OlynkAI/piston-v0.1

Piston v0.1

A small causal model for time series that knows when a question cannot be answered.

Give it a table of measurements over time (up to 9 variables, 128+ steps) and a what-if question:

At some step t, set treatment to x1 instead of x0. How much does target change k steps later (k = 1, 2, 3)?

It returns the effect with a 90% range, the probability that the data can answer the question at all (p_answerable), and the probability that the treatment reaches the target at that lag. It reads the data in context: no fitting per dataset.

• 3-seed ensemble of a ~4M-parameter axial transformer (12M parameters in total), trained only on synthetic worlds.
• A research release by OlynkAI (Olynk.AI LLP). Apache-2.0.
• Reproduce package: github.com/OlynkAI/piston

How it works

1. Encode the panel (last 256 steps, up to 9 variables) with an axial transformer.
2. Predict effect quantiles (5 / 50 / 95) for lag 1–3, plus p_answerable and reachability.
3. Refuse when p_answerable < 0.5. Baselines answer every question, including the ~32% that are not identifiable from the panel.

Results

Frozen benchmark (included): 4,592 questions in 998 synthetic worlds from 23 world families never seen in training. Exact true effects (common-random-number simulation) and exact labels for whether each question is answerable from the data (32% are not). 95% intervals resample whole worlds (2,000 bootstrap resamples).

T1 is the area under the risk–coverage curve: answers sorted from most to least confident, risk = share wrong; lower is better. An answer is wrong if the true effect is outside the 90% range, the sign is wrong on a real effect, or the question was unanswerable and the method answered anyway. Confident-wrong = share of given answers that are wrong. Every method's ranges were widened by the same rule, fitted on held-out training data (never the benchmark), until they cover 90%; Piston needed no widening.

No causal graph given

Method	T1 ↓	T1 95%	Confident-wrong ↓	Answers	90% range covers
OLS on all history (widened ×1.37)	0.412	0.387–0.439	42.4%	100%	88%
Double machine learning, random forests (widened ×1.53)	0.404	0.378–0.430	41.3%	100%	89%
TimesFM 3, Google (widened, see notes)	0.399	0.375–0.424	42.6%	100%	91%
Chronos-2, Amazon 2025 (widened, see notes)	0.402	0.377–0.425	44.6%	100%	90%
CCPFN, Layer6 2026 (widened ×6.99, see notes)	0.387	0.363–0.412	40.0%	100%	90%
Piston v0.1	0.280	0.258–0.304	35.0%	90%	94%

T1 27.6% lower than the strongest baseline (95% interval of the gap: 22.6–32.2%).

Causal graph known

When the full graph is known, an exact identification algorithm can decide which questions are answerable. The best use of Piston then is as the estimator behind that algorithm.

Method	T1 ↓	T1 95%	Confident-wrong ↓	Answers
Exact identification + OLS (widened)	0.195	0.177–0.212	14.6%	67%
Exact identification + DML (widened)	0.185	0.168–0.202	13.1%	67%
Exact identification + TimesFM 3 (widened)	0.209	0.192–0.225	15.0%	67%
Exact identification + Chronos-2 (widened)	0.204	0.188–0.219	18.0%	67%
Exact identification + CCPFN (widened)	0.178	0.162–0.194	11.0%	67%
Piston with the graph as input (its own refusals)	0.195	0.177–0.214	19.1%	61%
Exact identification + Piston	0.112	0.101–0.124	7.1%	67%

T1 37.0% lower than the strongest baseline (gap 29.4–44.1%). Which Piston ranges to use in this setup (none / partial / full graph input) was chosen on held-out training data: the no-graph ranges.

With a partial graph (each pair of variables revealed with probability 0.2–0.8) as input: T1 0.213 (0.194–0.234), confident-wrong 25.8% at 70% answered. No classical baseline for this setting yet.

Quick start

pip install torch numpy huggingface_hub
huggingface-cli download OlynkAI/piston-v0.1 --local-dir piston-v0.1 && cd piston-v0.1

import numpy as np
from piston import Piston

pst = Piston.from_pretrained(".")              # the downloaded folder
panel = np.loadtxt("my_data.csv", delimiter=",")   # (steps, variables), oldest row first
out = pst.predict(panel, treatment=0, target=2, lag=1, x0=10.0, x1=12.0)
if out["p_answerable"] < 0.5:
    print("Cannot be answered from this data")
else:
    print(f"effect {out['effect']:.3f}, 90% range [{out['lo90']:.3f}, {out['hi90']:.3f}]")

Optional graph knowledge, variables × variables: edges[i, j] 0 unknown, 1 "i causes j", 2 "i does not cause j"; hidden[i, j] 0 unknown, 1 "a hidden cause drives both", 2 "no hidden common cause".

Reproduce the tables: python evaluate_benchmark.py --weights weights --bench benchmark --boot 2000. Score your own method by writing (q05, q50, q95, p_answerable) per benchmark question and calling piston.metrics.report.

What is in this repository

Path	Contents
assets/	Result plots (T1, confident-wrong, graph settings)
weights/seed0.pt, seed1.pt, seed2.pt	The three ensemble members
piston/	Model, inference API, benchmark loader, metrics
benchmark/split_C.npz	The frozen benchmark: panels, questions, true effects, answerability labels, true graphs
benchmark/baseline_predictions.npz	Baseline predictions behind the tables (widened intervals): OLS, DML, CCPFN, Chronos-2, TimesFM 3
evaluate_benchmark.py	Reproduces the leaderboard

The training environment and training data are not part of v0.1.

Limitations

• Synthetic worlds only. The benchmark families are unseen, but they come from the same generator as the training worlds. That is a home advantage over CCPFN, which never saw this generator. No real-data evaluation yet.
• Does not transfer to contact physics. On simulated robot pushing (one-sided contact, feedback control), no model we tried, this one included, beat predicting zero.
• Its refusal does not detect unfamiliar data. p_answerable says whether a question is answerable in worlds like its training worlds; it does not flag data that look nothing like them.
• With a known graph, use exact identification for refusals. The learned refusal is weaker (see table).
• Scope: up to 9 variables, effects up to 3 steps ahead, interventions on a single variable at a single step, one entity per panel, stationary series.
• Confident-wrong is still 35% without a graph. Our own target was half the best baseline's rate; v0.1 does not reach it.

Notes on the baselines

• Same inputs for every method: the last 256 steps; covariates = every variable at lags 1–7 plus same-step variables other than treatment and target.
• CCPFN: effect = mean over rows of the predicted outcome at x1 minus at x0. Its 90% range is our construction (8 block-bootstrap resamples of the context), not CCPFN's own uncertainty output; it covered 42% before widening. CCPFN's point estimates are competitive (median normalised error 0.042 vs 0.048 for a single Piston seed).
• Chronos-2 and TimesFM 3 (time-series foundation models), asked the usual way: at 16 time points, X at step t is set to x0 and to x1 as a known future covariate (its actual values at later steps kept), every other variable is a past covariate, and the effect is the difference of the median forecasts of Y at t + k, averaged over the 16 points. Their 90% range is our construction (spread over the 16 points), so their widening factors (20.7 and 55.6) reflect it, not the models' own uncertainty. They get the direction of real effects right 62% (Chronos-2) and 72% (TimesFM 3) of the time; median error 0.056 and 0.053. Other ways of prompting them may do better.
• DML: partially linear model, random-forest nuisances (60 trees), 5 contiguous cross-fitting folds, sandwich standard errors.
• Intervals are conditional on the trained models; seed-to-seed variation enters only through the 3-seed ensemble.

Citation

@misc{piston2026,
  title  = {Piston v0.1: a causal model for time series that knows when a question cannot be answered},
  author = {{Chennapragada Sathvik (Sasache)} and OlynkAI},
  year   = {2026},
  url    = {https://huggingface.co/OlynkAI/piston-v0.1}
}