Scrub recipe/schema from card; keep results + honest framing (v0)

README.md

@@ -9,63 +9,60 @@ tags:
   - behavioral-attestation
   - payments
   - structured-output
-  - qlora
-  - lora
-  - peft
 ---
 
 # ZeroProof Intent Model 0 — `v0`
 
-> **Status: v0 — first iteration, private preview.** A LoRA adapter over `gemma-3-1b`
-> that reads a payment/agent conversation and emits **one structured intent object**.
-> This is the model half of ZeroProof's behavioral-attestation stack: it decides *what
-> the user actually intended* so a downstream layer can verify an agent's action matches it.
+> **Status: v0 — first iteration, private preview.** A compact adapter over `gemma-3-1b`
+> that reads a payment/agent conversation and returns **one structured intent object**.
+> It is the model half of ZeroProof's behavioral-attestation stack: it decides *what the
+> user actually intended* so a downstream layer can verify an agent's action matches it.
 
 ![Model 0 evaluation](eval_report.png)
 
-Built by **[ZeroProof](https://www.zeroproofai.com/)** — an LLM fine-tuning lab building the
-trust layer for AI agents. This is our own data, our own training pipeline, and our own
-rubric; that's the part that compounds.
+Built by **[ZeroProof](https://www.zeroproofai.com/)** — a lab building the trust layer for
+AI agents. The data, the training, and the rubric are ours; that's the part that compounds.
 
 ---
 
 ## What it does
 
-**In:** a short user↔assistant conversation. **Out:** one JSON object — `intent_detected`,
-`core_type`, a per-intent `details` object, `confidence`, `reason`, `source_message_seqs`.
-Ground truth comes from the **user's** turns only; assistant turns are context.
+**In:** a short user↔assistant conversation. **Out:** a single structured intent object —
+the detected intent, whether it's actionable, the salient details for that intent, and a
+confidence signal. Ground truth comes from the **user's** turns; assistant turns are context.
 
-**7 intents (`core_type`):** `spend` (buy from a merchant) · `send` (pay a person/wallet) ·
-`exchange` (swap/convert an asset) · `recur` (subscription/renewal) · `bill` (pay a
-specific bill) · `reverse` (refund/return/dispute) · `none` (no actionable intent — the
-safety class; a false positive here is a *false attestation*, our worst failure mode).
+It routes to one of **7 intents:** `spend` (buy from a merchant) · `send` (pay a
+person/wallet) · `exchange` (swap/convert an asset) · `recur` (subscription/renewal) ·
+`bill` (pay a specific bill) · `reverse` (refund/return/dispute) · `none` (no actionable
+intent — the safety class; a false positive here is a *false attestation*, the worst
+failure mode). This is **routing-then-extraction**, not one-word classification.
 
-This is **routing-then-extraction into a schema**, not one-word classification.
+*(The exact output schema and integration spec are part of the ZeroProof runtime and are
+not published here — reach out for access.)*
 
 ---
 
-## How well did the fine-tuning actually work?
+## How well did the fine-tuning work?
 
 Base and Model 0 scored in **one engine, on the identical 4,000 held-out conversations**
-(greedy decoding, zero train/eval leakage). Only the adapter differs.
+(greedy decoding, de-duplicated so nothing leaks from training). Only the adapter differs.
 
 | Metric | Base `gemma-3-1b` | **Model 0** | Δ |
 |---|---|---|---|
-| Valid JSON | 87.5% | **99.8%** | +12.3 |
+| Valid structured output | 87.5% | **99.8%** | +12.3 |
 | Intent detected (act vs. `none`) | 35.0% | **81.9%** | +46.9 |
-| **Core-type routing — macro** | **14.4%** | **70.0%** | **+55.6** |
-| Details score | 19.7% | **61.3%** | +41.6 |
+| **Intent routing — macro** | **14.4%** | **70.0%** | **+55.6** |
+| Detail extraction | 19.7% | **61.3%** | +41.6 |
 
 *Macro = equal weight per intent (the honest headline; the eval is 57% `none`, so a
-none-weighted micro number would flatter us — we don't report it as the headline).*
+none-weighted average would flatter us — we don't report it as the headline).*
 
-## Why this beats a base 1B — and why it's not luck
+## Why this beats a base 1B — and why it isn't luck
 
 The base model **doesn't fail randomly. It collapses to one class.** Across all 4,000 eval
-conversations, base predicted `spend` **80.7%** of the time — a `reverse` request, a
-status question, a "yo what's up" — all labeled `spend`. It can emit valid JSON (87.5%)
-but it **cannot route**: given the schema it latches onto the first/most-salient intent and
-marks nearly everything actionable.
+conversations it predicted `spend` **80.7%** of the time — a refund request, a status
+question, a "yo what's up," all called `spend`. It can produce well-formed output but it
+**cannot route**: it latches onto one intent and marks nearly everything actionable.
 
 Fine-tuning teaches the two things the base lacks — **routing across all seven intents**
 and **the `none` boundary**:
@@ -80,53 +77,21 @@ and **the `none` boundary**:
 | spend | 300 | 84 | 63 |
 | recur | 300 | 0 | **59** |
 
-(`spend` is the *only* column where the base looks good — because it guesses `spend` for
-everything. Model 0 trades that false strength for real coverage of all seven.)
+`spend` is the *only* column where the base looks good — because it guesses `spend` for
+everything. Model 0 trades that false strength for real coverage of all seven.
 
 ## Why a fine-tuned 1B instead of calling the frontier model?
 
-This model runs in the **hot path of every agent transaction**. Against the frontier parent
-(Opus-class), the argument is **economic and operational**, not "we're more accurate":
+This model runs in the **hot path of every agent transaction**. Against a frontier
+parent, the case is **economic and operational**, not "we're more accurate":
 
-- **Cost / latency / energy** — a self-hosted 1B is orders of magnitude cheaper and faster
-  per call than a frontier API. Attestation only works if you can afford to run it on
-  *every* action.
-- **Self-hostable & pinned** — a fixed adapter version you control, no external dependency
-  in the trust path.
-- **Accuracy vs. the parent** — this is the **next milestone (Model 1)**, measured on this
+- **Cost / latency / energy** — a self-hosted 1B is far cheaper and faster per call.
+  Attestation only works if you can afford to run it on *every* action.
+- **Self-hostable & pinned** — a fixed version you control, no external dependency in the
+  trust path.
+- **Accuracy vs. the parent** — that's the **next milestone (Model 1)**, measured on this
   exact eval. We do **not** claim to beat the parent on accuracy in v0; we claim to close
-  enough of the gap to be worth running everywhere. The honest comparison bars go here when
-  we run them.
-
----
-
-## Usage
-
-```python
-from peft import PeftModel
-from transformers import AutoModelForCausalLM, AutoTokenizer
-
-base = "unsloth/gemma-3-1b-it-bnb-4bit"
-tok = AutoTokenizer.from_pretrained(base)
-model = AutoModelForCausalLM.from_pretrained(base, device_map="auto")
-model = PeftModel.from_pretrained(model, "zero-proof-ai/intent-model-0")  # this adapter
-
-# prompt = system (intent schema) + the conversation; model returns the intent JSON
-```
-
-Served in production as an OpenAI-compatible vLLM endpoint (base + adapter loaded together).
-
-## Training
-
-- **Method:** QLoRA SFT (rank-16 adapter, response-only supervision) on a frozen,
-  pinned `gemma-3-1b` snapshot. 2 epochs, effective batch 16. Trained on Modal.
-- **Data:** 6,000 train / 4,000 eval, carved from a 13k label-blind distillation corpus.
-  Conversations are generated by a **two-agent role-play** where the models never see the
-  intent taxonomy, the JSON schema, or the word "classification" — so transcripts read
-  human, not schema-shaped. Generation and labeling are **separate passes**; labels come
-  from a tiered LLM-judge protocol (no regex/keyword filters). Near-duplicates removed
-  (MinHash-LSH) so eval gains aren't inflated. *Training data is proprietary; a hard,
-  balanced eval set will be released as an open benchmark.*
+  enough of the gap to be worth running everywhere.
 
 ---
 
@@ -134,25 +99,24 @@ Served in production as an OpenAI-compatible vLLM endpoint (base + adapter loade
 
 *(We publish these because the thinking is the product.)*
 
-- **Noticed:** the base's failure is *structured* (collapse-to-`spend`), which is exactly
-  why routing supervision moves the needle so much.
-- **Rubric weaknesses we own:** `details_score` is a strict exact-match and under-credits
-  near-right extractions; `confidence` is currently a teacher artifact (spikes on `0.95`),
-  not a calibrated signal — treat it as such.
+- **Noticed:** the base's failure is *structured* (collapse-to-`spend`) — which is exactly
+  why teaching it to route moves the needle so much.
+- **Rubric weaknesses we own:** detail scoring is strict exact-match and under-credits
+  near-right extractions; the confidence field isn't calibrated yet — treat it as a rough
+  signal, not a probability.
 - **What we're watching next (attestation metrics, not accuracy bars):** how often the
   model catches a **human switching intent mid-conversation**, and how often it correctly
-  flags **"this needs verification"** before an action fires. These matter more for a trust
-  layer than raw accuracy, and they're built on our own data + rubric.
+  flags **"this needs verification"** before an action fires. Those matter more for a trust
+  layer than raw accuracy.
 
 ## Limitations
 
-v0, single base family, **synthetic** training data (inherits teacher blind spots).
-Weakest intents are `recur` (59) and `spend` (63) — the status-vs-action and
-purchase-vs-defer boundaries — which the next data wave targets. No frontier/peer or
-long-context comparison is claimed yet (Model 1). Labels lean on LLM judgment.
+v0, single base family. Weakest intents are `recur` (59) and `spend` (63) — the
+status-vs-action and purchase-vs-defer boundaries — which the next iteration targets. No
+frontier/peer or long-context comparison is claimed yet (Model 1).
 
 ## Links
 
-- Whitepaper (draft): *Label-Blind Distillation and Honest Evaluation for 1B-Scale
-  Payment-Intent Detection*
-- Org: [Zero Proof AI](https://huggingface.co/zero-proof-ai) · [zeroproofai.com](https://www.zeroproofai.com/)
+- **ZeroProof** — [zeroproofai.com](https://www.zeroproofai.com/) ·
+  [huggingface.co/zero-proof-ai](https://huggingface.co/zero-proof-ai)
+- Technical report (v0): *Honest Evaluation for 1B-Scale Payment-Intent Detection* — coming soon.