> ## Documentation Index
> Fetch the complete documentation index at: https://doc.blueapi.ir/llms.txt
> Use this file to discover all available pages before exploring further.

# EMEP Evidence Tag Classification Map

> Map every EMEP research area to its evidence tag: Established research, Research hypothesis, Engineering assumption, Experimental result, or Future goal.

EMEP documentation uses evidence tags to distinguish proven claims from predictions, assumptions, and targets. This page maps every research area to its tag and shows the hierarchical evidence tree.

## Evidence Tag Definitions

| Tag                    | Meaning                               | Use in EMEP                                           |
| ---------------------- | ------------------------------------- | ----------------------------------------------------- |
| Established research   | Peer-reviewed, replicated finding     | Cite with authors and year; safe to depend on         |
| Research hypothesis    | Predicted but not yet validated       | Mark clearly; design experiments to test              |
| Engineering assumption | Pragmatic choice without formal proof | Document rationale; revisit if contradicted           |
| Experimental result    | Finding from an EMEP experiment       | Tag with EXP-TBD-### ID; provisional until replicated |
| Future goal            | Target state, not yet implemented     | Use for roadmap and planning only                     |

## Research Area to Evidence Tag Mapping

| Research Area                                  | Evidence Tag           | Justification                                                                                              |
| ---------------------------------------------- | ---------------------- | ---------------------------------------------------------------------------------------------------------- |
| Model Soups (averaging weights)                | Established research   | Wortsman et al. 2022; replicated in multiple follow-ups                                                    |
| Task Arithmetic (task vectors)                 | Established research   | Ilharco et al. 2022; widely adopted in merging literature                                                  |
| TIES-Merging                                   | Established research   | Yadav et al. 2023; validated on multi-task benchmarks                                                      |
| DARE                                           | Established research   | Yu et al. 2023; independent replication confirmed                                                          |
| SLERP for model weights                        | Research hypothesis    | Shoemake 1985 is established for rotations; application to high-dimensional weight vectors is hypothesized |
| Evolutionary Model Merge                       | Established research   | Akiba et al. 2024; demonstrated on diverse model families                                                  |
| LoRA for task vector production                | Established research   | Hu et al. 2021; standard practice in PEFT                                                                  |
| QLoRA for resource-constrained fine-tuning     | Established research   | Dettmers et al. 2023; replicated at scale                                                                  |
| NSGA-II for merge optimization                 | Engineering assumption | Deb et al. 2002 is established; applicability to model merging search space is assumed                     |
| NSGA-III for many-objective merging            | Research hypothesis    | Deb & Jain 2014 is established; benefit for 5+ merge objectives is predicted                               |
| CMA-ES for merge hyperparameters               | Research hypothesis    | Hansen 2001 is established; continuous merge ratio optimization is untested in EMEP                        |
| Catastrophic forgetting in merged LLMs         | Research hypothesis    | McCloskey & Cohen 1989 is established for small networks; LLM-scale merge forgetting is under study        |
| EWC for merge-time protection                  | Research hypothesis    | Kirkpatrick et al. 2017 is established for sequential learning; merge application is speculative           |
| Benchmark contamination detection              | Established research   | Sainz et al. 2023; methods are validated but not exhaustive                                                |
| Hidden Test Set isolation prevents overfitting | Engineering assumption | Logical consequence of ML best practice; not formally proven for evolutionary merging                      |
| ed25519 signing for model provenance           | Engineering assumption | Cryptographic primitive is established; supply chain application in ML is emerging                         |
| Safetensors security advantage over pickle     | Established research   | Format specification and security analysis are published                                                   |
| Offline-first enterprise mode                  | Future goal            | Design target for Phase 7; not yet implemented or validated                                                |
| Pareto front diversity guarantees              | Research hypothesis    | NSGA-II/III provide diversity mechanisms; guarantee for model merge space is unproven                      |
| Semantic versioning for models                 | Engineering assumption | Standard practice in software; formalized for ML artifacts in EMEP                                         |

## Evidence Tree

```mermaid theme={null}
mindmap
  root((Evidence Tree))
    Established
      Model Soups
      Task Arithmetic
      TIES
      DARE
      LoRA
      QLoRA
      Benchmark Contamination
      Safetensors Security
    Hypothesis
      SLERP for Weights
      NSGA-III for Merging
      CMA-ES for Ratios
      Forgetting in LLMs
      EWC in Merges
      Pareto Diversity
    Assumption
      NSGA-II for Merges
      Test Set Isolation
      ed25519 Provenance
      Semantic Versioning
    Result
      EXP-TBD-001..018
    Goal
      Offline-First Mode
      Full Reproducibility
```

## Tagging Rules

1. Every claim in specification pages must carry exactly one tag
2. Untagged claims are flagged in documentation review
3. Hypotheses must link to an experiment ID that will test them
4. Assumptions must name the condition under which they would be revisited
5. Results must include the EXP-TBD-### identifier and replication status

## Next Steps

See [Component Mapping](/research/component-mapping) for how these research areas justify specific EMEP components, or [References](/research/references) for the full paper annotations.