/ primarily model synthetic agents/outcomes, while parts of experiments/ and E016/E017 consume real observed verdicts or public execution-graded data. Neither synthetic mechanisms, exploratory analysis of already seen data, a passing schema fixture, nor a matched simulation budget automatically demonstrates general real-world agent value. A metadata-only verifier does not execute candidate code; explicit behavioral harnesses operate in separate controlled contexts. interop/ is an in-process protocol binding, not a network service. All paths grant no direct permission to approve, push, or merge.
experiments/harness.py + project_contracts.py -> contract/config validity
WorkUnit + bound source + worker candidate -> ResultManifest
experiments/evaluator_plan_runner.py -> evaluator-owned verification
experiments/run_evidence_report.py -> non-selecting evidence
experiments/replay_run.py -> reproducibility comparison
experiments/record_human_decision.py -> explicit local decision record (no merge)
interop/bindings.py + adapters.py -> external agent protocol normalization
sim/ + randomness_lab/ -> falsifiable research models/analyses
Use source-derived architecture, protocol specifications, research index and test practice to distinguish runnable components, tests, and evidence claims.
experiments/ (27 source files)Executable contracts, calibration and bounded two-attempt/reference experiments. Main experiments guide; see run evidence, EvaluatorPlan and research design.
| Source module | Actual responsibility | Evidence / authority boundary |
|---|---|---|
ace_population_sim.py |
Seeded ACE open-work capacity simulation | Synthetic only; not live community actuation |
agent_value_forecast.py |
Preregistered next-agent value forecasts on pinned public solve matrices | Public-data analysis; changes after the registered code freeze require amendments |
criticality_susceptibility.py |
Matched-seed synthetic queue-stress and recovery study | Engineering queue model, not a physical criticality claim |
evaluator_plan_runner.py |
Bind WorkUnit/source/validator coverage and dispatch verifier adapters | Evaluator-owned metadata checks; not merge or worker authority |
evaluator_plan_v03.py |
Additive v0.3 patch-substring evaluator compatibility layer | Binds original plan digest; candidate code stays unexecuted |
forecast_intervals.py |
Exploratory task/agent bootstrap intervals for next-agent forecast coverage | Post-hoc study of opened splits; frozen E046 forecaster unchanged |
free_compute_router.py |
Select one admitted, policy-compatible compute offer | Selection only; does not allocate resources or run tasks |
harness.py |
Phase 0 manifest/schema preflight and deterministic smoke | Smoke executes built-in fixtures only, not manifest-supplied commands |
local_compute_offer.py |
Conservative local capacity discovery to a compute-offer pool | Discovery only; no remote registration or workload execution |
local_validator.py |
Path, provenance, artifact and request validation | Metadata-only verifier; does not execute candidate code |
local_verifier.py |
Versioned JSON and unified-diff independent checks | Requires jsonschema; no candidate execution or integration |
patch_verifier_v020.py |
Add strict added-line substring semantics to legacy patch checks | Preserves original scope, provenance and integrity constraints |
project_contracts.py |
Validate ProjectManifest and DomainPack composition | Configuration only; never loads named adapters |
provenance_integrity.py |
Cross-object WorkUnit, result and verification digest checks | Detects binding inconsistencies, not semantic correctness |
public_agent_dependence.py |
E045 exploratory analysis of public execution-graded agent outcomes | Observed upstream data; not confirmatory due to earlier data exposure |
record_human_decision.py |
Emit and verify an evidence-bound local decision record | Separate explicit human action; no Git push or merge |
replay_run.py |
Capture and replay bounded two-attempt evidence bundles | Checks reproducibility; does not decide integration |
run_evidence_report.py |
Validate and render non-selecting run evidence and replay comparisons | Presentation/recommendation, never acceptance authority |
submission_provenance.py |
Reanalyze pinned submission results using model/scaffold metadata | No inferred independent-agent identity from shared submissions |
substring_patch_verifier.py |
Metadata-only added-line substring verifier | Never executes candidate code; no path/intent oracle |
temporal_holdout.py |
E048 preregistered new-submission cohort and frozen analysis replay | Uses only post-pin submissions; not a new statistical method |
transition_patch_verifier.py |
v0.4 added/removed-line transition verifier | Metadata-only semantics; behavioral test is separate |
two_attempt_orchestrator.py |
Deterministic isolated attempts via fixtures/result-bundle adapters | Not a general worker runtime; no candidate code or merge |
verification_backpressure.py |
Risk-weighted verifier scheduling and generation throttling | Research reference policy, no actual verification authority |
verification_backpressure_benchmark.py |
Temporal seeded backpressure comparison across schedules | Synthetic hidden truth, not real defect prevalence |
verification_scaling_matrix.py |
Seven-mode matched synthetic generation/verification benchmark | Simulated acceptance never changes repository state |
work_unit_decomposition.py |
Validate and aggregate five-arm WorkUnit decomposition cohorts | Committed examples are infrastructure, not observed efficacy |
interop/ (5 source files)Transport-free A2A/MCP adapter boundary. Main interop guide and mapping spec. Optional official SDK tests are conditional; skipped tests confer no conformance.
| Source module | Actual responsibility | Evidence / authority boundary |
|---|---|---|
__init__.py |
Export semantic protocol binding package | No transport or authority |
adapters.py |
Common worker-adapter interface, local/mock A2A normalization | Worker completion distinct from independent verification |
bindings.py |
Round-trip WorkUnits through A2A and MCP wire envelopes | Pure in-process mapping, not a network server |
identity.py |
Optional namespaced identity evidence on ResultManifest | Identity never authorizes verification or merge |
sdk_conformance.py |
Optional round trips through official pinned protocol SDK types | Unavailable SDKs skip; a skip is not conformance |
sim/ (42 source files)Verifier-error models, coordination, search, and congestion experiments. Main simulation guide and published experiment records. E016/E017 use fixed executable/oracle evidence in addition to synthetic studies.
| Source module | Actual responsibility | Evidence / authority boundary |
|---|---|---|
aco_stigmergy_sim.py |
Seeded ACO task-routing ecology compared with simple baselines | Synthetic workers/tasks; no live task assignment |
adaptive_verification_ecology_ablation.py |
Cumulative AVE component ablations | Budget-matched model, not live reviewer evidence |
adaptive_verification_ecology_sim.py |
Adaptive Verification Ecology task/worker/verifier toy simulator | Deterministic synthetic mechanism test |
adaptive_verification_ecology_stress.py |
AVE-3 parameter and environment stress cases | Synthetic stress grid, no deployment claim |
adaptive_verification_ecology_targeted.py |
Targeted verifier-diversity/risk/backpressure policy comparison | Seeded study; probes remain diagnostic |
e015_analyze.py |
Effective-independent-panel-size analysis of E015 grids | Model-derived quantities depend on error assumptions |
e015_quorum_frontier.py |
Cost-weighted quorum optimization over E015 grids | Assumption-bound optimum, not universal acceptance rule |
e015_worker.py |
Verifier-size, accuracy, correlation and quorum sweep | Synthetic error-dependence grid |
e016_agent.py |
Run one local open-weight LLM verifier against a fixed corpus | Node-local model call; output is an untrusted verdict |
e016_analyze.py |
Analyze observed verifier decisions and shared-error models | Corpus-specific observational analysis, not broad competence |
e016_corpus.py |
Build candidate corpus with executable hidden-test ground truth | Executes fixed corpus code; not an untrusted sandbox service |
e017_analyze.py |
Fit/compare dependence shapes against measured oracle panels | Panel findings tied to frozen corpus and instrument |
e017_oracles.py |
Generate region-scoped executable partial test oracles | Region labels do not prove independence |
e017_verify.py |
Evaluate partial-oracle panel against executable corpus | Executes controlled candidates; not arbitrary worker integration |
e018_dependence_models.py |
Recompute quorum/error models under shared-shock versus item difficulty | Closed-form model comparison |
e020_quorum_frontier.py |
Empirical E017 quorum frontier with explicit base rate/costs | Corpus- and cost-dependent optimality only |
e025_learned_verifiers.py |
Learn synthetic reliability and dependence, evaluate held-out streams | Synthetic calibration, not production reputation |
e026_archive_contamination.py |
Count false accepts retained in Quality-Diversity archives | Replays synthetic archive with an imperfect panel |
e027_defect_propagation.py |
Test cost of downstream accepted-defect propagation | Synthetic catastrophe/quality proxies |
e028_latent_defect_dimension.py |
Hide integrity truth from observable candidate quality | Synthetic confound-removal experiment |
e030_supplied_goal_membership.py |
Test whether supplied plausible goals contain future target | Goal-set intervention on synthetic search |
e031_learned_goal_filter.py |
Replace supplied goal hypotheses with learned beliefs | Synthetic learned-goal counterfactual |
e032_population_scaling.py |
Compare agent count at fixed and growing evaluation budgets | Separates compute budget from population size |
e033_goal_distance.py |
Sweep distance from plausible-goal set | Synthetic extrapolation over goal geometry |
e034_goal_direction.py |
Match distance while varying direction of goal change | Controlled directional sensitivity study |
e035_direction_across_shells.py |
Replicate goal-direction effects across multiple distances | Synthetic robustness extension |
e036_adversarial_contributors.py |
Model contributors optimizing against verification gates | Synthetic adversary, not a live red team |
e037_ladder_under_panels.py |
Repeat goal-distance/direction effects with imperfect panels | Panel settings are modeled, not field measurements |
e038_symmetric_gate_asymmetric_burden.py |
Study unequal consequences of symmetric gate errors | Synthetic selection-burden mechanism |
e039_content_addressed_blind_spot.py |
Inspect whether verifier blind spots depend on artifacts | Identifies structural limits of the modeled threat |
emergence_sim.py |
Minimal random/scalar/Quality-Diversity goal-change search | Falsifiable synthetic landscape only |
homeostatic_stigmergy_sim.py |
Congestion-governed stigmergic routing controller | Synthetic capacity regulation |
matched_budget_emergence.py |
Five-arm emergence comparison at matched evaluation budgets | Controlled model, not real software productivity |
physarum_compute_routing_sim.py |
Physarum-inspired path choice after hard resource admission | Graph assumes all offers already admitted |
physarum_compute_routing_stress.py |
Compute-path routing under shocks, outages and burden | Synthetic topology; no unreviewed live resource use |
replay_task_routing.py |
Historical/shadow task-routing replay without look-ahead | Read-only; never assigns GitHub work |
review_capacity_governor.py |
E044 bounded logistic review-capacity feedback model | Synthetic queue, no ACE policy mutation |
run_aco_parameter_sweep.py |
ACO parameter grid and Pareto-nondominated policy report | Research tuning on synthetic data |
run_aco_sweep.py |
Repeated-seed ACO comparisons across routing baselines | Simulator comparison, not measured contributors |
run_emergence_sweep.py |
Multiseed vague-goal landscape outcomes | Aggregate synthetic seeds and assumptions |
run_verifier_correlation_sweep.py |
Sensitivity of emergence outcomes to verifier correlation | Correlation modeled, not measured |
verification_aggregation_sim.py |
Independence-group-aware versus naive vote aggregation | Synthetic correlated-panel counterfactual |
randomness_lab/ (23 source files)Dependency-free stochastic outcome and policy models, R1-R4. Main lab guide, R1 readiness, and research program.
| Source module | Actual responsibility | Evidence / authority boundary |
|---|---|---|
__init__.py |
Public exports for lightweight coordination simulation | Pure research API |
__main__.py |
CLI for single or repeated simulated worker selection | Does not dispatch actual workers |
experiments.py |
Repeated seeded simulation trials and descriptive intervals | Normal-approximation summaries are not external validity |
model.py |
Worker, Bernoulli and item-difficulty outcome environments | Synthetic success distributions |
policies.py |
Greedy, exploration, UCB and Thompson-style selection policy objects | No real provider execution |
r1.py |
R1 worker-diversity and verifier-panel experiment core | Assumptions must be stated before empirical inference |
r1_correlation_threshold.py |
Dependence-threshold sweep for diversity benefit | Parameterized synthetic correlation |
r1_dependence_shape.py |
Reassess R1 advantage under an alternative joint-failure shape | Model-shape sensitivity, not measured new panel |
r1_low_diversity_threshold.py |
Descriptive low-diversity marginal-gain thresholds | Synthetic sample, not real scaling law |
r1_panel_frontier.py |
Full quorum/dependence/billing tradeoff sweep | Cost assumptions change answers |
r1_readiness.py |
Fail-closed preflight for real replay benchmark cohort | Readiness checks do not create observations |
r1_replay.py |
Analyze observed candidate/result bundles as a real R1 cohort | Requires independent evidence; no outcome fabrication |
r1_scaling.py |
Matched-attempt R1 population and diversity scaling | Synthetic benchmark family |
r1_sweep.py |
Worker/verifier correlation and quality-penalty parameter sweep | Synthetic sensitivity evidence |
r1_threshold_robustness.py |
Bootstrap/sign-test robustness of synthetic thresholds | Diagnostic intervals, not population guarantees |
r1_verifier_dependence.py |
Inspect what verifier-correlation parameter actually couples | Corrects panel-versus-single-verifier interpretation |
r2.py |
R2 load/capability-aware task-routing model and traces | Synthetic queues and outages |
r2_capability_rarity.py |
Rare-capability assignment stress study | Modeled rarity, not live HPC capacity |
r2_factor_sweep.py |
Staleness, outage and other R2 factor studies with instrumentation | Controlled model/runtime costs |
r2_scale.py |
Multi-scale R2 routing regime sweep | Synthetic worker counts, not production scale |
r3.py |
R3 evolutionary search over orchestration genomes | Simulated task families, not a production optimizer |
r4.py |
R4 stigmergy and nonstationarity policy comparison | Synthetic shifts and worker outcomes |
simulator.py |
Seeded worker/outcome execution kernel | Outputs simulated booleans, not code test verdicts |
# From a checked-out repository with Python 3.11+:
python -m pip install -e '.[verify]'
python scripts/demo.py # synthetic contract fixtures
python experiments/two_attempt_orchestrator.py --help
python experiments/replay_run.py --help
python -m randomness_lab --help
python sim/aco_stigmergy_sim.py --help
python -m pytest -q tests/test_nonpackage_module_maps.py
Some simulators, behavioral harnesses and real-data studies require additional local artifacts, pinned upstream datasets, controlled runtimes or optional SDKs. Consult their owning guides before invoking them; this map is not blanket approval to download data, execute candidate code, or dispatch agents. The CLI quick reference lists installable-package commands separately.
The matching drift test compares file paths in this table with current direct *.py sources. New/moved/removed modules should update this page and the owning guide within the same PR; changing historical scientific results still needs the appropriate preregistration/amendment process. The operational utility map covers tools/ and scripts/ separately.