# Supplementary data: experimental outputs

This archive contains the raw outputs of the case study reported in the manuscript "Archives as wellspring, verifier and generator of knowledge: a vision for the age of AI". The case study compares two ways of combining the same archival materials with a language model:

- **Condition A (trained language model):** the base model was further trained on the records with LoRA adapters and answered from memory alone, without access to the records.
- **Condition B (knowledge-generating archive):** the base model, without further training, answered by using the archive's database as a set of tools (metadata lookup, a relational graph, vector search and reading the text of individual records).

Both conditions received the same two requests:

- **T1:** list the people who sent letters to Sim Neunggyu, with the number of letters, in order from those who sent the most.
- **T2:** write a narrative of about 5,000 characters reconstructing his life on the basis of the letters he received.

Web search was disabled in all runs.

## Decoding

Each configuration was run six times:

- one greedy run (temperature 0, top-p 1), and
- five sampled runs (temperature 0.7, top-p 0.95; top-k disabled).

## Directory structure

All runs are in `runs/`. Each directory name begins with the start time of the run (UTC).

### Condition A

| Directory | Base model | Contents |
|---|---|---|
| `20261016T083350Z_gemma-4-31B-it_A-cite` | Gemma 4 31B-it | `base/` (before training) and `checkpoint-480/`, `checkpoint-640/`, `checkpoint-720/`, `checkpoint-800/` (after the given number of training steps) |
| `20261016T090959Z_Qwen3.8-27B_A-cite` | Qwen3.8-27B | `base/` (before training) and `checkpoint-560/` |

Within each subdirectory, files are named by request and run, for example `t1-greedy` or `t2-s3` (the third sampled run). Each response is provided in two forms:

- `.md`: the response text, including the grounds tags the model was asked to attach to each sentence;
- `.json`: the same response with run metadata (base model, decoding settings, number of generated tokens, hashes of the prompts used).

### Condition B

| Directory | Base model | Request | Notes |
|---|---|---|---|
| `20261016T042652Z_gemma-4-31b-it_b` | Gemma 4 31B-it | T1 | |
| `20261016T044101Z_gemma-4-31b-it_b` | Gemma 4 31B-it | T2 | original request |
| `20261016T051130Z_gemma-4-31b-it_b_t2` | Gemma 4 31B-it | T2 | revised request (chronological scenes) |
| `20261019T064312Z_qwen3.8-27b_b_t1` | Qwen3.8-27B | T1 | |
| `20261019T064810Z_qwen3.8-27b_b_t2` | Qwen3.8-27B | T2 | revised request (chronological scenes) |

Each directory contains:

- `config.json`: run settings (model, decoding, output limit, maximum number of refinement steps, hashes of the prompts used);
- `units/*.json`: one file per run (`t1-greedy`, `t1-sample-1` … `t1-sample-5`, and likewise for T2);
- `provenance.jsonld`: a provenance record of the run in JSON-LD (W3C PROV).

In each unit file:

- `sentences` holds the response, sentence by sentence. Each sentence has `text`, a grounds `label` (for example `holding` or `inference`) and the `record_ids` cited for it;
- `trace` and `scratchpad` record the tool calls and the tool results returned to the model, including the output of the counting tool (for T1, the per-sender counts and the number of letters with unknown senders).

The character counts reported in the manuscript are based on `sentences[].text` only; labels and record identifiers are not counted.

## Notes

- **Record identifiers** (for example `A004_01_A00267_001`) refer to records in the archive from which the materials were drawn. The source materials themselves (metadata, summaries and transcribed texts) are not included in this archive and are available from the corresponding author on reasonable request, as stated in the manuscript.
- **Server addresses.** The `endpoint` field in `config.json` refers to temporary inference servers used for the runs, which are no longer active.
- **Code version.** The Qwen runs of Condition B were executed from a working tree that differed from the git revision recorded in `config.json` (`887f976`). The differences concerned the handling of the model's output, including disabling Qwen's thinking mode, recovering truncated citations, re-planning after empty results or errors, and one correction step when the output was not a sentence array. The model and decoding settings were as recorded.
- **Preliminary run.** If present, the directory `20261016T040140Z_gemma4-e4b_b` is a preliminary run with a different, smaller model. It is not reported in the manuscript.
- **Failed partial runs** were removed; only the complete runs reported in the manuscript are retained.
