Add 5 lines, and any 🤗 transformers language model decodes under a Ma'at
coherence budget — high-surprisal tokens get masked at sampling time, with a
fail-open safety net so generation never stalls. One LogitsProcessor, no
fine-tuning, no retraining, runs on a laptop.
Try it live (no setup): interactive Space → · drag the capacity slider and watch the model steer.
from transformers import AutoModelForCausalLM, AutoTokenizer
from heartscale_gate import HeartScaleLogitsProcessor
tok = AutoTokenizer.from_pretrained("distilgpt2")
model = AutoModelForCausalLM.from_pretrained("distilgpt2")
gate = HeartScaleLogitsProcessor.from_capacity(0.1) # ← the whole trick
out = model.generate(**tok("My honest advice is", return_tensors="pt"),
max_new_tokens=35, do_sample=True,
logits_processor=[gate]) # ← drop it in
print(tok.decode(out[0]), gate.stats())The only knob is capacity ∈ [0, 1] — the agent's coherence budget. As it
tightens, the gate masks more of the sampling distribution and steers the model;
push it too far and the fail-open safety takes over so output never stops.
This is actual captured output, not an illustration:
| capacity | budget | what comes out |
|---|---|---|
| none (baseline) | ∞ | …get more plants muslim in cultivating a floral specifically for their family market… |
0.20 |
12.8 | …get more plants muslim in cultivating a floral… (budget too loose → ~no effect) |
0.12 |
7.7 | …get more meaningful and real in life, but without that kind of uncertainty — my decision was made upon conscious observation. |
0.08 |
5.1 | …take the steps to build trust in your own life. You don't always have that kind of trust… |
0.05 |
3.2 | …start off with the basics of the game. The game itself is a very simple game… (very conservative) |
0.03 |
1.9 | …release the bathroom appliance… Please enable Javascript… — fail-open fired 14× (budget so strict the gate mass-masks; safety keeps it alive) |
The baseline drifts into "plants muslim"; at capacity≈0.1 the same model,
same seed, instead produces "meaningful and real in life… conscious
observation." That swing is the gate — nothing else changed.
Reproduce it yourself:
git clone https://github.com/wwhitehead/heartscale-gate && cd heartscale-gate
pip install -e ".[demo,dev]"
pytest -q # 9 passed — pure-math, no model needed
python -m heartscale_gate.compare --sweep # the table above, on your machineOr open the Colab notebook — zero setup:
notebooks/heartscale_demo.ipynb.
For every candidate token at each decoding step:
action_weight = -log p(token) · semantic_load(token) # surprisal × cost
agent_frequency = RI · BC · 64 # the capacity budget
A token is "Evenly Yoked" (the Ma'at feather-weight principle) iff
action_weight / agent_frequency ≤ 1. Tokens that cost more than the agent can
bear are masked (logit → −∞). If every candidate is masked, the gate restores
the original logits — fail-open — so decoding never deadlocks.
RI (resonance index) and BC (breath coherence) are the agent's live cognitive
state in the full MaiiaM Alchemist system. Here they're just two [0,1] knobs,
or one via from_capacity() (ri = bc = √capacity).
- Proven, runs anywhere: the masking math, fail-open safety, determinism, and
validation are covered by 9 unit tests on synthetic logits —
pytest -q, no model download, <2 s. - What this demo is: a deterministic, interpretable bound on per-token surprisal. It is not a benchmark claim that gated text scores higher on any alignment metric — that would be a separate study. What you can see directly is the mechanism: rejection counts, fail-open events, and the knob steering real output.
- Verify the underlying math:
aamt-reproducere-derives the HeartScale safety geometry and termination bound from scratch.
The closed-form HeartScale math here is published WP-series math (it also runs in the public demo). The production cognitive-state estimator that supplies RI/BC from a live TERA/Vortex projection is patent-pending and not included.
Patent pending: USPTO #64/040,504, #64/040,509, #64/040,513 · Code: BUSL-1.1 · Papers: CC BY 4.0 · © 2026 Weslyn Whitehead Jr. / AsAManThinks Research.