# SOURCE_MAP — 086

Accessed 2026-09-29. Exact revisions verified by public downloads, existing official repository snapshot and normalized source comparison. All local orchestration/metrics/contrasts/plots are independent teaching implementation, not original paper code. We execute official Sentence Transformers CrossEncoder and unchanged BEIR Rerank; we do not reproduce original BERT-LARGE training.

| Module / formula | Official fixed source | Local use | License / modifications |
|---|---|---|---|
| Candidate pair creation and score-to-ID association | [BEIR Rerank.rerank L15–49](https://github.com/beir-cellar/beir/blob/ef83d29307061c65d04b035b4f4e7c18bd8374af/beir/reranking/rerank.py#L15-L49) | `official_rerank.py`, called in `run.py` | Apache-2.0, exact bytes copied, license preserved in licenses/BEIR-LICENSE. No scoring changes. Importing this standalone module avoids loading unrelated BEIR dense APIs. |
| Title-space-abstract | [same L24–35](https://github.com/beir-cellar/beir/blob/ef83d29307061c65d04b035b4f4e7c18bd8374af/beir/reranking/rerank.py#L24-L35) | `run.py` altered corpus and actual fast-tokenizer offsets | Apache-2.0, official execution; abstract-only intervention sets title to empty; official source remains unmodified |
| Joint tokenization, longest_first, max_length | [ST CrossEncoder.smart_batching_collate_text_only L164–179](https://github.com/UKPLab/sentence-transformers/blob/7d52a069e0b37d976b3ed3f674a6180436c27574/sentence_transformers/cross_encoder/CrossEncoder.py#L164-L179) | installed sentence-transformers3.4.1, `run.py` | Apache-2.0; entire installed source equals official after newline normalization; no source edits |
| Inference and scalar score | [ST CrossEncoder.predict L366–462](https://github.com/UKPLab/sentence-transformers/blob/7d52a069e0b37d976b3ed3f674a6180436c27574/sentence_transformers/cross_encoder/CrossEncoder.py#L366-L462) | `run.py`, official `predict` inside BEIR | Apache-2.0; explicit Identity matches model config; eval/no_grad official implementation |
| Joint encoding score and paired input | [model card](https://huggingface.co/cross-encoder/ms-marco-MiniLM-L6-v2/blob/233902d25c440f23af6f7d6e94d2946bac0bee0a/README.md#L17-L56), [config](https://huggingface.co/cross-encoder/ms-marco-MiniLM-L6-v2/blob/233902d25c440f23af6f7d6e94d2946bac0bee0a/config.json) | frozen weights/tokenizer via `model.lock.json`, `prepare.py` | model card Apache-2.0; model files not distributed.6layers, hidden384,1label, max512. config _name_or_path refers toL12 ancestor; actual num_hidden_layers=6. |
| nDCG / paired contrasts | [BEIR metric evaluation L68–115](https://github.com/beir-cellar/beir/blob/ef83d29307061c65d04b035b4f4e7c18bd8374af/beir/retrieval/evaluation.py#L68-L115) | pytrec_eval primary in `run.py`; independent `common.metric` and `audit.py` | metric/contrast code independent; no formula claimed from original paper |
| Fixed candidate source | [084 runs](https://github.com/distance539/ai-research-lab/blob/804da74ba3484e0156183951e89228d8e71c780b/cycle06/084-rrf-rank-changes/results/runs.json.gz) | `inputs/candidates.json` | independent project code Apache-2.0; exact hash selection/subsetting in input.lock.json; full retrieval not rerun |

SciFact original repository commit68b98a56d93e0f9da0d2aab4e6c3294699a0f72e. BEIR dataset archive lacks a Git revision; actual public URL and SHA-256 in resources.lock.json are its version identity. Original license distinguishes claim annotations CC BY4.0 and abstracts ODC-By1.0; full corpus remains in cache, not this package. `licenses/SCIFACT-LICENSE.md` retains provenance and text. qrels subset is attributed in NOTICE.md. No manual annotation was performed.

Official modules were actually downloaded and read; web browser cache misses did not replace this verification. Raw SHA-256 evidence is in upstream_verification.json. Full package requirements freeze runtime; environment.json records all installed versions. Model card numbers and V100 throughput are not our results.

## 086 additions

`run.py` independently constructs2×2conditions, counts actually retained query/title/abstract tokens using fast tokenizer `sequence_ids` and `offset_mapping`, and saves SHA256 of actual input IDs. Offsets crossing the title/abstract boundary count as abstract. This character-offset audit is an independent diagnostic, not part of original BEIR scoring. `audit.py` checks saved metrics and token accounting; `plot.py` plots persisted summary only. The difference-in-differences equation is arithmetic of paired nDCG values, independently implemented, no upstream formula attribution.

Local common.py/prepare.py/official_rerank.py and model/data/dependency/license locks reused from[085 fixed source](https://github.com/distance539/ai-research-lab/tree/4a2c80e8b39305f1ac411bcd80c0c688d0660c5e/cycle06/085-reranker-recall-ceiling). All required self-written modules are included, no implicit local dependency. Inputs/candidates.json and qrels.json are exact085 bytes (input.lock.json).085 traces back to084 RRF and082/083 original retrieval.086 does not rerun retrieval or alter candidate generation.

The shared installed ST3.4.1 source is compared to official source after newline normalization. Raw source downloads and HTTP/hash evidence are preserved in upstream_verification.json. Full source text was read at this run, not inferred from a search snippet. Model/tokenizer exact revision and every consumed file hash are fixed in model.lock.json; data lacks a Git revision and uses archive/file SHA256 from resources.lock.json. The original SciFact commit is historical provenance, not a fabricated revision of BEIR export.