File size: 2,498 Bytes
01717aa
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
---
pretty_name: Cross-Form Learning Consequence
tags:
- machine-learning
- online-learning
- plasticity
- reproducibility
- synthetic-data
---

# Cross-Form Learning Consequence

A small controlled experiment asking whether a useful consequence of learning can survive after the **task-specific predictive representation that produced it is destroyed**.

The task family is deliberately transparent: 16-dimensional synthetic binary classification with four active coordinates and manual NumPy logistic SGD.

## Experiment

1. Learn a task-specific predictor `R`.
2. Derive a lower-bandwidth four-coordinate structural outline `S` from `|R|`.
3. Destroy `R` and reset the future predictor to zero.
4. Use `S` only to bias the **directional allocation** of future gradient updates.
5. Compare against neutral, shuffled-history, wrong-family, random-outline, oracle, no-history, and direct-preservation controls.

Structured updates are rescaled every step so their L2 norm matches the ordinary gradient norm. The intervention therefore changes allocation/direction rather than step magnitude.

## Result

Across **1,024 paired episodes**:

- the destroyed predictor returned to **0.500000** accuracy on every episode;
- the structural outline recovered the hidden support exactly on **98.83%** of episodes;
- transformed-minus-neutral normalized accuracy-AUC was **+0.080524**;
- oracle-minus-neutral was **+0.081000**;
- the fixed transform recovered **99.42%** of oracle structural headroom across seed means.

This is an **instrument-level existence result**, not a learned memory or architecture result. The transformation was fixed by the experimenter.

## Run

```bash
python code/run_experiment.py --self-test
```

Full exposed cohort:

```bash
python code/run_experiment.py --output-root ./runs
```

## Repository contents

- `code/run_experiment.py` — public-release copy of the frozen NumPy implementation
- `protocol/FROZEN_PROTOCOL_PUBLIC_RELEASE.md` — complete protocol with public naming
- `results/aggregate_summary.json`
- `results/per_seed_summary.csv`
- `results/episode_summary.csv`
- `results/assertion_log.txt`
- `results/development_result.md`
- `notebooks/RUN_IN_COLAB.ipynb`
- `docs/SOURCE_CUSTODY.md`

## Scope

The result applies to this deliberately constructed synthetic instrument. It does not establish general continual-learning performance, learned persistence-form selection, architecture independence, or superiority to standard transfer/meta-learning methods.