# Aleph Initial Alpha: stable-signifier projection

## Submission status

This method generated all 81 test predictions at Predict Future Selves
benchmark commit `9b6a766712583fec8d3182957260b1123fbfa146`. The prediction artifact was
frozen on 2026-08-30 with SHA-256:

```text
a463d9e314069357f050c9f2270acfad59165db2c0bab19322d46517444d9ab3
```

The test references are private, so no test score is reported here.

## Method

The method is a deterministic, extractive stable-signifier projection. It uses
the 150 public training pairs and no external training or retrieval data.

1. Tokenize text exactly like the shared evaluator: Unicode word tokens after
   case-folding and whitespace normalization.
2. For every unique source token, estimate the document-level probability that
   it appears in the paired follow-up. Smooth this probability toward the
   corpus-wide token-retention rate with 10 equivalent prior observations.
3. Split each source response using its strongest repeated boundary: nonblank
   lines, then sentences, then comma-separated phrases, or finally the whole
   response if none produces at least three units.
4. Score each unit by the mean retention probability of its unique content
   tokens. Thirteen structural tokens such as “I,” “am,” and “person” are
   omitted from scoring.
5. Fit an ordinary least squares regression of follow-up word count on source
   word count using the training pairs only. For each new case, greedily retain
   the highest-scoring units while adding a unit moves predicted length closer
   to the regression target. Restore retained units to source order.

The model does not use demographic fields. It makes no random choices and has
no random seed. Test predictions were generated once after public-development
model selection and were not manually edited or reviewed case by case.

## Implementation

- Software: Python 3.9.23, standard library only
- Script: `analysis/stable_signifier_projection.py`
- Prior strength: `10.0`
- AI model or provider used for generation: none
- Access and run date: 2026-08-30

Reproduce from a checkout of the pinned benchmark commit:

```bash
python3 analysis/stable_signifier_projection.py \
  --benchmark-dir /path/to/predict-future-selves \
  --dev-output results/stable_signifier_dev_predictions.csv \
  --test-output submissions/aleph_initial_alpha_submission.csv
```

Validate the test artifact with the benchmark's pinned validator:

```bash
python3 scripts/validate_submission.py \
  /path/to/aleph_initial_alpha_submission.csv
```

## Public development scorecard

The 50-case scorecard below was generated by the benchmark's authoritative
evaluator. There is no composite score, weighting, ranking, or claim of an
overall winner. Repeat-2024 values are the benchmark's published baseline.

| Group | Metric | Stable projection | Repeat 2024 | Direction |
| --- | --- | ---: | ---: | --- |
| Agreement | Normalized exact-match rate | 0.000000 | 0.000000 | Higher |
| Agreement | Normalized edit similarity | 0.298061 | 0.291966 | Higher |
| Agreement | Token Jaccard similarity | 0.142768 | 0.141930 | Higher |
| Agreement | Token-overlap F1 | 0.307444 | 0.312202 | Higher |
| Agreement | ROUGE-L F1 | 0.227552 | 0.225683 | Higher |
| Agreement | Character n-gram F1 | 0.292293 | 0.296534 | Higher |
| Form | Word-count MAE | 41.640000 | 56.680000 | Lower |
| Form | Line-count MAE | 9.760000 | 6.800000 | Lower |
| Form | Mean predicted word count | 77.320000 | 108.040000 | Descriptive |
| Form | Mean reference word count | 94.760000 | 94.760000 | Descriptive |
| Change | Prediction repeat-2024 rate | 0.280000 | 1.000000 | Descriptive |
| Change | Observed repeat-2024 rate | 0.000000 | 0.000000 | Descriptive |
| Change | Source-similarity MAE | 0.678215 | 0.774317 | Lower |
| Change | Mean prediction-to-source similarity | 0.903898 | 1.000000 | Descriptive |
| Change | Mean observed follow-up-to-source similarity | 0.225683 | 0.225683 | Descriptive |

Relative to repeating 2024 verbatim, the projection improves normalized edit
similarity, token Jaccard, ROUGE-L, word-count error, and source-similarity
error. It reduces token-overlap and character n-gram F1 and worsens line-count
error. Its mean source similarity remains far above the observed mean
(`0.903898` versus `0.225683`), so the method still predicts too much textual
continuity even after projection.

The development prediction CSV has SHA-256
`0b43ad574543cb128fa51470432467a34998c675b945665bb44141d6e04bca39`.
The machine-readable scorecard has SHA-256
`67b495d069668c9dd9c8007325f8215bd82bd76bfa16738884f7a4aac62f5f05`.

## Licensing

The prediction CSV adapts *You Can Predict Future Selves with AI (or Without
AI)* by Dr. Jason Jeffrey Jones and is licensed CC BY-NC-SA 4.0. Changes are the
extractive predictions described above. This method document and its source
code are offered under the MIT License used for non-data materials in the
benchmark repository.