# E129 — Governed Recursive Self-Improvement (measured)

**Verdict:** POSITIVE_BOUNDED: self-generated library extends sealed tier reach beyond frozen, random and flat controls
**AGI status:** NOT ACHIEVED
**Gates:** 11/11
**Actual elapsed:** 540.507 s
**Sealed answer commitment:** `f48cbd947192d06e3dcc360798848e77fc638b4ae77e5de22593c54e2f1dd8f8`

## What ran

A frozen engine (bounded search over a library of integer programs) was evaluated on 48 sealed
tasks before any self-modification, then four times more.  Between generations the engine used
its **own search** on development tasks and wrote recurring fragments of its own solutions into
its executable library.  The library it ran on generation N+1 contained programs that did not
exist at generation N.

Every sealed answer was withheld until scoring.  Sealed results were never used for improvement
or promotion.  Promotion used development data only.

## Compounding curve (main arm)

| generation | library | max depth | T1 | T2 | T3 | T4 | T6 | T8 |
|---:|---:|---:|---:|---:|---:|---:|---:|---:|
| 0 | 21 | 1 | 1.000 | 0.000 | 0.000 | 0.000 | 0.000 | 0.000 |
| 1 | 35 | 2 | 1.000 | 1.000 | 1.000 | 0.375 | 0.000 | 0.125 |
| 2 | 67 | 4 | 1.000 | 1.000 | 1.000 | 1.000 | 0.125 | 0.250 |
| 3 | 96 | 8 | 1.000 | 1.000 | 1.000 | 1.000 | 0.125 | 0.250 |

Each cell is the fraction of query predictions correct for that nominal tier.  A tier is a chain
of K two-primitive concepts; tiers are nominal because functional equivalences can make a task
reachable one step earlier, which is disclosed.

## Self-written library

- generation 0: 21 entries, max primitive depth 1
- final: 96 entries, max primitive depth 8, 75 macros
- emergent sealed tasks (not solved at generation 0): **27**, tiers [2, 3, 4, 6, 8]

## Controls at the same generation count

- `compute`: 20/48 sealed tasks fully solved; library 21 entries, max depth 1
- `flat`: 28/48 sealed tasks fully solved; library 36 entries, max depth 2
- `frozen`: 8/48 sealed tasks fully solved; library 21 entries, max depth 1
- `random`: 16/48 sealed tasks fully solved; library 53 entries, max depth 2

## Gates

- **G10_journal_chain**: PASS — chain head 26c9f0058e0ff00f...
- **G11_cost_reported**: PASS — 2386246 evaluations at final sealed evaluation
- **G1_self_written_library**: PASS — 75 macros at final generation
- **G2_compounding_depth**: PASS — max primitive depth 1 -> 8
- **G3_emergent_sealed_tasks**: PASS — 27 sealed tasks solved after generation 0, tiers [2, 3, 4, 6, 8]
- **G4_tier_reach_vs_gen0**: PASS — highest tier with >=1 full solve 1 -> 8
- **G5_total_solved_beats_controls**: PASS — main 35; compute 20, flat 28, frozen 8, random 16
- **G6_tier4_beats_flat**: PASS — T4 accuracy main 1.000 vs flat 0.375 (T6 0.125 vs 0.000)
- **G7_depth_beats_flat**: PASS — max depth main 8 vs flat 2
- **G8_no_dev_regression**: PASS — no generation rolled back for development regression
- **G9_protected_targets_refused**: PASS — 2 refusals recorded

## Honest limits

- Constructed integer-rule domain, not professional work and not an external benchmark.
- Macros are data programs, not model weights. No model was trained; no cloud model was called.
- The same team wrote the engine, generator and verifier. There is no external reviewer.
- Tier labels are nominal; equivalences let a few higher-tier tasks be solved earlier.
- Coverage limits compounding: a macro can only be learned if a development task exercises it.
- The evidence key is self-generated: it proves integrity of the bundle, not the author's identity.
- This is not AGI, not VAGI, and not unrestricted recursive self-improvement. It is one measured,
  governed loop in a constructed domain.

## Reproduce

```bash
python3 -m cain.cain2027.e129.cli campaign --seed 128 --generations 4
python3 artifacts/e129/verify_e129_offline.py artifacts/e129
```

The verifier imports no CAIN code. It re-applies every recorded program to the sealed inputs,
recomputes every accuracy, checks the answer commitment, the library expansion chain, the Merkle
root and the signature.
