You say what must be true before something can be done. Messages arrive, get corrected, get withdrawn. EASE-Delta reads each message against each requirement once, re-reads only what a change touches, and tells you what is ready, what is blocked and why, and which single question is worth asking. It proposes; it never acts.
The reader at the heart of EASE-Delta answers supports, refutes or not enough info, and it was trained to say not enough info when a passage is about something else. It runs here, in your browser: the text you type is not sent anywhere.
A client hand-off: four requirements, two actions, twelve events, including a duplicate, a late copy of an old message, a withdrawal and a bounced payment. Each state below is what the released 395M model produced. It is a recording; nothing was edited.
Learned parts read language. Exact code handles versions, precedence, logic and planning.
Readings are cached per (requirement, message). A change re-reads the pairs it touches: about 2–7% of the tokens of re-reading the task. Duplicates and stale copies cost nothing.
The cached state equals a full rebuild bit for bit (0 of 195,434 values differed). A person's correction takes effect at once and can be undone exactly.
Trained with unrelated passages as not enough info: 0.6% read as evidence, against 45% for a public NLI model.
Every proposal shows the records it rests on. It names the one question worth its cost, and a READY action still waits for a person.
Hypotheses and decision rules were written and hashed before any test data was read; the failures are reported too. Tasks are generated; their text is human-written.
| What was measured | Result |
|---|---|
| Correct ready / blocked / needs-info decisions | 91.6% standard tasks · 86.8% larger tasks |
| Against a dense reader of the whole task (same backbone and training) | +2.2 to +5.5 points, fewer stale decisions |
| Larger reader (395M) against the base reader (149M) | +2.0 and +2.2 points, close to the prediction made beforehand |
| Update after a change, Apple M4 Max | 59 ms GPU · 0.2–0.6 s CPU |
| VitaminC test (claims against Wikipedia revisions) | 91.5% |
| Conflicts that follow only from a consequence | 45% recognised (a known weakness) |
Python 3.11 or later. The model downloads from the Hugging Face Hub on first use.
pip install "ease-delta @ git+https://github.com/jithinsaireddy/ease-delta"
from ease import EvidenceReader, Tracker
reader = EvidenceReader.from_pretrained() # jithinpothireddy21/ease-delta
task = Tracker.define(
requirements={"approved": "Acme has approved the final design.",
"date": "Acme has confirmed the delivery date."},
actions={"send_packet": "approved and date"},
reader=reader)
task.add("mail-1", "Email from Acme: we approve the final design, please go ahead.")
task.add("mail-2", "Email from Acme: we confirm delivery on 12 March.")
print(task.status()) # ready / blocked / needs info, and why
task.update("mail-1", "Email from Acme: we withdraw our approval.")
print(task.explain("send_packet"))
from transformers import pipeline
reader = pipeline("text-classification", model="jithinpothireddy21/ease-delta-reader", top_k=None)
reader({"text": "The client has approved the final design.", # claim first
"text_pair": "Email from the client: we approve the final design."}, # evidence second
truncation=True)
# also: sentence_transformers.CrossEncoder("jithinpothireddy21/ease-delta-reader")
ease serve --model release/edge --aggregator release/aggregator # one person, 127.0.0.1:8791
ease serve --model release/edge --aggregator release/aggregator --multi # workspaces with keys, quotas, webhooks
curl -s localhost:8791/tasks/from-template -H 'content-type: application/json' \
-d '{"template": "client-onboarding", "task_id": "acme", "params": {"client": "Acme"}}'
Docs: README · Guides · Models: system, reader, reader (base)