Repository navigation
Conversation
… out of them A seal that re-stamps itself on every run proves nothing. Three separate causes made the versioned artefacts churn on every test run: - seal_splits() always wrote sealed_at = now(); it also contained a latent bug where sealed_at was briefly assigned the _sha256_str function object. - pre_register() always wrote registered_at = now(), twice per call. - register_dataset() always wrote registered_at = now(), and the registry cache was a single module-level dict, so a redirected registry leaked into the real one. Now: - sealed_at / registered_at only advance when the sealed payload actually changes, so re-sealing an unchanged record is a true no-op. - seal_splits/verify_seal/unlock_historical_test/get_split accept splits_path. - register_dataset/verify_integrity/get_registry accept registry_path; the cache is keyed by resolved path and _reset_registry_cache() exists. - Backtester accepts prereg_dir and registry_path, and writes the pre-registration once instead of twice. - Tests redirect all three mutable artefacts to a temp root via a session fixture, and a new session guard fails the suite if any sealed artefact changes during the run. pytest now leaves the working tree clean (previously 6 tracked files were rewritten on every run).
Two defects in the multilayer engine, both verified by measurement.
R1 — the /v1 multilayer endpoint returned the SAME global opinion trajectory
three times under the keys social/digital/economic, so the flagship
'multilayer' output was a three-fold duplicate of a single array. Each layer is
a distinct topology, so the quantity that genuinely differs per layer is the
mean opinion an agent is exposed to through that layer: mean(A_l . opinion).
- New MultilayerEngine.layer_opinion_profile(); because layers are
row-normalised this reduces to dot(colsum(A_l)/N, opinion), an O(N) scalar
product per step rather than a second O(N.k) matvec.
- series is now {opinion, layer_signal{social, digital, economic}} with real,
distinct values (social -0.019, digital -0.015, economic +0.252 on the same
run), replacing three identical arrays.
- llm_orchestrator reads series['opinion'].
R11 — dynamic_rewiring() refreshed only the dense _layers_flat, while step()
reads _layers_sparse. Measured: 90% censorship removed 720 edges from the dense
view and ZERO from the sparse view, so censorship and viral-hub interventions
had no effect on the simulation whatsoever.
- New _rebuild_layer_views() rebuilds dense, sparse and column-mass together;
dynamic_rewiring calls it. Censorship intensity now measurably changes the
outcome.
…ling matrix Before this change the model was not a coupled system. Measured on the real engine: only 2 of 20 possible cross-variable couplings existed, both positive, with no feedback cycle anywhere, and 4 of the 5 state variables were inert — perturbing cooperation, income or info_access moved no output by more than 0.0007, and polarization was 0.697 +/- 0.005 for every seed, population size, dt and layer weighting. The coupling is now an explicit table (DEFAULT_COUPLING_GAINS) where every term carries its mechanism (COUPLING_RATIONALE) and is read directly by multi_potential_gradient. Density 10% -> 30%, and the graph now closes a cycle op -> coop -> info -> op. Design defects fixed: - income was dead: it relaxed to a flat 0.5 with no in-edges. Hierarchy now shifts its level (not just its rate); per-agent spread 0.099 -> 0.199. - hierarchy was bistable but disconnected: it now drives cooperation and income. - the network force only touched opinion. It now reaches all five dimensions via NETWORK_SUSCEPTIBILITY, using one sparse matmul for the K columns instead of K separate matvecs. - the opinion coupling modulates the double-well POSITION instead of applying a linear push, which is the correct reading of the filter-bubble effect on a bimodal potential and finally gives a real response surface. Measured after the rebuild: - echo-chamber gain 0.0 -> 0.75 moves polarization 0.6909 -> 0.7820 - coupling 0.0 -> 1.0 moves polarization 0.7249 -> 0.8408 - layer weights move cooperation and info_access - ablating any one of the six declared terms changes the outcome Also vectorises _multilayer_langevin_step_core, replacing a triple Python loop over (layer, agent, neighbour) with a matmul. Adds tests/test_coupling_structure.py (23 behavioural tests) that fail if the model regresses to a feed-forward chain, if a declared term becomes decorative, if a negative gain stops inverting the relationship, or if the response surface disappears. Suite: 908 tests, 0 failures. BREAKING CHANGE: default dynamics change. Existing calibration artefacts (configs/calibrated, pre-registration seals) describe the previous potential and must be re-derived before being quoted as validation evidence.
The API exposed n_agents, steps and seed — none of which moves the multilayer result — while coupling, the one parameter that genuinely changes polarization, was unreachable. layer_weights was exposed but had a <0.005 effect. - MultilayerRequest DTO (extra='forbid', anti-DoS bounds): n_agents, steps, seed, dt (bounded to the explicit-Euler stability limit), coupling (0..5), layer_weights, and coupling_gains as 'source->target' overrides that make ablated variants and counterfactuals expressible over HTTP. - POST /v1/engine/multilayer returns landscape, series (opinion + real per-layer signals), diagnostics and the resolved coupling matrix. - run_multilayer_simulation accepts coupling, dt and coupling_gains; malformed or unknown override keys surface as 422 with the offending name. - 10 contract tests: 401 without a key, 422 on unknown fields, 422 on an unstable dt, per-layer series are not duplicates, coupling is an effective dial, and the matrix override actually ablates the feedback loop. Also removes the duplicate dense (L, N, N) _layers_flat stack, whose only consumer was the drift function (now sparse). Measured init peak at N=2000 drops from 293 MB to 132 MB (-55%); the removed copy would have cost 2.4 GB at N=10000. BREAKING CHANGE: /v1 engine response for the multilayer path is now served by a dedicated endpoint with a different series shape.
…e CLI Fase 2 (obsolete) and Fase 3 (coverage/quality) of the audit plan. CfC is now opt-in and lazy. The residual corrector was disqualified by walk-forward validation (R2 = -18.7, 8.2x worse than persistence), yet every MultilayerEngine instantiation imported torch, computed a tau matrix and logged a disclaimer in order to apply no correction at all. It is now behind use_cfc=True / MASSIVE_ENABLE_CFC=1, and a rejected model no longer pays for the load. Packaging contract: - tests/test_packaging_contract.py pins the exact defect class that reached main: a py-module declared in pyproject.toml that no longer exists, which setuptools only warns about. It also fails on include-globs that resolve to no package, unresolvable console-script targets, missing package-data, and any Python import of a known-removed module. - It immediately found two real defects: metrics/py.typed was declared as package-data but absent (so type checkers ignored the package), and the adapters* glob resolved to nothing. Both fixed; the dead monitoring* glob is gone (monitoring/ holds only YAML/JSON assets, no Python package). Reproducibility: DEFAULT_CONFIG had no seed, so two identical CLI runs returned different results even though the README promises 'seeded RNG everywhere'. simulate now defaults to seed 42 with a --seed override, and --pasos/--n-runs are bounded like their HTTP DTO counterparts. Coverage: - tests/test_cli_main.py: massive/cli/main.py was at 0% while the README presents the CLI as the server-less quick start. - tests/test_sentiment.py: massive/core/sentiment.py was at 0% despite backing the stance arm that decides compare_arms outcomes; now 100%. - docs/archive/README.md marks the archive as historical and non-normative, which is what made ~120 references to removed modules misleading rather than merely old.
…h the purge The advertised type gate was inert. .github/workflows/lint.yml ran 'mypy --config-file mypy.ini', but mypy.ini was deleted when [tool.mypy] moved into pyproject.toml. mypy exits 2 with 'Cannot find config file', so the Lint & Type Check job failed on every run — and it duplicated the working gate that typecheck.yml already had. - lint.yml now delegates to scripts/typecheck_slice.py, giving one authoritative definition of the type slice. - The slice grows from 25 to 32 files and now includes multilayer_engine.py, which sat outside every gate and is exactly why the duplicate dense layer stack and the inert dynamic_rewiring survived. - PRODUCTION_ARCHITECTURE_SPEC.md no longer points at the deleted mypy.ini. Security (bandit 48 -> 41 findings, 7 MEDIUM -> 0): - simulator.py: HTTP timeout coerced through _positive_timeout(); a null llm_timeout previously left the request without a timeout and could hang the simulation loop. - compare_arms.py: URL scheme validated before urlopen (rejects file://, data:, ftp:// supplied via configuration). - docs/examples: torch.load(..., weights_only=True). - massive serve: default bind 0.0.0.0 -> 127.0.0.1; exposing on all interfaces is now explicit. - False positives documented with # nosec and a reason (Host-header allow-list, indirect timeouts). Reproducibility: DEFAULT_CONFIG had no seed, so two identical CLI runs returned different results despite the README promising 'seeded RNG everywhere'. Detected by a determinism test; simulate now defaults to seed 42 with a --seed override, and --pasos/--n-runs are bounded like their HTTP DTOs. Purge: CfC no longer loads by default (a torch import, a tau matrix and a disclaimer per instantiation, to apply no correction). docs/archive/README.md marks the archive non-normative and lists the dead module names. Coverage: massive/cli/main.py 0% -> 74%, massive/core/sentiment.py 0% -> 100%, total 69.02% -> 70.03%, and fail_under raised 30 -> 65 because a 30% floor could never detect a regression. tests/test_packaging_contract.py pins the packaging defect class that reached main and immediately found two live ones: metrics/py.typed was declared as package-data but absent (so type checkers ignored the package) and the adapters* glob resolved to no package. Both fixed; the dead monitoring* glob is gone.
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
|
Important Review skippedAuto reviews are disabled on base/target branches other than the default branch. Please check the settings in the CodeRabbit UI or the
|
|
You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard. |
Four jobs were red on main for reasons unrelated to any recent change. R12 (critical) — the Docker image could not be built at all. Dockerfile:61-62 runs 'COPY nginx.conf' and 'COPY supervisord.conf', and neither file existed in the repository: failed to compute cache key: failed to calculate checksum ...: /supervisord.conf: not found Both are created, written for the constraints the Dockerfile itself imposes: - nginx.conf runs non-root (USER appuser + setcap for :80): no user directive, pid and every *_temp_path in /tmp, logs in the chowned /var/log/nginx. Reverse-proxies /v1, /api/v1, /health, /ready, /metrics, /version, /docs and /openapi*.json; immutable caching for Vite's hashed /assets/ and no-store for index.html; SPA fallback. - supervisord.conf runs uvicorn on loopback (nginx is the only public entry) and nginx in the foreground. Deliberately no user= (the container already starts as appuser and supervisord can only drop privileges from root) and no [eventlistener] (listeners must speak the stdin protocol; a plain 'supervisorctl status' would block the event loop). Verified for real: nginx -t accepts the config; supervisor's own ServerOptions.realize parses the other and reports both programs. Lint & Type Check, two independent causes: - Type Check (MyPy) ran 'mypy --config-file mypy.ini' against a deleted file and exited 2 on every run. Now delegates to scripts/typecheck_slice.py. - Python Lint installed unpinned 'ruff black mypy', so CI got a newer black than the one pyproject pins (black==24.10.0) and demanded reformatting of files that are correctly formatted per the project's own pin. The versions are now read from [project.optional-dependencies].dev so they cannot drift again. Secret scan: gitleaks scans BASE^..HEAD, but the default fetch-depth of 1 leaves the base commit's parent absent -> 'ambiguous argument', ~0 bytes scanned, exit 1. The incremental job now checks out full history. R13 (reported, not changed): requirements.txt pins torch>=2.2.0 without a CPU index, so the CPU simulator image pulls ~2.6 GB of CUDA stack (torch 539 MB, cudnn 519 MB, cublas 405 MB, triton 286 MB). Not touched here because it changes dependency resolution for every user. api.py — corrected my own earlier classification. The frontend does NOT use it: frontend/src/services/api.ts uses baseURL '/v1' and states that 'the deprecated /api/* aliases are no longer referenced by this client'. The local-development runbook claimed the opposite, which is what kept the legacy module alive. The runbook is fixed, api.py now emits DeprecationWarning naming the canonical app, 'massive serve --legacy' warns loudly, and three tests pin all of it. Docs: new docs/architecture/coupling_matrix.md (state vector, declared matrix with mechanisms, network susceptibility, measured response surface, ablation table, invariants, how to change the model); /v1/engine/multilayer documented in docs/api.md; canonical-vs-historical trust table in docs/index.md; mkdocs nav updated and the site build verified.
Every workflow filtered pull_request to [main, master], so a PR into any other long-lived branch (prod-readiness, release/*) received no pytest, no ruff, no mypy, no Docker build and no secret scan. That is precisely how the broken type gate (mypy --config-file mypy.ini), the unbuildable Dockerfile (missing nginx.conf / supervisord.conf) and the unpinned lint tools accumulated: nothing ever ran on the branch where the work happened. push stays restricted to the default branch; validate_ts_types keeps its paths filter. Only the pull_request branch filter is removed.
The blocking typecheck job installs a minimal dependency set, so 'requests' is absent there and requests.post is Any. With warn_unused_ignores = true the '# type: ignore[arg-type]' on the json= argument is then flagged as unused — it is needed only in the environment that has requests installed. Declaring the payload as dict[str, Any] satisfies requests' JsonType directly, so no suppression is required in either environment. Verified both ways: the full local slice (requests present) and a run with requests blocked to reproduce the minimal CI environment.
The routers carrying the weakest coverage are the product's public surface: benchmarks 18%, forecast 29%, engine 42%, llm 40%. 'It returns 200' is not a contract, so these tests pin what actually matters to a client: - every documented router rejects an unauthenticated caller (401) and a wrong key; - the benchmark router rejects absolute paths and directory traversal, and fails closed (503) when LLM mode is requested without a provider key rather than silently degrading to an offline run; - the forecast router requires simulation_state, validates mode and the n_runs bound, and returns a response matching the real ForecastResponse shape (sim_id / horizon_ticks / points / feasibility) instead of whatever the engine happens to emit; - the engine router rejects an out-of-range coupling, a population below 2, and a missing user_goal; - the llm router rejects an empty prompt and a content-less extract; - unknown paths are 404. Coverage of the targeted routers: forecast 29% -> 87%, benchmarks 18% -> 62%.
docker-compose publishes '8000:8000' and the Docker E2E workflow probes http://127.0.0.1:8000/health from the host. Docker port publishing forwards to the container's external interface, so binding uvicorn to the container's loopback (as the first version of this config did) would leave the published port dead and the health check would time out after 40 attempts. nginx on :80 remains the intended public entry for browser traffic (SPA + /v1 proxy); :8000 is the direct API surface that the compose file advertises. --forwarded-allow-ips stays at 127.0.0.1 because that is the peer uvicorn sees for requests proxied by nginx, so X-Forwarded-* is trusted only from it. Re-parsed with supervisor's own ServerOptions.realize: both programs load.
…PU (R1), CLI forecast defaults (R7) R2 (critical) — MultilayerEngine was O(N²) in two independent places, measured on the real engine: 1. build_layers returned dense (N, N) adjacency and MultilayerEngine kept them: resident 1536 MB at N=8000. Generators now build CSR natively from the networkx graph (nx.to_scipy_sparse_array), with no dense intermediate, behind dense_layers=False (opt-in; default behaviour unchanged). 2. generate_hierarchical connected every pair of hubs with p=0.7, a near-dense clique O(n_hubs²): 2.36 M edges and ~4.3 GB transient peak at N=16000. The hub core now has bounded degree (_HUB_HUB_DEGREE=8), preserving the hierarchy while keeping the graph O(N). Measured before/after: - N=8000 resident: dense 1536 MB -> sparse 1.2 MB (1280x) - N=16000 init peak: 8.4 GB -> 31 MB (270x) - numerical parity is exact: dense vs sparse produce bit-identical state (pol 0.721955 vs 0.721955 at N=3000) Note: the economic layer topology changes (bounded hub degree vs the old dense clique). It is deterministic per seed, but it is a different network than before; this is the intended fix, not a silent regression. R1 — Dockerfile now defaults the torch index to CPU (build arg TORCH_INDEX), so the CPU simulator no longer pulls ~2.6 GB of nvidia-* wheels; GPU builds pass --build-arg TORCH_INDEX=https://download.pytorch.org/whl/cu124. R7 — the CLI forecast command constructed TemporalConfig(event_type=...) which used the generic 7/90 defaults for EVERY event (policy_adoption should be 365 days, viral_online 14). Now uses from_event_type(), so per-event-type defaults apply. Also fixes the 2 mypy errors in massive/cli/main.py. Tests: tests/test_sparse_layers.py (7) pins dense/sparse parity, O(N) hierarchy and CSR rewiring; test_cli_main.py pins the event-type defaults. Suite: 0 failures, coverage 70.45%.
…g gate
R3 — the README advertises a 'blocking' type check, but it only covered 32
hand-picked files while /home/adlg/.local/lib/python3.14/site-packages/numpy/__init__.pyi:737: error:
Type statement is only supported in Python 3.12 and greater [syntax]
type _Falsy = L[False, 0] | bool_[L[False]]
^
Found 1 error in 1 file (errors prevented further checking) had 26
errors across 6 files. All are now resolved:
- The dominant pattern was float(x.value) over CalibratedValue
(float | list[float]): a latent TypeError (float([a,b]) raises). Added
scalar_value() in massive/core/schemas.py (midpoint for intervals) and used it
in schemas.py (3) and physics_calibration.py (8).
- convergence_certifier.py: initial_state typed np.ndarray | dict but handled a
scalar branch; widened to include float and cast the scalar in the
np.isscalar branches. Also dropped a no-redef annotation.
- abcsbi.py: _means/_stds/_weights are Optional ndarray set by fit(); both
sample() and density() now assert non-None locals after the _fitted guard. A
loop variable idx was reassigned from an ndarray to a scalar — renamed pick.
- reconstruct.py: return float(residuals.sum()) instead of a numpy int.
The blocking gate now covers the FULL clean scope: massive/ backend/ services/
massive_core/ plus the clean root modules (multilayer_engine, cfc_*, etc.) —
98 files, 0 errors, verified in both the full-dependency and minimal-CI
environments. Remaining root modules (document_intelligence 40, energy_runner
17, simulator 5, api 4, …) are tracked for a follow-up, not gated yet.
Verified: mypy full scope 0 errors; suite 0 failures.
ground_truth/_generate.py is the generator of the SEALED ground-truth data and had ZERO test coverage. Its pure, deterministic core is now covered (0% -> 50%): - Gini coefficient (uniform -> 0, extreme -> ~1, order/scale invariance) - IPF synthetic-microdata fitter (marginals honoured, normalisation, zero marginal does not diverge) - generate_microdata (shape, bipolar opinion bounds, per-seed reproducibility) - compute_network_metrics (key set, reproducibility of the scientific metrics; metadata.generated_at is intentionally non-deterministic and checked apart) Found and fixed a real bug while testing: _iterative_proportional_fitting used np.where(current > 0, target / current, 1.0), which computes the division for EVERY element (0/0 -> NaN) and only then masks it, emitting a RuntimeWarning on every sweep. Replaced with np.divide(..., where=current > 0) which never divides by zero. Also documented (not changed) that compute_network_metrics stamps metadata.generated_at = now(), so the returned dict is intentionally not byte-deterministic even though its scientific metrics are.
The Makefile api-legacy target claimed '(used by frontend/)' — false: the frontend uses baseURL /v1 and no longer references /api/*. install.sh's fallback to api:app now warns loudly instead of presenting the legacy app as equivalent. api.py itself already emits DeprecationWarning naming backend.app.main:app; the entry points that still reference it are now honest about what they serve.
Executes the remediation plan from a full audit of
origin/main. Six commits, each independently verifiable.Depends on #160 (the
mainsync), otherwise the packaging fix here conflicts with the stalesocial_connectorsdeclaration that still lives onmain.0. Pre-registration seals were not seals
Running
pytestrewrote six versioned artefacts on every run —prereg_*.yaml(registered_at),datasets/ground_truth/splits.json(sealed_at) and_provenance_registry.json. A seal that re-stamps itself cannot prove that thresholds were locked before the data was seen, which is the entire claim of the pre-registered anti-leakage protocol.Three causes, all fixed at the semantic level:
seal_splits()always wrotesealed_at = now(); it also contained a latent bug wheresealed_atwas briefly assigned the_sha256_strfunction object. It now only advances when the sealed payload changes.pre_register()wroteregistered_at = now()twice per call. It now preserves the original timestamp unless the registered content changes, and writes once.register_dataset()always re-stamped, and the registry cache was a single module-level dict, so a redirected registry leaked into the real one. The cache is now keyed by resolved path.Tests redirect all three mutable artefacts to a temp root, and a session guard fails the suite if any sealed artefact changes (verified to fire).
1. The model was not a coupled system
Measured on the real engine: 2 of 20 possible cross-couplings, both positive, no feedback cycle, and 4 of 5 state variables inert. Perturbing
cooperation,incomeorinfo_accessmoved no output by more than 0.0007. Polarization was0.697 ± 0.005for every seed, population size,dtand layer weighting — the model was a near-constant with nothing to optimise.The coupling is now an explicit, auditable table where every term carries its mechanism (
COUPLING_RATIONALE): density 10% → 30%, and it closes a cycleopinion → cooperation → info_access → opinion.Three design defects fixed:
incomewas dead (relaxed to a flat 0.5 with no in-edges; spread 0.099 → 0.198),hierarchywas bistable but disconnected, and the network force touched onlyopinion(it now reaches all five dimensions).Measured: echo-chamber gain 0.0 → 0.75 moves polarization 0.6909 → 0.7820;
coupling0.0 → 1.0 moves it 0.7249 → 0.8408; ablating any single term changes the outcome. Ablating all of them reproduces the old pathology exactly (incomespread collapses back to 0.0885).2. Two bugs found while implementing
/v1returned the same opinion trajectory three times undersocial/digital/economic. Nowseries.opinionplus a real per-layer signalmean(A_l · opinion), computed as an O(N) dot product because the layers are row-stochastic.dynamic_rewiring()refreshed only the dense copy whilestep()reads the sparse view: 90% censorship removed 720 edges from the dense view and ZERO from the one the dynamics use. Censorship and viral-hub interventions had no effect whatsoever.3. Memory
Removed the duplicate dense
(L, N, N)stack: init peak at N=2000 293 MB → 132 MB (-55%); the removed copy would have cost 2.4 GB at N=10000. The dense kernel went from a triple Python loop over(layer, agent, neighbour)to a matmul.4. The API could not influence the model
n_agents,stepsandseeddo not move the result;couplingdoes, and it was unreachable. NewPOST /v1/engine/multilayerwithcoupling,dt(bounded to the explicit-Euler limit),layer_weightsandcoupling_gainsas"source->target"overrides, so ablated variants and counterfactuals are expressible over HTTP. 10 contract tests.5. The advertised type gate was broken
.github/workflows/lint.ymlranmypy --config-file mypy.ini, but that file was deleted when[tool.mypy]moved intopyproject.toml. mypy exits 2 with "Cannot find config file", so Lint & Type Check failed on every run and the "Type-check: MyPy (blocking)" badge backed nothing. It now delegates toscripts/typecheck_slice.py, and the slice grows 25 → 32 files — includingmultilayer_engine.py, which sat outside every gate and is exactly why the defects above survived.6. Security, obsolete code, coverage
llm_timeoutcould hang the simulation loop), URL scheme validation beforeurlopen,torch.load(weights_only=True)in examples, andmassive servenow binds127.0.0.1instead of0.0.0.0by default.DEFAULT_CONFIGhad no seed, so two identical CLI runs returned different results despite the README promising "seeded RNG everywhere". Found by a determinism test; now seed 42 with a--seedoverride.tests/test_packaging_contract.pypins the defect class that reachedmainand immediately found two live ones:metrics/py.typeddeclared as package-data but absent (type checkers ignored the package), and theadapters*glob resolving to no package.sentiment.py0% → 100%, total 69.02% → 70.03%;fail_under30 → 65, because a 30% floor could never detect a regression.Verification
ruff check .black --check .mypy(blocking slice)pytestbanditpytestconfigs/calibrated/, pre-registration seals) describe the previous potential and must be re-derived before being quoted as validation evidence. This is the honest consequence of fixing the coupling, not an oversight./v1multilayer series shape changed ({opinion, layer_signal}), served by a dedicated endpoint.🤖 Generated with Claude Code
Summary by cubic
Executes the audit remediation plan: fixes seals that re-stamped themselves, replaces the near-constant engine with a genuinely coupled system, exposes the dials that actually matter over HTTP, makes the layers memory-efficient, and repairs the inert CI type gate plus a Docker image that could not be built.
Breaking changes: default engine dynamics change, so calibration artefacts (
configs/calibrated/, pre-registration seals) must be re-derived before being quoted as validation evidence; the/v1multilayer series shape changed and is now served by a dedicated endpoint; and the economic layer now has a bounded hub degree, so a given seed yields a different network than before.Engine and seals
registered_at/sealed_aton every test run; they now only advance when the sealed payload changes, and a session guard fails the suite if any sealed artefact changes.opinion → cooperation → info_access → opinioncycle.incomewas dead (spread 0.099 → 0.198); the network force now reaches all five dimensions.dynamic_rewiring()modified the dense layer copy whilestep()reads the sparse view, so censorship had no effect; both views now rebuild together./v1returned the same opinion trajectory three times; series now returnsopinionplus a real per-layermean(A_l · opinion)signal.sentiment.py→ 100%,ground_truth/_generate.py0% → 50% (the sealed-data generator), total 70.45% withfail_underraised 30 → 65; the IPF fitter no longer divides by zero.API, CI, and Docker
POST /v1/engine/multilayerexposescoupling,dt,layer_weights, andcoupling_gainsoverrides; the previously exposedn_agents,steps, andseeddo not move the result.[main, master], so PRs into other long-lived branches got no checks; they now run on any target branch.lint.ymlranmypy --config-file mypy.iniagainst a deleted file; it now delegates toscripts/typecheck_slice.py, which graduated from hand-picked files to the full clean scope (0 errors after fixing 26 type errors, notablyfloat()overCalibratedValue), and installs lint tools from the pinneddevdependencies.Dockerfilecopiednginx.confandsupervisord.conf, which did not exist. Both are now included, written for the non-root container user, uvicorn binds0.0.0.0so the published :8000 port is reachable, and the torch index defaults to CPU instead of pulling ~2.6 GB of nvidia-* wheels.DEFAULT_CONFIGhad no seed, so identical CLI runs returned different results; defaults to seed 42 now, and the CLI forecast command applies per-event-type temporal defaults.torch.load(weights_only=True), andmassive servebinds127.0.0.1by default.api.pymodule has no frontend consumers; it now emits aDeprecationWarning, theapi-legacyMakefile target and theinstall.shfallback warn loudly, and the runbook no longer claims the React app depends on it.Written for commit 2d4504c. Summary will update on new commits.