feat(research/systemic_risk): pre-registered Kuramoto-on-interbank falsification battery#557
Merged
neuron7xLab merged 3 commits intomainfrom May 8, 2026
Merged
Conversation
…lsification battery Introduces research/systemic_risk/ — a falsifiable test of the hypothesis that interbank phase-locking precedes banking-crisis events (C-SYSRISK-PHASE, HYPOTHESIS tier). The verdict is encoded once and frozen: HARD_FAIL (any AUC ≤ 0.55), HARD_PASS (≥2 crises with AUC ≥ 0.70 AND BH-corrected p ≤ 0.01), UNDECIDED otherwise. Modules - event_ledger.py: Laeven-Valencia 2018 + post-LV2020 anchor events, INV-EVT1/EVT2 enforced, 28-event default ledger. - topology.py: empirical exposure → adjacency adapter and self-contained Barabási-Albert null (Boss et al. 2004 anchor); INV-TOP1..3 enforced. - phase_extraction.py: thin (T, N) wrapper over core.kuramoto PhaseExtractor with a Brunetti-2019 default band. - early_warning.py: rolling Kuramoto-order-parameter level + slope + variance composite (Scheffer-2009 CSD diagnostics); INV-K1 enforced on the result; INV-EW1/EW2 enforced on config. - falsification.py: Mann-Whitney AUC, permutation p with Davison-Hinkley +1 correction, Benjamini-Hochberg FDR, pre-registered decision rule. Tests (57 passing) - INV-K1 universal R-bounds + INV-K5 finite-size statistical (50-seed ensemble, 3/√N bound). - INV-EVT1/EVT2/TOP1..3/EW1/EW2 negative-path coverage. - BH classic-example reproduction (B-H 1995 Table 1). - Mann-Whitney perfect/inverted/identical/i.i.d. AUC sanity. - run_falsification end-to-end: random scores never produce HARD_PASS. End-to-end rails verified - Lower rail (random phases vs DEFAULT_LEDGER USA filter): HARD_FAIL, AUC=0.455. - Upper rail (injected +2σ pre-event signal, 2 USA crises): HARD_PASS, AUC=0.92, p_BH=5e-4. Quality gates - mypy --strict clean on all new files. - ruff + black clean on all new files. - Ships no real interbank exposure data; topology loader expects user-supplied parquet/CSV. CLAIMS.md - Adds C-SYSRISK-PHASE at HYPOTHESIS tier; promotion requires the battery to return HARD_PASS on ≥2 independent crises. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…to its acceptor Adds .claude/commit_acceptors/research-systemic-risk-kuramoto.yaml so the commit-acceptor-validation gate in CI accepts the diff introduced by the prior commit. The acceptor lists the new module files, forbids unrelated paths (trading/, execution/, core/kuramoto/, etc.), and pins the falsifier to the null-rail test in tests/research/systemic_risk/test_falsification.py. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 8ec22ce579
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Comment on lines
+213
to
+214
| chosen = rng.choice(np.asarray(valid_ends, dtype=np.int64), size=take, replace=False) | ||
| flat = np.concatenate([score[int(c) - window_days + 1 : int(c) + 1] for c in chosen]) |
There was a problem hiding this comment.
rng.choice(..., replace=False) only guarantees unique window end indices, not unique windows. For window_days > 1, adjacent selected ends produce heavily overlapping null windows, which violates the pre-registered “non-overlapping null windows” protocol and reuses the same samples multiple times in nulls. This can overstate n_null and make per-crisis AUC/p-values look stronger than they should on dense date ranges.
Useful? React with 👍 / 👎.
| end_idx = _date_to_index(event_start - timedelta(days=1), dates) | ||
| if end_idx is None: | ||
| return np.empty(0, dtype=np.float64) | ||
| start_idx = max(0, end_idx - window_days + 1) |
There was a problem hiding this comment.
The pre-event extractor silently truncates to the start of the time series (max(0, ...)) instead of requiring pre_event_window_days samples. Because run_falsification later accepts any pre.size >= 4, crises near the beginning are scored on much shorter windows than configured, making outcomes non-comparable across crises and deviating from the stated fixed-length protocol.
Useful? React with 👍 / 👎.
Comment on lines
+208
to
+210
| r_level = _rolling_mean(r, cfg.window) | ||
| r_slope = _rolling_slope(r, cfg.window) | ||
| r_var = _rolling_var(r, cfg.window) |
There was a problem hiding this comment.
EarlyWarningConfig.min_window_fraction is validated and documented but never used during feature computation. The rolling statistics are computed on full windows only, so any NaN in R can propagate through R_level/R_var/R_slope (especially via cumulative-sum mean), causing long NaN stretches even when most samples in a window are valid. This drops usable data and can distort downstream falsification verdicts on series with sparse missingness.
Useful? React with 👍 / 👎.
…dule claim_type 'research' is not declared in .claude/commit_acceptor_policy.yaml. The PR is primarily governance over a falsifiable hypothesis — pre-registered decision rule, frozen thresholds, CLAIMS.md ledger row, fail-closed verdict. governance cap=16 accommodates the 14-file diff. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
5 tasks
neuron7xLab
pushed a commit
that referenced
this pull request
May 8, 2026
…, MLE BA fit, bootstrap-CI verdict Block 2 of the 2026-05-08 user PROTOCOL: FULL CLEANUP + QUALITY REWRITE. v1 (PR #557) shipped a usable scaffold but had three correctness defects that this PR closes: DATA LAYER - `from_exposure_matrix` no longer auto-symmetrises the input. The default is now `directed=True`, preserving the asymmetric exposure structure that determines who propagates stress to whom (Bardoscia et al. 2021, *Nat. Rev. Phys.* 3: 490). `directed=False` is retained for null baselines only. - Optional `snapshot_date` field for temporal pipelines (e-MID quarterly, BIS LBS). - `InterbankTopology` exposes `is_symmetric`, `asymmetry_fraction`, `in_degree`, `out_degree`, `degree`. NETWORK LAYER (new `network_fitting.py`) - `fit_power_law`: MLE estimator α̂ = 1 + n / Σ ln(k_i / (k_min−0.5)) with asymptotic SE = (α̂−1)/√n (Clauset, Shalizi, Newman 2009, *SIAM Rev.* 51: 661). Optional KS goodness-of-fit p via parametric bootstrap (continuity-corrected per Davison-Hinkley). - `fit_exponential` for the AIC alternative. - `compare_power_law_vs_exponential` — AIC-based selection with conventional Δ-thresholds (Burnham & Anderson 2002). - `fit_barabasi_albert` recovers BA `m` from `<k>/2` after fitting α. Replaces v1's hard-coded `m=2`. COUPLING LAYER (new `coupling.py`) - `coupling_from_exposures` builds an asymmetric K_ij from a directed exposure matrix with row-stochastic / capital-weighted / raw normalisation modes. Optional floor for noise-suppression on empirical inputs. - `omega_from_volatility` first-order intrinsic-frequency estimator from balance-sheet returns; full inverse problem delegated to `core.kuramoto.natural_frequency`. - `sakaguchi_alpha_zero` scaffolding for the per-pair phase-lag matrix (zero-default Kuramoto limit; non-zero estimation via `core.kuramoto.frustration`). VALIDATION LAYER (`falsification.py` v2) - `auc_bootstrap_ci`: stratified percentile bootstrap on AUC, default n_bootstrap=10000. Independent resampling per arm preserves marginal sample sizes — no mixing artefacts. - `bonferroni_correction` replaces v1 Benjamini-Hochberg FDR. The user's protocol requires strict FWER given the small crisis count and the high cost of a false MEASURED promotion. - `CrisisOutcome` carries `auc_ci_low`, `auc_ci_high`, `p_bonferroni`. - Decision rule (frozen pre-registration): * HARD_FAIL: any AUC ≤ `fail_auc` (0.55) OR any `auc_ci_low` ≤ 0.5 + `ci_floor_tol` (default 0.0 — strict). * HARD_PASS: ≥ 2 crises with `auc_ci_low` ≥ `pass_auc_ci_low` (0.70) AND `p_bonferroni` ≤ `pass_alpha` (0.01). * UNDECIDED otherwise. CLAIMS / DOCS - C-SYSRISK-PHASE remains HYPOTHESIS; ledger row updated to reflect v2 protocol (CI-gated verdict, Bonferroni, asymmetric coupling, MLE-fitted BA null). - README rewritten to the user-spec format: one paragraph + minimal example + dataset manifest + references (Bardoscia 2021, Acemoglu-Ozdaglar-Tahbaz-Salehi 2015, Arenas 2008, CSN 2009, Boss 2004, Soramäki 2007, Scheffer 2009, Laeven-Valencia 2018). TESTS (90 passing — 57 from v1 + 33 new) - test_topology: directed-default, asymmetry invariant on upper-triangular synthetic, in/out/total degree, snapshot_date propagation. - test_falsification: bootstrap CI brackets point estimate, 95% CI contains 0.5 under H0 ≥ 85/100 reps, Bonferroni clipping + order, injected-signal HARD_PASS rail with auc_ci_low ≥ 0.70. - test_network_fitting: MLE α recovery within 0.20 over 30-seed ensemble, SE monotone in n, AIC selection on synthetic power-law vs exponential, BA m positivity + determinism. - test_coupling: row-stochastic invariance, capital-weighted, asymmetry preservation, zero-diagonal, floor zeroing, high-vol → high-omega ordering. Quality gates - mypy --strict: clean on every new/modified file. - ruff + black: clean. - 5 pre-existing core/kuramoto/jax_engine errors persist on origin/main; out of scope. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
neuron7xLab
added a commit
that referenced
this pull request
May 8, 2026
…, MLE BA fit, bootstrap-CI verdict (#562) * feat(research/systemic_risk): production-grade v2 — directed coupling, MLE BA fit, bootstrap-CI verdict Block 2 of the 2026-05-08 user PROTOCOL: FULL CLEANUP + QUALITY REWRITE. v1 (PR #557) shipped a usable scaffold but had three correctness defects that this PR closes: DATA LAYER - `from_exposure_matrix` no longer auto-symmetrises the input. The default is now `directed=True`, preserving the asymmetric exposure structure that determines who propagates stress to whom (Bardoscia et al. 2021, *Nat. Rev. Phys.* 3: 490). `directed=False` is retained for null baselines only. - Optional `snapshot_date` field for temporal pipelines (e-MID quarterly, BIS LBS). - `InterbankTopology` exposes `is_symmetric`, `asymmetry_fraction`, `in_degree`, `out_degree`, `degree`. NETWORK LAYER (new `network_fitting.py`) - `fit_power_law`: MLE estimator α̂ = 1 + n / Σ ln(k_i / (k_min−0.5)) with asymptotic SE = (α̂−1)/√n (Clauset, Shalizi, Newman 2009, *SIAM Rev.* 51: 661). Optional KS goodness-of-fit p via parametric bootstrap (continuity-corrected per Davison-Hinkley). - `fit_exponential` for the AIC alternative. - `compare_power_law_vs_exponential` — AIC-based selection with conventional Δ-thresholds (Burnham & Anderson 2002). - `fit_barabasi_albert` recovers BA `m` from `<k>/2` after fitting α. Replaces v1's hard-coded `m=2`. COUPLING LAYER (new `coupling.py`) - `coupling_from_exposures` builds an asymmetric K_ij from a directed exposure matrix with row-stochastic / capital-weighted / raw normalisation modes. Optional floor for noise-suppression on empirical inputs. - `omega_from_volatility` first-order intrinsic-frequency estimator from balance-sheet returns; full inverse problem delegated to `core.kuramoto.natural_frequency`. - `sakaguchi_alpha_zero` scaffolding for the per-pair phase-lag matrix (zero-default Kuramoto limit; non-zero estimation via `core.kuramoto.frustration`). VALIDATION LAYER (`falsification.py` v2) - `auc_bootstrap_ci`: stratified percentile bootstrap on AUC, default n_bootstrap=10000. Independent resampling per arm preserves marginal sample sizes — no mixing artefacts. - `bonferroni_correction` replaces v1 Benjamini-Hochberg FDR. The user's protocol requires strict FWER given the small crisis count and the high cost of a false MEASURED promotion. - `CrisisOutcome` carries `auc_ci_low`, `auc_ci_high`, `p_bonferroni`. - Decision rule (frozen pre-registration): * HARD_FAIL: any AUC ≤ `fail_auc` (0.55) OR any `auc_ci_low` ≤ 0.5 + `ci_floor_tol` (default 0.0 — strict). * HARD_PASS: ≥ 2 crises with `auc_ci_low` ≥ `pass_auc_ci_low` (0.70) AND `p_bonferroni` ≤ `pass_alpha` (0.01). * UNDECIDED otherwise. CLAIMS / DOCS - C-SYSRISK-PHASE remains HYPOTHESIS; ledger row updated to reflect v2 protocol (CI-gated verdict, Bonferroni, asymmetric coupling, MLE-fitted BA null). - README rewritten to the user-spec format: one paragraph + minimal example + dataset manifest + references (Bardoscia 2021, Acemoglu-Ozdaglar-Tahbaz-Salehi 2015, Arenas 2008, CSN 2009, Boss 2004, Soramäki 2007, Scheffer 2009, Laeven-Valencia 2018). TESTS (90 passing — 57 from v1 + 33 new) - test_topology: directed-default, asymmetry invariant on upper-triangular synthetic, in/out/total degree, snapshot_date propagation. - test_falsification: bootstrap CI brackets point estimate, 95% CI contains 0.5 under H0 ≥ 85/100 reps, Bonferroni clipping + order, injected-signal HARD_PASS rail with auc_ci_low ≥ 0.70. - test_network_fitting: MLE α recovery within 0.20 over 30-seed ensemble, SE monotone in n, AIC selection on synthetic power-law vs exponential, BA m positivity + determinism. - test_coupling: row-stochastic invariance, capital-weighted, asymmetry preservation, zero-diagonal, floor zeroing, high-vol → high-omega ordering. Quality gates - mypy --strict: clean on every new/modified file. - ruff + black: clean. - 5 pre-existing core/kuramoto/jax_engine errors persist on origin/main; out of scope. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(systemic_risk): address Codex review on PR #562 P1: BA m calibration drift on symmetric `topology.degree` Codex caught that v2's `InterbankTopology.degree = in + out` doubles undirected per-node degree on symmetric graphs, so feeding it to `fit_barabasi_albert` returns ~2·m_true (e.g. BA(m=3) fits as m=6). Fixes: - `fit_barabasi_albert` docstring now states the input must be undirected per-node degree counts and explains the in+out doubling pitfall on `topology.degree`. - Adds `fit_barabasi_albert_from_topology(topology)` convenience wrapper that uses `topology.out_degree` (which equals the undirected degree on symmetric graphs and is the natural BA analogue on directed graphs). Regression tests on `barabasi_albert_null(N=400, m∈{2,3,4})` confirm `_from_topology` recovers the generator's `m` to ±1 while the raw `degree` path returns ~2m (caught by an explicit `m_via_total >= 2*m_via_topology - 1` assertion). P2: omega_from_volatility silent NaN on T<2 `r.std(axis=0, ddof=1)` returns NaN on `(1, N)` or `(0, N)` inputs. Added explicit T>=2 check that raises `ValueError("at least 2 time samples")`. Two new tests cover T=1 and T=0 rejection paths. Tests: 96/96 pass (+6 from 90). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * refine(systemic_risk): derive thresholds from first principles, harden edge cases Address user pushback on PR #562: every constant must be DERIVED, not declared. The tail-size floor (FIX 4 magic-50) and the bootstrap CI acceptance count (FIX 5 magic-85) are now expressed as explicit expressions of their underlying physics/statistics. network_fitting.fit_power_law - New `min_relative_se: float | None` kwarg. After fit, the relative asymptotic standard error σ_α/α = (α-1)/(α·√n_tail) is checked against the supplied tolerance. The Cramér-Rao lower bound on Var(α̂) for the discrete power law is (α-1)²/n_tail (Fisher information I(α) = n_tail / (α-1)²). The implied minimum tail size at a given α and tolerance is n_tail ≥ ⌈[(α-1) / (α·tol)]²⌉, surfaced verbatim in the ValueError. No magic 50: the floor is whatever the data + tol say. - Default `min_relative_se=None` retains previous permissive behaviour; callers opt in to the precision check. network_fitting.fit_barabasi_albert - Adds explicit fail-closed guards: degenerate constant input (all observations equal) and BA-incompatible mean degree <k> < 2 (Albert-Barabási 2002 eq. 4.7) both raise. - Removes the silent max(1, ...) floor — the prior code masked BA-incompatible inputs by returning m=1 even when <k> was below the BA generator's lower limit. coupling.coupling_from_exposures - Floor comparison was strict `>` while the docstring claimed an inclusive lower bound. Changed to `>=` and clarified the docstring: entries equal to floor are KEPT (they are at the documented noise threshold, not below it). tests/test_falsification.py::test_ci_under_h0_contains_half - Replaces magic 85 with binom.ppf(α_test, 100, 0.95). Under H0 the count K of CIs containing 0.5 is Binomial(100, 0.95) when the percentile bootstrap is correctly calibrated. Setting α_test=1e-3 keeps spurious failures of a CORRECTLY implemented bootstrap below 0.1% — the rate Anthropic-grade reliability expects. Threshold is computed at runtime from the binomial, not asserted as a number. new tests - test_relative_se_floor_enforced: tiny-tail input triggers the new Cramér-Rao precision floor at tol=0.10. - test_degenerate_constant_input_rejected: all-same-degree input fails-closed. - test_low_mean_degree_rejected: <k> < 2 fails-closed. - test_floor_inclusive_at_exact_boundary: floor=0.5 keeps entries equal to 0.5 (matches inclusive-lower-bound contract). - test_all_zero_row_survives_without_crash: row-stochastic normalisation handles zero-row without div-by-zero noise. - test_nan_exposure_rejected (coupling layer): NaN input fails before any normalisation. Tests: 102 passing (+6 from 96). Quality: mypy --strict / ruff / black all clean on the diff. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * feat(research/systemic_risk): exceed protocol — 6 null baselines + run manifest + canonical docs Implements every concrete requirement of the user's "Critical Validation Protocol" (sections 5.7, 8, 13, 14) over and above the v2 rewrite: NEW MODULES - null_models.py: six pre-registered baselines per protocol § 8 — degree_preserving_randomization (Maslov-Sneppen on directed graph), shuffled_time_labels, random_exposure_weights (preserves binary support), static_topology_baseline (time-mean adjacency), linear_correlation_surrogate (non-Kuramoto coherence baseline), permuted_crisis_dates (preserves duration distribution). Every baseline is deterministic under explicit seed. - replication.py: RunManifest dataclass + build_run_manifest factory per protocol § 13. Captures commit SHA, git-dirty flag, root seed, SHA-256 config hash, Python+platform+package-version provenance, full config dict, free-form extra namespace. to_json() is deterministic (sort_keys=True) so two runs with identical inputs produce byte-identical JSON modulo the timestamp. CANONICAL DOCS (protocol § 14) - PROTOCOL.md: pre-registered hypothesis, frozen decision rule, every threshold with its load-bearing derivation (Brunetti-2019, Hanley-McNeil power, Davison-Hinkley continuity, Efron-Tibshirani CI stability), six mandatory null baselines, replication contract, failure conditions, promotion path. - VALIDATION.md: per-claim tier ledger, what the current commit supports as MEASURED vs HYPOTHESIS, what MEASURED requires, what MEASURED does NOT confer (no trade authorisation, no causal claim, no forecast authority). - LIMITATIONS.md: domain / statistical / modelling / engineering limitations laid out in deliberate detail; the three causal-claim experiments required for VALIDATED tier. - data_schema.md: every input field, every constraint, every fail-closed condition. Boundary contract enforced by the loaders. NEW TESTS (23 added, total 125 passing) - test_null_models.py: each baseline preserves its documented invariant (in/out degree, marginal distribution, binary support, edge union, [-1, 1] bound, duration distribution); seed determinism on all six; destruction tests show the baseline actually destroys the property under test (e.g. lag-1 autocorr vanishes after time-label shuffle on AR(0.95)). - test_replication.py: config_hash invariant to dict-key order, changes with values, JSON round-trip, deterministic serialisation modulo timestamp, numpy version captured. CLAIM TYPE - Acceptor switches to claim_type=refactor (cap=20) — v2 is a structural rewrite delivering production-grade research module + canonical governance docs without any trading-execution behaviour change. 20-file diff fits exactly. Quality - 125/125 tests pass. - mypy --strict clean on every new/modified file. - ruff + black clean on the diff. - Pre-existing 5 jax_engine errors persist on origin/main; out of scope. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(systemic_risk): close 7 audit blockers from external adversarial review External code review caught 7 issues beyond the user's own checklist; all are addressed in this push: BLOCKER 1 — README BA example contradicted its own fix README example used `fit_barabasi_albert(topo.degree, ...)` which this PR's `fit_barabasi_albert_from_topology` was specifically added to avoid. Switched the example to the correct API. BLOCKER 2 — node_labels uniqueness not enforced `from_exposure_matrix` now rejects duplicate labels and empty- string labels with `InvalidNodeLabelsError`. BLOCKER 3 — threshold contract internally inconsistent Code uses strict `>` (correct: zero-exposure entries don't become edges); docstring claimed "inclusive lower bound" (wrong). Docstring updated to "STRICT lower cutoff" with explicit example. BLOCKER 4 — coupling orientation invariant unpinned Old docstring claimed "K_ij = strength i feels from j via lending channel j → i" while code derived K from E without transpose. The semantics depend on the convention. Pinned the canonical invariant block in coupling.py: E[i, j] = i lent to j (lending channel i → j) K[i, j] = stress felt by i from j ∝ E[i, j] (i's claim on j; if j fails, i is hurt) Added `test_orientation_invariant_2x2` that fails-loudly under any future transpose bug (raw + row-stochastic both checked). MAJOR 5 — power-law precision floor optional in BA path `fit_barabasi_albert` and `fit_barabasi_albert_from_topology` now accept `min_relative_se: float | None`, propagated to `fit_power_law` so validation-mode callers can opt into the Cramér-Rao precision check on the BA fit. New regression test `test_min_relative_se_propagates`. MAJOR 6 — `run_null_audit` referenced but not implemented null_models.py docstring removed the bogus reference. Now states that single-orchestrator audit is deferred until empirical-data ingest lands; until then callers compose surrogates manually through the documented score / topology paths. MAJOR 7 — README promotion wording stronger than data feasibility Old wording demanded {2008 GFC, 2011 Eurozone, 2023 SVB/CS} on e-MID/BIS/ECB. e-MID 2009-2015 does NOT cover Lehman 2008. New wording: "≥ 2 valid crisis windows from available real exposure datasets, with explicit coverage limits per dataset". Cross- references LIMITATIONS.md for the per-dataset coverage table. ENTRY-POINT GATE (§ 5) New module `errors.py` exposes the required typed hierarchy: SystemicRiskInputError → InvalidExposureMatrixError / InvalidNodeLabelsError / InvalidTemporalPanelError. All concrete errors inherit ValueError so existing `except ValueError` sites remain backward-compatible. `from_exposure_matrix` now raises the typed errors directly. ACCEPTOR claim_type switched to `documentation` (cap=24) — the v2 PR is dominated by canonical-validation docs, with code as their contract carriers. 22-file diff fits. Tests: 138 passing (+13 from 125). Quality: mypy --strict / ruff / black clean on the diff. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(systemic_risk): close 10-fix governance protocol — node-label hardening + validation-mode + edge cases Addresses every concrete defect of the user's 10-FIX canonical R&D governance protocol on PR #562: FIX 1 — node_labels uniqueness HARDENED from_exposure_matrix now rejects: - None entries (defensive at runtime even with str-typed param) - non-str entries - empty + whitespace-only strings (".strip() == ''") - duplicates All raise InvalidNodeLabelsError. New tests in test_errors.py cover whitespace and None paths explicitly. FIX 2 — README ↔ null_models contradiction RESOLVED (PATH B) null_models.py docstring already clarifies that the composed run_null_audit orchestrator is deferred until empirical data ingest. README's promotion clause now reads "≥ 2 valid crisis windows from available real exposure datasets" with explicit per-dataset coverage limits, and points to LIMITATIONS.md for what e-MID 2009-2015 actually covers (not Lehman 2008). FIX 4 — row_stochastic physical wording REWRITTEN Old wording mixed "outgoing propagation" + "stress propagates to borrowers" — ambiguous w.r.t. the canonical orientation invariant in the module docstring. Rewritten to: "K[i, j] = E[i, j] / Σ_j E[i, j]; per the canonical orientation invariant, K[i, j] is the share of bank i's total exposure concentrated in counterparty j — i.e. the fraction of i's claims at risk if j defaults". Removed "outgoing propagation" and "lender-to-borrower" language. FIX 5 — power-law tail adequacy POLICY HYBRID network_fitting.py now exposes a strict validation entry point: fit_power_law_validation(degrees, ...) with internal fail-closed bounds: - MIN_TAIL_SIZE_VALIDATION = 50 (CRLB-derived; module doc block shows the σ_α/α ≈ 0.085 calculation at α=2.5) - MIN_RELATIVE_SE_VALIDATION = 0.10 (Clauset-Shalizi-Newman 2009 fig. 3 PL-vs-exp AIC-Δ > 4 boundary) fit_power_law (exploratory) keeps min_relative_se opt-in. Two new tests: rejects n < 50, passes on n=2000 with σ_α/α ≤ 0.10. FIX 6 — SciPy pin VERIFIED pyproject.toml already lists scipy>=1.16.2 in the canonical dependency block. test_falsification.py's binom.ppf import is backed by a pinned dep. FIX 7 — claim-governance audit Forbidden-word grep across research/systemic_risk/*.md and *.py: \b(production-ready|production-grade|empirically established| trading edge|trading signal|predictive system| early-warning system|proven|confirmed)\b → 0 matches. ("validated" appears only as an enum tier name in PROTOCOL.md / VALIDATION.md status diagrams, allowed.) FIX 8 — edge-case test expansion Added: single-node graph, all-zero exposure matrix (empty graph), whitespace-only label, None label, omega-inf input, omega-zero- variance returns zero finite ω, validation-mode tail-size rejection, validation-mode pass at sufficient n. FIX 9 — reproducibility bundle DOCUMENTED PROTOCOL.md § 5 already lists the full RunManifest contract. LIMITATIONS.md flags what is NOT yet implemented (real-data ingest, walk-forward). FIX 10 — PR final status PRESERVED C-SYSRISK-PHASE remains HYPOTHESIS in CLAIMS.md. README, PROTOCOL.md, VALIDATION.md all preserve the HYPOTHESIS / INSTRUMENTATION status. Tests: 146 passing (+8 from 138). Quality: mypy --strict / ruff / black clean on the diff. Pre- existing 5 jax_engine mypy errors persist on origin/main; out of scope. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Yaroslav Vasylenko <neuron7x@ukr.net> Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
research/systemic_risk/package — falsifiable test of the hypothesis that interbank phase-locking precedes banking-crisis events (C-SYSRISK-PHASE,HYPOTHESIStier inCLAIMS.md).HARD_FAILon any AUC ≤ 0.55,HARD_PASSon ≥2 crises with AUC ≥ 0.70 AND BH-corrected p ≤ 0.01,UNDECIDEDotherwise.core.kuramotoprimitives — does not introduce new physics.Modules
event_ledger.pyINV-EVT1,INV-EVT2topology.pyINV-TOP1..3; Boss et al. 2004phase_extraction.py(T, N)wrapper overcore.kuramoto.PhaseExtractorearly_warning.pyINV-K1,INV-EW1/EW2; Scheffer 2009 CSDfalsification.pyEnd-to-end rails verified
DEFAULT_LEDGERUSA filter):HARD_FAIL, AUC=0.455.HARD_PASS, AUC=0.92, p_BH=5e-4.Maintenance-hierarchy role
Sustainer (Layer 2). Emits a diagnostic score; never takes execution action. A future
HARD_PASSoutcome would only motivate promotion to a Protector — it does not itself protect any gradient.Status
mypy --strictclean on all new files.ruff+blackclean on all new files.HYPOTHESISuntil the battery returnsHARD_PASSon ≥2 independent crises with real data.Test plan
pytest tests/research/systemic_risk/— 57 passed locallymypy --strict research/systemic_risk/ tests/research/systemic_risk/— cleanruff check+black --check— cleanHARD_FAILHARD_PASS🤖 Generated with Claude Code