fintfm 0.5.5__tar.gz

fintfm-0.5.5/LICENSE

@@ -0,0 +1,202 @@
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

fintfm-0.5.5/PKG-INFO

@@ -0,0 +1,407 @@
+Metadata-Version: 2.4
+Name: fintfm
+Version: 0.5.5
+Summary: A tabular foundation model for financial risk — pretrained on synthetic company financials, so it predicts credit default with no training on your data
+Author: Mukharbek Organokov
+Author-email: Mukharbek Organokov <mukharbek.organokov@gmail.com>
+License-Expression: Apache-2.0
+License-File: LICENSE
+Classifier: Development Status :: 2 - Pre-Alpha
+Classifier: Intended Audience :: Financial and Insurance Industry
+Classifier: Intended Audience :: Science/Research
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Programming Language :: Python :: 3.13
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Classifier: Typing :: Typed
+Requires-Dist: numpy>=2.0
+Requires-Dist: pandas>=2.2
+Requires-Dist: scikit-learn>=1.5
+Requires-Dist: pyyaml>=6.0
+Requires-Dist: scipy>=1.18.1
+Requires-Dist: torch>=2.4
+Requires-Dist: catboost>=1.2 ; extra == 'bench'
+Requires-Dist: lightgbm>=4.0 ; extra == 'bench'
+Requires-Dist: xgboost>=2.0 ; extra == 'bench'
+Requires-Dist: pyarrow>=17 ; extra == 'bench'
+Requires-Dist: huggingface-hub>=0.34 ; extra == 'hf'
+Requires-Dist: kaggle>=1.6 ; extra == 'kaggle'
+Requires-Dist: kagglehub>=0.3 ; extra == 'kaggle'
+Requires-Dist: pyarrow>=17 ; extra == 'kaggle'
+Requires-Python: >=3.12
+Project-URL: Homepage, https://github.com/kabartay/fintfm
+Project-URL: Repository, https://github.com/kabartay/fintfm
+Project-URL: Changelog, https://github.com/kabartay/fintfm/blob/main/docs/CHANGELOG.md
+Project-URL: Checkpoint, https://huggingface.co/kabartay/fintfm-binary
+Project-URL: Measurement log, https://github.com/kabartay/fintfm/blob/main/docs/results/FINDINGS.md
+Provides-Extra: bench
+Provides-Extra: hf
+Provides-Extra: kaggle
+Description-Content-Type: text/markdown
+
+# FinTFM
+
+[![CI](https://github.com/kabartay/fintfm/actions/workflows/ci.yml/badge.svg)](https://github.com/kabartay/fintfm/actions/workflows/ci.yml)
+[![License](https://img.shields.io/badge/license-Apache--2.0-green.svg)](LICENSE)
+[![Python](https://img.shields.io/badge/python-3.12%20%7C%203.13-blue.svg)](pyproject.toml)
+[![Checkpoint](https://img.shields.io/badge/%F0%9F%A4%97%20weights-fintfm--binary-yellow)](https://huggingface.co/kabartay/fintfm-binary)
+[![TabArena](https://img.shields.io/badge/TabArena-93rd%20of%2095-critical)](docs/results/TABARENA.md)
+
+**A tabular foundation model for corporate credit risk, built from scratch and measured in
+public.**
+
+Given a table of labelled rows, FinTFM predicts new rows in a **single forward pass**, with your
+data supplied as context rather than trained on. There are no gradient steps at fit time and no
+per-dataset tuning — the model is pretrained once, on **synthetic data only**, and never sees a
+real table during training.
+
+That last property is the point. A model that provably never saw your benchmark cannot have
+memorised it, and in a regulated domain the ability to *demonstrate* that is worth more than a
+few points of accuracy.
+
+```python
+from huggingface_hub import hf_hub_download
+from fintfm.inference import FinancialTFMClassifier
+
+ckpt = hf_hub_download("kabartay/fintfm-binary", "v4-cellattn-labels.pt")
+clf = FinancialTFMClassifier(ckpt, device="cpu")
+clf.fit(X_train, y_train)            # stores the table as context; no training happens
+pd_estimates = clf.predict_proba(X_test)[:, 1]
+```
+
+The checkpoint is on Hugging Face at
+[**kabartay/fintfm-binary**](https://huggingface.co/kabartay/fintfm-binary) — 885K parameters,
+binary, up to 136 features, **Apache-2.0**. It is the checkpoint every published binary number
+below was measured on, so those results are reproducible against this file rather than a
+variant of it.
+
+## Status, stated plainly
+
+**FinTFM is a research codebase with a public claims ledger, not a product.** It exists to find
+out whether in-context tabular learning can be made competitive with gradient-boosted trees on
+financial tables, and to say honestly where it currently is not.
+
+| | |
+| --- | --- |
+| **External benchmark** | [TabArena](https://tabarena.ai), 27 binary datasets against 94 other methods: **rank 93 of 95** |
+| **On real credit panels** | calibration consistently among the best measured; discrimination consistently loses to tuned gradient boosting — both, on every panel tried |
+| **Licence** | Apache-2.0, **code and weights** — chosen separately, not inherited (see [Licensing](#licensing--provenance)) |
+| **Checkpoint** | [kabartay/fintfm-binary](https://huggingface.co/kabartay/fintfm-binary) — 885K, binary, ≤136 features |
+| **Tests** | 246, plus `ruff`, the openspec validator, a dependency-licence check and a documentation-link check — all in CI |
+| **Measurement log** | 121 numbered findings, each declaring how its numbers were produced |
+| **Problem types** | binary declared; multiclass and regression implemented but **not** declared (§121) |
+
+**Read [`docs/paper/CLAIMS.md`](docs/paper/CLAIMS.md) before anything else.** Every claim this
+project could make is tagged SURVIVES / SINGLE DRAW / SUPERSEDED / RETRACTED / OPEN against its
+evidence. That file, not this one, is the source of truth — and it records what stopped being
+true as prominently as what holds.
+
+> Throughout this repository, **§N** refers to entry N in
+> [`docs/results/FINDINGS.md`](docs/results/FINDINGS.md), a measurement log in which every number states how it
+> was produced (MEASURED, SMOKE-TEST, SIMULATED or ESTIMATED) and negative results are kept
+> deliberately, because they are what stops the same wrong conclusion being reached twice.
+
+## Results, as measured
+
+TabArena, 27 binary datasets, one fold each, against 94 other methods:
+
+| change | mean ROC-AUC | Elo | rank |
+| --- | --- | --- | --- |
+| label encoding (§98) | 0.7642 | 662 | 93 / 95 |
+| **+ out-of-fold target statistics** (§101) | **0.7823** | 765 | 93 / 95 |
+| + multiclass-capable prior (§105) | 0.7817 | **813** | 93 / 95 |
+| + 5.0M parameters, matched task count (§114) | 0.7774 | 751 | 93 / 95 |
+| + tree-structured prior (§115), two seeds | 0.7926 / 0.7913 | 842 / 826 | 93 / 95 |
+
+**One change has moved the number on the benchmark that matters, and it was preprocessing
+rather than architecture.** The tree prior adds a further +0.0094 here and replicates across two
+seeds — but costs **−0.0221 average precision on credit data** (§116), negative on 5 of 5 folds
+with three surviving multiple-comparison correction at p < 0.001, so it ships off by default.
+Everything else tried — parameter scale at matched task volume, training volume, a widened
+structural-causal prior — has been null or negative. **The rank has never moved.**
+
+## What this project does that others do not
+
+Three things, stated at their true weight:
+
+**Incoherent PD term structures are impossible by construction.** The field's standard
+per-horizon construction produces non-monotone cumulative default curves for **39% of firms** on
+real data — a firm whose 3-year default probability is below its 2-year. A hazard head makes
+that unrepresentable, at zero measured accuracy cost.
+
+**A severe capacity defect was found by measurement, not review.** The original architecture had
+capped discrimination regardless of true task difficulty, caught against an exactly-known
+Bayes-optimal AUC. Its cause was isolated to the architecture rather than the prior after
+eliminating seven other candidates one at a time, then closed by a two-way cell-attention change
+worth +0.042 average precision on real data.
+
+**The record corrects itself in public.** This README has been wrong about its own results at
+least twice (§107, §112). Both are retracted in place, with the reasoning kept. No peer project
+publishes a document that tags its own claims RETRACTED as prominently as its wins, and in a
+domain where a model must arrive with its own validation evidence, that is the differentiator.
+
+## How it works
+
+1. **Priors** (`src/fintfm/prior/`) — synthetic task generators, sampled and mixed per batch
+   (`mixture.py`).
+   - `financial.py` — a structural generative story for company balance sheets, P&L and
+     default labels: accounting identities, sector and macro effects, a wide derived-ratio
+     family, missingness, a sharpness (signal-to-noise) knob, and a default-rate envelope that
+     can reach real low-default-portfolio rates (~0.2%).
+   - `scm.py` — a generic random-graph structural-causal-model prior (the TabPFN/TabICL-style
+     idea): random layered functions with several nonlinearities, for general nonlinear and
+     multiclass structure the financial prior does not cover.
+   - `tree.py` — ensembles of oblivious decision trees, generating the **axis-aligned,
+     piecewise-constant** structure real tabular data is full of and the other priors do not.
+     Added on *distinctiveness* grounds after `fintfm-priorscore` measured it as the only
+     member of the mixture with a positive tree-versus-linear gap (§112). Off by default
+     (`p_tree=0.0`).
+   - `trivial.py`, `crossed.py` — diagnostic-only priors used to isolate specific hypotheses
+     (whether the architecture can learn at all; whether a prior's features or its label
+     mechanism carries a measured effect). Not part of the default training mixture.
+2. **Model** (`src/fintfm/modeling/model.py`) — a from-scratch Transformer. Cells are embedded
+   individually and carry a random per-task column identity (so the model can tell columns
+   apart without a positional embedding, which would break invariance to column order).
+   Columns attend to each other within a row; optionally — `ModelConfig.n_cell_blocks`,
+   currently experimental — cells also attend across rows *within one feature* before that,
+   giving each column a distribution-derived identity rather than only a random tag. Rows are
+   then pooled and attend to context rows to perform in-context learning. An optional hazard
+   head (`modeling/hazard.py`) produces a cumulative PD term structure that is monotone by
+   construction. Independent implementation of ideas described in the public TabPFN / TabICL
+   literature — no code or weights from any existing project (see Licensing below).
+3. **Training** (`src/fintfm/modeling/train.py`, `fintfm-train`) — infinite synthetic data, one
+   gradient step per fresh batch, cosine LR schedule, checkpointing, a held-out quality metric
+   that reports per-task AUC (not the pooled-across-tasks number, which is inflated whenever
+   task base rates differ) and Brier skill against a base-rate-only predictor.
+4. **Inference** (`src/fintfm/inference/`) — an sklearn-compatible `FinancialTFMClassifier`:
+   `fit()` stores the table as context, `predict_proba()` runs the frozen network. Several
+   context-construction strategies (uniform, balanced, hybrid, retrieval, prototype) and a
+   base-rate correction, plus ensembling over column-identity draws to recover the
+   distributional column-order invariance a random identity trades for expressiveness.
+   `inference/categorical.py` encodes categorical columns as **out-of-fold** smoothed target
+   statistics, because the model reads every cell as an ordered scalar and label encoding's
+   arbitrary order is measurably worse than no order at all (§100). The out-of-fold step is
+   not a refinement: the naive form puts a row's own label into its own encoding, which makes
+   the context self-predictive and the feature absent at query time — it harms the model
+   rather than flattering the score, so it survives careless validation.
+   `inference/regressor.py` adds a `FinancialTFMRegressor` on the same frozen network and no
+   new architecture: a continuous target cut into K quantile bins is an integer index over K
+   outcomes, so the existing classification head regresses as-is, and the output is natively
+   **distributional** — quantiles and prediction intervals come free, and the predicted
+   density can be bimodal, which is what loss given default actually is and what a Gaussian
+   head cannot represent.
+5. **Evaluation** (`src/fintfm/evaluation/`, `src/fintfm/experiments/`) — real corporate-default
+   panels (V4FinBench via its published protocol, `fintfm-v4protocol`; Polish and Taiwan
+   bankruptcy, `fintfm-bench`), an out-of-time harness, a synthetic capability suite
+   (`fintfm-capability`, including a class-count sweep for multiclass, §99) with probes of a
+   *known* ceiling — some with a closed-form
+   Bayes-optimal AUC — specifically built to catch a model that looks fine on average while
+   capped in a way an aggregate score cannot see. Average precision is reported alongside
+   ROC-AUC everywhere, and read first at low base rates (ROC-AUC's chance floor is 0.5
+   regardless of prevalence; AP's floor is the prevalence itself, so it stays legible at the
+   base rates this project's target segment actually has).
+   `fintfm-priorscore` scores a *prior* rather than a model, on the three criteria the
+   literature converged on — performance, diversity, distinctiveness — using **fitted**
+   baselines only, so it cannot confuse "the prior lacks this structure" with "our model cannot
+   learn it". Its first version could, and §112 is the record of what that cost.
+
+## Quickstart
+
+```bash
+uv sync --extra bench --extra hf   # naming one extra uninstalls the others
+uv run pytest                      # 246 tests; 1 skip is expected, more means look
+```
+
+### Using a checkpoint
+
+The estimators are scikit-learn compatible. `fit()` **stores** the table as context — it takes
+no gradient steps — and `predict_proba()` runs the frozen network once.
+
+```python
+from fintfm.inference import FinancialTFMClassifier, FinancialTFMRegressor
+
+clf = FinancialTFMClassifier(ckpt, device="mps")   # ckpt from hf_hub_download, above
+clf.fit(X_train, y_train)          # stores context; no training happens
+proba = clf.predict_proba(X_test)[:, 1]
+
+reg = FinancialTFMRegressor("runs/v4-regression.pt", n_bins=10)
+reg.fit(X_train, y_cont)
+point = reg.predict(X_test)                     # distribution mean
+lo, hi = reg.predict_interval(X_test, 0.8)      # 80% interval, free from the same head
+```
+
+Categorical columns need encoding before they reach the model — it reads every cell as an
+ordered scalar, and label encoding is measurably worse than no order at all (§100). Use
+`fintfm.inference.categorical.CategoricalTargetEncoder`, which is out-of-fold on the context
+rows for reasons that are **not** optional; see [How it works](#how-it-works).
+
+### Training a checkpoint
+
+```bash
+uv run fintfm-train --steps 300 --d-model 32 --d-cell 16 --n-layers 2 --max-features 16 \
+    --max-classes 2 --device cpu --out runs/v0-smoke.pt   # pipeline check, a couple of minutes
+uv run fintfm-bench --model runs/v0-smoke.pt --credit      # real corporate-default panels
+```
+
+That smoke config exists to check the pipeline runs, not to produce a usable checkpoint — see
+[`docs/infra/COMPUTE.md`](docs/infra/COMPUTE.md) for measured step costs at real scale. A real pretraining
+run (thousands of steps, `--d-model 128`+, `--max-features` matching your data) takes
+hours-to-a-day and needs a GPU: `--device mps` on Apple Silicon, `--device cuda`, or see
+[`docs/infra/HF_JOBS.md`](docs/infra/HF_JOBS.md) for the Hugging Face Jobs recipe this project actually
+uses for training runs. **Check `uptime` before running anything heavy locally** — see
+[`CLAUDE.md`](CLAUDE.md) for why.
+
+Both `fintfm-bench --credit` (Polish/Taiwan need 64/95 features) and `fintfm-v4protocol`
+(V4FinBench needs 136) will print `SKIPPED for fintfm: model takes 16 features, data has N` —
+correct, expected behaviour for the smoke checkpoint above, not a bug. Baselines still run and
+score normally; only the FinTFM arm needs a checkpoint pretrained with a matching
+`--max-features` to be evaluated.
+
+## Configuration
+
+The numbers experiments use — split years, context sizes, strategies, seeds, scoring
+thresholds, the prior's default-rate and sharpness envelope — live in
+[`src/fintfm/configs/default.yaml`](src/fintfm/configs/default.yaml), not scattered through
+the code. Every entry point takes `--config` with a file that is **deep-merged** over that
+default, so an override carries only what it changes:
+
+```bash
+uv run fintfm-ctxsweep --model runs/m.pt --config configs/context-sweep-3seed.yaml
+uv run fintfm-v4oot    --model runs/m.pt --config configs/retrieval-best.yaml
+```
+
+Resolution order is packaged default → `--config` (or `$FINTFM_CONFIG`) → explicit CLI flag.
+Runs print the layers they used and record them as `config_sources` in their output JSON.
+Unknown keys are refused by name rather than ignored, and an overlapping train/test split is
+rejected before the run starts. [`configs/README.md`](configs/README.md) explains what
+belongs in configuration and what deliberately stays in code; several files under
+[`configs/`](configs/) are **deliberately diagnostic, not production** (their own headers say
+so), used to isolate a single variable while chasing a specific finding.
+
+## Licensing / provenance
+
+Everything here — the prior, the architecture, the training loop — is original code written
+for this project. **No weights, datasets, or source from Neuralk (Seldon), Fundamental
+(NEXUS), Google TabFM, TabPFN, TabICL, TabDPT, or any other tabular-foundation-model product
+are used or may be added.** Those are cited in project discussion purely as public
+research/product context. Every third-party dataset or dependency's licence is checked before
+use, and a *weight* licence is checked separately from its *code* licence, every time — never
+assumed from a prior check. Direct dependencies are all permissive — numpy, pandas and
+scikit-learn BSD-3, scipy BSD-3, torch BSD-3, PyYAML MIT, and the optional benchmark extras
+lightgbm MIT, xgboost Apache-2.0, catboost Apache-2.0, pyarrow Apache-2.0 — and
+`scripts/check_licences.py` enforces that in CI rather than leaving it to this paragraph.
+
+**One honest exception, which CI found and this paragraph previously did not mention.** On
+**Linux**, `torch` pulls in around fifteen NVIDIA CUDA runtime packages that are **NVIDIA
+Proprietary**, not permissive. They are a transitive runtime dependency rather than a choice
+made here, this project does **not redistribute them** — `uv build` produces a pure-Python
+wheel and pip fetches them from PyPI under NVIDIA's own terms — and they are absent on macOS,
+which is why a check run only on a developer's Mac reported everything permissive while CI did
+not. **Anyone shipping this in a product must read NVIDIA's EULA themselves**; it is not
+something this repository can assert on their behalf.
+
+## What would change the picture
+
+Stated so the project is falsifiable rather than open-ended. The ~0.035 deficit to the field
+has no measured axis of variation, and these are the remaining candidates:
+
+| still open | why it is a candidate | status |
+| --- | --- | --- |
+| **factorized attention** | the encoder is memory-bound at every turn, and three peers independently chose the cheaper form | proposed (44.x), prior art recorded |
+| **a joint objective** | `p(x, y \| D)` rather than `p(y \| x, D)` makes every column a training signal, not just the target | proposed (48.15), scoped as a measurement before a rewrite |
+| **prior design** | the only lever a peer's own ablations identify as decisive | instrument built and partly falsified — see below |
+
+### Closed by measurement
+
+Not abandoned — **tested and ruled out**, which is the more useful half of the record and the
+reason the list above is short:
+
+| candidate | what the measurement said |
+| --- | --- |
+| parameter scale | **−0.0049** at 5.7× the parameters on *matched* task volume (§114). A peer's published curve returns +0.0049 R² for 16.7×, against a 0.035 deficit. |
+| training volume | null at 5× (§93); **+0.0028** at 2× (§114). |
+| depth over width | **could not be scored at all** — the harness's per-dataset time limit, after 8 of 27 datasets (§114). Depth costs inference time and this model is already slow. |
+| `column_id_dim` | peaks at the value chosen by accident; all four non-peak arms below it on all five folds (§104). |
+| a widened structural-causal prior | **p = 1.000** against its own matched control. |
+| a tree-structured prior | +0.0094 on general tabular data, replicated across two seeds (§115) — and **−0.0221 average precision on credit** (§116). Ships off by default. |
+
+**The prior-scoring instrument is half-confirmed and half-falsified.** It predicted the
+general-tabular gain above before it was measured (§112 → §115). A second arm, added to ask the
+credit question, then predicted the wrong direction on its first real test — raising the
+structural-causal prior's share cost **−0.0339 AP**, 0 of 5 folds, all five significant (§118).
+So it predicts **breadth, not fit**, and no claim rests on its credit arm.
+
+## Reproducing the measurements
+
+Every number in [`docs/results/FINDINGS.md`](docs/results/FINDINGS.md) names the command that produced it.
+The entry points:
+
+| command | what it measures |
+| --- | --- |
+| `fintfm-v4protocol` | V4FinBench under its **published** five-fold protocol — the benchmark this project's claims rest on |
+| `fintfm-bench --credit` | Polish and Taiwan bankruptcy panels |
+| `fintfm-v4oot` | out-of-time split, which the published protocol is not |
+| `fintfm-capability` | synthetic probes with a **known** ceiling, some with a closed-form Bayes-optimal AUC |
+| `fintfm-priorscore` | scores a *prior*, not a model, on performance / diversity / distinctiveness |
+| `fintfm-ctxsweep` | context construction, which explains more variance than model family (§5) |
+
+Two conventions worth knowing before reading any of it. **Every number declares how it was
+produced** — SMOKE-TEST, MEASURED, SIMULATED or ESTIMATED — because in an ML repository a wrong
+number does not crash, it looks like a result. And **negative and superseded results are kept**,
+because they are what stops the same wrong conclusion being reached twice.
+
+## Where this is going
+
+[`docs/`](docs/) holds the project's reasoning, indexed in [`docs/README.md`](docs/README.md).
+
+- [`docs/paper/CLAIMS.md`](docs/paper/CLAIMS.md) — **start here.** Every claim this project
+  could make, tagged by status, newest evidence wins.
+- [`docs/results/FINDINGS.md`](docs/results/FINDINGS.md) — the full measurement log, numbered sequentially
+  (121 entries and counting), each declaring how its numbers were produced.
+- [`docs/design/DECISIONS.md`](docs/design/DECISIONS.md) — why the project is built the way it is, and what
+  would reverse each choice.
+- [`docs/roadmap/STRATEGY.md`](docs/roadmap/STRATEGY.md) — the plan of record.
+- [`docs/design/ARCHITECTURE.md`](docs/design/ARCHITECTURE.md) — how the model works, in more depth than
+  this file.
+- [`docs/results/TABARENA.md`](docs/results/TABARENA.md) — the external-evaluation recipe, the coverage
+  fraction any score must carry, and four silent failure modes including a results cache that
+  returns stale numbers after a preprocessing change.
+- [`docs/paper/RELATED_WORK.md`](docs/paper/RELATED_WORK.md) — the competing models, read from
+  primary sources, with what each costs this project's positioning. Its digest is the shortest
+  useful summary of why the remaining lever is the prior.
+- [`docs/results/POSTMORTEM.md`](docs/results/POSTMORTEM.md) — wrong diagnoses, each caught by measurement
+  rather than review, kept on the record deliberately.
+- [`openspec/changes/`](openspec/changes/) — the live roadmap as structured proposals with
+  verifiable tasks, not prose. `uv run python openspec/tools/validate.py` checks every
+  proposal is well-formed and every finding declares its provenance.
+
+Short version: the mechanism claims ("no training on your data", "no feature engineering") are
+already owned by better-funded competitors, so the thesis is not the mechanism. It is a credit
+model that arrives with its own validation evidence — calibrated, auditably free of benchmark
+contamination, and eventually backed by a pre-registered forward track record that cannot be
+bought — plus a public, self-correcting record of what has and has not been shown to be true.
+
+## Citing
+
+A `CITATION.cff` is in the repository root, so GitHub's **"Cite this repository"** button
+renders BibTeX and APA directly.
+
+Cite the **repository** for the method, the measurement log or any finding; cite the
+**checkpoint** ([kabartay/fintfm-binary](https://huggingface.co/kabartay/fintfm-binary)) when
+the specific weights matter to what you are reporting. They are different artifacts and a
+reader can only check the one you name.
+
+## Contributing, and what cannot come in
+
+Issues and discussion are welcome. Two hard boundaries, both of which protect the only claim
+this project has that competitors do not:
+
+**No code, weights, or training data from any tabular-foundation-model product** — TabPFN,
+TabICL, TabDPT, LimiX, Nori, MITRA, or any other — may enter this repository. Their published
+papers are legitimate to read and cite, and [`docs/paper/RELATED_WORK.md`](docs/paper/RELATED_WORK.md)
+does exactly that for ten of them. Ingesting any of it would destroy the provenance argument
+above, and nothing would warn us.
+
+**Check a weights licence separately from its code licence, every time.** Four of the ten peer
+projects read for this release ship permissive code with **non-commercial weights**, a split
+invisible from a repository's headline licence badge. One of them restricts commercial use of
+the model's *output*, not merely the weights.