chatpbc1/pally-endpoint

Pally — Business Consulting AI with Verifiable Math and Self-Correcting Rules (Docs)

Pally (Pally-1.7B) is a 1.7B-parameter business-consulting language model built by ChatPBC Labs. It is unusual for its size in three ways:

1. It shows its math. When a question involves money, margins, months, or ratios, Pally writes the arithmetic as visible [calc: 412000 / 31000] steps. A sandboxed runtime (not the neural network) executes each step and substitutes the true result before you see the answer.
2. It verifies and revises. In System-2 mode the runtime checks every [calc:]/[verify:] marker; failed expressions are fed back to the model, which revises its answer (up to 8 iterations).
3. It edits its own operating rules. Pally's behavior rules live in an external, versioned config ("constitution", "tool_rules", "metacognitive_rules", "refusal_rules", "playbook"). The model can rewrite them at inference time through a sandboxed [cortex_edit] channel. Every edit is logged, capped, persisted, and revertible. This is called POWER MODE (round cortex1-v8).

Weights: chatpbc1/Pally · Paper: RESEARCH_CORTX1.md · Developer guide: DEVELOPER_GUIDE.md · Chat UI (when demo is online): pally-console Space

---

What Pally is good at — and what it is not

Strength	Weakness (honest)
Unit economics: LTV, CAC, payback, runway, margins — with worked, code-verified steps	Not a frontier model (≈49/100 vs GPT-OSS 20B's 82/100 on a consulting bench)
Uses your digits, not memorized training examples (v7 digit-binding training)	[calc:] firing is not 100% — on missed cases it reverts to mental math, which a 1.7B model often gets wrong
Stable persona: warm, direct consultant; never invents statistics; one concrete next action per answer	Knowledge is frozen at training time; no live data
Tiny and private: one consumer GPU, or CPU (slow); Apache-2.0 weights	Short answers by design; not for 2,000-word memos
Transparent self-improvement: watch every rule edit it makes	Self-edits need human review in production (_edit_log)

Rule of thumb: for any number that matters, require the [calc:] steps or verify independently.

---

Using the model

Quickstart (transformers)

from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("chatpbc1/Pally", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("chatpbc1/Pally", trust_remote_code=True,
                                             torch_dtype="auto")
msgs = [{"role": "user", "content": "We have $412k and burn $31k/mo. Runway?"}]
inputs = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt")
out = model.generate(inputs, max_new_tokens=400, do_sample=False)
print(tok.decode(out[0][inputs.shape[1]:], skip_special_tokens=True))

Plain transformers gives you the persona, not the calculator or verifier. For those, serve it with the CORTX-1 runtime (below).

CORTX-1 serving (calculator + verifier + POWER MODE)

The serving stack is a single Python file (FastAPI). Serve it on any GPU/CPU box:

pip install fastapi uvicorn pydantic torch transformers
python cortx-1-power-mode-v8.py     # loads the model, starts API on :8000

Then call the OpenAI-compatible API:

curl http://localhost:8000/v1/chat/completions -H 'Content-Type: application/json' \
  -d '{"model":"pally-1.7b",
       "messages":[{"role":"user","content":"AOV $37, margin 52%. Contribution per order?"}],
       "cortex":{"system2":true}}'

from openai import OpenAI
c = OpenAI(base_url="http://localhost:8000/v1", api_key="none")
r = c.chat.completions.create(
    model="pally-1.7b",
    messages=[{"role": "user", "content": "We burn $31k/mo with $412k saved. Runway?"}],
    extra_body={"cortex": {"system2": True}})
print(r.choices[0].message.content)

Full endpoint and config reference: see the Developer Guide.

---

POWER MODE — the self-correcting loop (v8)

user question
   |
   v
gate: System-1 (one pass) or System-2 (deliberate)
   |
   v
generate --> parse [calc:]/[verify:] markers --> sandboxed AST evaluator
   |                    | fails?                      |
   |                    +--> REVISION request <-------+
   |                              |
   v                              v
[cortex_edit] blocks? --> rewrite own rule files (up to 8, <=8,000 chars each)
   |                              |
   v                              v
answer to user            config versioned & persisted (HF repo), edit logged,
                          recent edits + recent failures injected into the
                          next prompt -- the loop closes.

Editable files: constitution, tool_rules, metacognitive_rules, refusal_rules, playbook. Static identity (name, lab, "not ChatGPT/Claude/...") is not editable. Set "power_mode": false in cortex/config.json to fall back to classic v7 limits (4 sections, 4,000 chars, 1 edit per self-improve turn).

---

Repository map

Path	Purpose
chatpbc1/Pally	Model weights, paper, developer guide, benchmarks
chatpbc1/Pally/DEVELOPER_GUIDE.md	How to serve, call, configure, and extend Pally
chatpbc1/Pally/RESEARCH_CORTX1.md	Research paper: CORTX-1 architecture + self-correction
This repo (chatpbc1/pally-endpoint)	CORTX runtime config (cortex/config.json), metrics, docs
chatpbc1/pally-console	Chat UI Space

Built by ChatPBC Labs. Apache-2.0.