docs/CLAIMS.md

# CausalAtlas Fact-Locked Claim Map

This file maps every public claim to evidence, code, scope, and a falsifiable
limitation. The canonical narrative is
[`results/move1/MOVE1_REPORT.md`](../results/move1/MOVE1_REPORT.md).
Evaluation-code paths refer to the source
[GitHub repository](https://github.com/jang1563/causalatlas); this dataset is an
artifact mirror and does not duplicate the `eval/` tree.

## Claim summary

| ID | Supported claim | Primary evidence |
|---|---|---|
| C1 | GEARS loses to observed additive in aggregate on held-out Norman combination edges but wins on the held-out non-additive subset. | `results/move1/RESULT_layerA.md`, `results/move1/layerA_norman.json` |
| C2 | Arc STATE/Tahoe is below a no-change floor on edge-call accuracy while its cosine-PDS is 0.793, so the assessment is metric-dependent. | `results/move1/RESULT_layerA.md`, `results/move1/layerA_tahoe.json` |
| C3 | Trust/verify utility is cost-conditioned in the tested benchmark; model names are experimental conditions, not a general capability scale. | `results/move1/MOVE1_REPORT.md`, linked Verify-or-Trust release |
| C4 | At λ = 0.5, always-additive beats three tested Claude policies in one neutral three-tier prompt setting; at λ = 0.2, the model policies win. | `results/move1/MOVE1_3TIER_RESULT.md` |
| C5 | Uncertainty-based selection loses to random for uniform map reconstruction, while GO-predicted effect-magnitude selection improves recall for the tested hit-discovery objective. | `results/move1/EXPERIMENT_SELECTION_RESULT.md` |

## C1. Norman model-versus-baseline regime map

Five-fold leave-combination-out results:

- all held-out combinations: GEARS 0.790 AUROC, additive 0.897, gap -0.107
  [95% CI -0.132, -0.083];
- additive-residual subset: GEARS 0.755, additive 0.945, gap -0.190
  [-0.217, -0.162];
- non-additive subset: GEARS 0.749, additive 0.606, gap +0.142
  [+0.050, +0.233].

Evidence:

- `results/move1/RESULT_layerA.md`
- `results/move1/layerA_norman.json` for the original split and threshold sweep
- `eval/move1_gears_cv.py`
- `eval/move1_synergy_cv.py`

Scope and limitations:

- one GEARS configuration;
- Norman K562 CRISPRa combinations;
- observed additive requires measured single perturbations;
- edge discrimination does not measure every biological-use objective;
- the released JSON does not contain the large CV prediction table.

## C2. Tahoe / Arc STATE metric split

Released artifact values:

- 1,135 perturbations and 2,274,000 edges;
- STATE call accuracy 0.732 [0.728, 0.736];
- no-change call accuracy 0.805 [0.800, 0.809];
- STATE edge AUROC 0.439;
- STATE cosine-PDS 0.793.

Evidence:

- `results/move1/RESULT_layerA.md`
- `results/move1/layerA_tahoe.json`
- `eval/move1_layerA_tahoe.py`

Scope and limitations:

- upstream Arc model/data terms apply;
- call accuracy and cosine-PDS answer different questions;
- drug novelty depends on upstream training-set documentation;
- the result does not transfer automatically to other cell states or drugs.

## C3. Cost-conditioned trust and verification

The linked Verify-or-Trust benchmark scores correctness minus assay cost. It
reports that verification allocation and net utility depend on the chosen cost,
prompt information, model version, and reliability signal supplied in the
benchmark.

Evidence:

- `results/move1/MOVE1_REPORT.md`
- [Verify-or-Trust repository](https://github.com/jang1563/verify-or-trust)
- [Verify-or-Trust dataset](https://huggingface.co/datasets/jang1563/verify-or-trust)

Scope and limitations:

- controlled benchmark rather than general LLM tool use;
- model versions are not a controlled capability axis;
- no universal model-size or capability-ordering claim is made here.

## C4. Three-tier default-to-baseline result

At λ = 0.5:

- always-additive net 0.802;
- trust-all-GEARS net 0.661;
- Haiku 4.5 net 0.725, assay rate 32%;
- Sonnet 4.6 net 0.722, assay rate 37%;
- Opus 4.8 net 0.730, assay rate 40%.

At λ = 0.2, the net scores for Haiku, Sonnet, and Opus are 0.821, 0.833, and
0.851, respectively, all above always-additive. This establishes a
cost-conditioned policy comparison, not a universal baseline rule.

Evidence:

- `results/move1/MOVE1_3TIER_RESULT.md`
- `eval/move1_3tier_baselines.py`
- `eval/move1_3tier_router.py`
- `eval/move1_3tier_env.py`

Scope and limitations:

- 52 held-out combination panels;
- one neutral prompt family;
- three Claude model versions;
- Opus additive-follow is based on four non-assayed disagreements;
- no independent prompt replication.

## C5. Goal-dependent experiment selection

In the offline Replogle replay:

- uniform-map uncertainty-minus-random AULC is -0.018, 95% CI
  [-0.030, -0.005];
- hit-discovery recall at budget is 0.19 for GO-predicted value selection and
  0.07 for random, approximately 2.7 times as high.

Evidence:

- `results/move1/EXPERIMENT_SELECTION_RESULT.md`
- `eval/move1_replogle_prep.py`
- `eval/move1_expsel.py`
- `eval/move1_expsel_checks.py`

Scope and limitations:

- offline, corpus-bound replay;
- Replogle K562 essential screen;
- kNN/ridge world-model family;
- no LLM or RL selector result in this release.

## Retired claims

The additive-disagreement competence classifier, the associated verification
replay, broad model-ordering language, and general conclusions about RL are not
public claims. Their provenance and retirement reasons are documented in
[`results/move1/RETIRED_ANALYSES.md`](../results/move1/RETIRED_ANALYSES.md).

## Machine-readable companion

[`artifact_manifest.json`](../artifact_manifest.json) contains the same five
supported claims and the retired-analysis registry. Full source validation is
documented in [`REPRODUCIBILITY.md`](REPRODUCIBILITY.md).