reports/PROVENANCE.md

# Kuiper-R1 — Provenance Report

Kuiper-R1 is a reasoning-trace **inversion** model: it expands compressed reasoning bubbles into full, explicitly-labeled **synthetic** traces. It does not reproduce any proprietary model's hidden reasoning, and every output is wrapped in `<synthetic_trace>` — declared synthetic by construction.

## Asset audit

| Asset | Role | License | Verdict |
|---|---|---|---|
| `empero-ai/Qwythos-9B-Claude-Mythos-5-1M` | BASE MODEL (used, per owner instruction) | apache-2.0 (weights only) | QUESTIONABLE |
| `Qwen/Qwen3.5-9B` | CLEAN CONTROL BASE (available; recommended for a provenance-clean re-base) | apache-2.0 | CLEAN |
| `open-thoughts/OpenThoughts3-1.2M` | TRAINING CORPUS | apache-2.0 | CLEAN |
| `Jackrong/DeepSeek-V4-Distill-8000x` | HELD-OUT VALIDATION BENCHMARK | mit | CLEAN |
| `lordx64/fable-sft-combined-v2` | OPTIONAL STRESS SET — REJECTED, NOT USED | agpl-3.0 | BLOCKED |
| `Jackrong/Trace-Inverter-4B` | METHODOLOGY REFERENCE (not used as data/weights) | apache-2.0 | CLEAN as a reference; its authors' separate application to Claude data is not replicated here |

### Detail

**`empero-ai/Qwythos-9B-Claude-Mythos-5-1M`**  
- Role: BASE MODEL (used, per owner instruction)  
- License: apache-2.0 (weights only)  
- Teacher/lineage: self-declared: post-trained on ~500M tokens of Claude Mythos/Fable traces with UNDISCLOSED collection method; base weights from Qwen/Qwen3.5-9B  
- Verdict: QUESTIONABLE — FAILS clean-provenance gate  
- Note: Apache-2.0 covers the Qwen-derived weights; it does not grant rights to the Claude-derived training content, whose sourcing the card does not disclose. Kuiper inherits this taint at the weights level. Recorded here explicitly; the release decision is deferred to post-eval review (owner directive).  

**`Qwen/Qwen3.5-9B`**  
- Role: CLEAN CONTROL BASE (available; recommended for a provenance-clean re-base)  
- License: apache-2.0  
- Teacher/lineage: Qwen foundation model (open)  
- Verdict: CLEAN  

**`open-thoughts/OpenThoughts3-1.2M`**  
- Role: TRAINING CORPUS  
- License: apache-2.0  
- Teacher/lineage: reasoning traces annotated by QwQ-32B (open weights)  
- Verdict: CLEAN — no proprietary-frontier teacher; permissive license; lineage disjoint from the benchmark  

**`Jackrong/DeepSeek-V4-Distill-8000x`**  
- Role: HELD-OUT VALIDATION BENCHMARK  
- License: mit  
- Teacher/lineage: DeepSeek-V4-Flash (DeepSeek license permits distillation); prompts from GLM-5.1-Reasoning-1M-Cleaned  
- Verdict: CLEAN  

**`lordx64/fable-sft-combined-v2`**  
- Role: OPTIONAL STRESS SET — REJECTED, NOT USED  
- License: agpl-3.0  
- Teacher/lineage: harvested from 481 Claude Fable 5 sessions; reasoning reconstructed post-hoc  
- Verdict: BLOCKED — AGPL copyleft + Anthropic-output provenance; excluded from the pipeline  

**`Jackrong/Trace-Inverter-4B`**  
- Role: METHODOLOGY REFERENCE (not used as data/weights)  
- License: apache-2.0  
- Teacher/lineage: n/a (technique only)  
- Verdict: CLEAN as a reference; its authors' separate application to Claude data is not replicated here  

## Training-data cleanliness guarantee

- Training traces come solely from `OpenThoughts3-1.2M` (Apache-2.0, QwQ-32B teacher).
- Compression to bubbles is **deterministic and extractive** — no model call, no proprietary teacher touches the data, fully reproducible.
- A source-row provenance filter rejects any trace containing frontier-model self-identification markers (defense in depth).
- The benchmark lineage (GLM-5.1 prompts / DeepSeek-V4-Flash) is **disjoint** from the training lineage, so eval measures generalization, not memorization.
- No Quark-derived data is used (contained Opus/GPT rows); excluded per owner instruction.

## Known limitation (recorded, not resolved)

The base model carries undisclosed Claude-trace provenance and **fails the clean-provenance release gate**. A provenance-clean rebuild is available by swapping the base to `Qwen/Qwen3.5-9B` (identical architecture) — see the model card. Publication is gated on owner review of this report plus the eval report.

## Built dataset (from summary.json)

- Examples: train=3000, val=1000, test=1000, benchmark=3529
- Train domains: {"math": 1001, "science": 1531, "code": 468}
- Avg trace chars (train): 5780 · avg bubble chars: 1168
- Compression config: {"strategy": "extractive", "target_ratio": 0.28, "min_bubbles": 3, "max_bubbles": 24, "bubble_prefix": "- "}