# Reific public-data reference replays

Verified 15 September 2026. These files support the private website preview.
They contain attributed public research data and deterministic computations,
not customer data or an agent/model evaluation.

## Sources and attribution

- Candanedo, L. (2017). *Appliances Energy Prediction* [Dataset]. UCI Machine Learning Repository. https://doi.org/10.24432/C5VC8G
  - Source description: https://archive.ics.uci.edu/dataset/374/appliances+energy+prediction
  - Original download: https://archive.ics.uci.edu/static/public/374/appliances%2Benergy%2Bprediction.zip
  - Original file: `source/energydata_complete.csv`; retrieved 2026-09-15.
- Candanedo, L. (2016). *Occupancy Detection* [Dataset]. UCI Machine Learning Repository. https://doi.org/10.24432/C5X01N
  - Source description: https://archive.ics.uci.edu/dataset/357/occupancy+detection
  - Original download: https://archive.ics.uci.edu/static/public/357/occupancy%2Bdetection.zip
  - Original file: `source/datatraining.txt`; retrieved 2026-09-15.

UCI lists both datasets under Creative Commons Attribution 4.0 International:
https://creativecommons.org/licenses/by/4.0/
The original files are unchanged. Derived CSVs select a period or columns and
normalize serialization to UTF-8, comma delimiters, and LF newlines. Evidence
JSON and charts are derived by Reific. No affiliation or endorsement is implied.
The original archive and file SHA-256 digests are in `source/provenance.json`.

## Reproduce

For a standalone replay, download `/evidence/replay-bundle.zip` from the preview
and extract it to a new directory. It contains all source data and artifacts:

```sh
python3 -m venv /tmp/reific-evidence
/tmp/reific-evidence/bin/pip install -r requirements.txt
/tmp/reific-evidence/bin/python replay.py --check
```

This verifies the downloaded source hashes and exact output files without
overwriting them. No private repository or credential is required.

From the website repository root:

```sh
python3 -m venv /tmp/reific-evidence
/tmp/reific-evidence/bin/pip install -r scripts/evidence/requirements.txt
/tmp/reific-evidence/bin/python scripts/evidence/replay.py --check
```

The same script is available as `replay.py` in this folder. To use it separately,
preserve this directory layout including `source/provenance.json`, both source
files, and the generated study folders. Run `python replay.py --check` from an
environment with the pinned requirements. Without `--check`, the script writes
the computed artifacts. The tokenizer may fetch its public encoding table on
first use; there are no language-model or paid API calls.

## Exact comparisons

| Replay | Full CSV | Timestamp + signal CSV | Evidence JSON |
| --- | ---: | ---: | ---: |
| Building demand | 250,068 tokens | 16,838 tokens | 529 tokens |
| Room occupancy | 294,203 tokens | 133,963 tokens | 915 tokens |
| Sum of the two separate payloads | 544,271 tokens | 150,801 tokens | 1,444 tokens |

Tokenization is `tiktoken==0.11.0`, `o200k_base`, using
`len(encoding.encode(text, disallowed_special=()))`. Tokens include the exact
file headers, delimiters, whitespace, and final newline. The report also counts
the same task question prefixed to each payload for a text-envelope comparison.
It does not count a chat/API message envelope, system prompt, or tool schema.

Reduction (%) = (baseline payload tokens - evidence payload tokens) /
baseline payload tokens * 100. The aggregate is the sum of each separately
tokenized file, not an average of percentages or a concatenated single prompt.

The projected CSV is a stronger simple baseline than the full sensor CSV. A
query-enabled agent can also use the same direct SQL/Python computation; these
examples do not establish an advantage over that baseline. Exact serialized
payload counts do not establish total agent token burn, price savings, task
resolution, accuracy, latency, or a customer outcome. Agent calls: 0.

## Questions, scope, and limitations

Building demand selects the first seven complete days in the source, from
2016-01-12 00:00:00 inclusive to 2016-01-19 00:00:00 exclusive. It compares all
1,006 rolling groups of three consecutive ten-minute energy measurements.
The 2,390 Wh maximum is recorded at 2016-01-16 18:40, 18:50, and 19:00. The
source does not specify whether its timestamps label interval starts or ends.
Maximum energy is not a fault diagnosis or identification of a specific device.

Room occupancy uses all 8,143 rows of `datatraining.txt`, removes `Occupancy`
before selection and payload construction, and compares all 8,113 endpoint
changes across thirty sampling intervals. The largest rise is 715 ppm in
1,801 seconds. All 31 separately inspected source occupancy labels are zero.
The label audit is separate and is never passed into a counted payload. This is
a retrospective single-window check, not a model score or causal conclusion.
The published source images behind the occupancy labels were not independently
inspected. Source timestamps in both datasets have no explicit timezone.

The general reference implementation is deliberately simple. It illustrates
inspectable evidence contracts and claim boundaries; it is not the production
Reific model or a claim of learned-model quality.