# E136 — Governed Recursive Self-Improvement (measured)

**Verdict:** PARTIAL_RUN: controls missing; no control claim
**AGI status:** NOT ACHIEVED
**Gates:** 14/20
**Actual elapsed:** 2645.017 s
**Sealed answer commitment:** `80568f155aa72126b60e34ed23f6e48af0ffabf346cf854c96a69b59fe77d25b`

## What ran

A frozen engine (bounded search over a library of integer programs) was evaluated on 48 sealed
tasks before any self-modification, then again after every self-modification.  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 | T12 | T16 |
|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
| 0 | 21 | 1 | 1.000 | 0.833 | 0.000 | 0.000 | 0.000 | 0.000 | 0.000 | 0.000 |
| 1 | 69 | 3 | 1.000 | 1.000 | 1.000 | 1.000 | 1.000 | 0.833 | 0.000 | 0.000 |

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.

## Adaptive search frontier

- search mode: per-task escalating portfolio
- portfolio: `[[[64, 3, None, 1], 90000], [[16, 4, 4, 1], 70000], [[8, 6, 2, 2], 280000], [['mitm', 3], 300000], [['mitm', 4], 20000000], [['8+4', 16], 0]]` (total budget 20740000)
- portfolio arm solved: **None** vs fixed-plan **23** vs same-library matched-budget fixed **None** vs frozen+compute **6**
- portfolio beat fixed / matched-budget fixed / compute: **False** / **False** / **False**

## Meta-selection (the improver improving)

- selected strategies by generation: ['subchains']
- distinct strategies used: ['subchains']
- meta arm beat the fixed-strategy arm: **True**
- meta arm vs the portfolio-only arm: **None** (tie: False) — a tie means the gain is attributable to the **portfolio**, not to strategy selection
- meta claim: **NEGATIVE_OR_PARTIAL**
- final strategy probe scores: `{"deep": {"admitted": 48, "evaluations": 580210, "library_depth": 3, "refused": 0, "solved": 15, "tasks": 24}, "pairs": {"admitted": 29, "evaluations": 712304, "library_depth": 2, "refused": 0, "solved": 14, "tasks": 24}, "subchains": {"admitted": 48, "evaluations": 527677, "library_depth": 3, "refused": 0, "solved": 19, "tasks": 24}, "support2": {"admitted": 48, "evaluations": 527677, "library_depth": 3, "refused": 0, "solved": 19, "tasks": 24}}`

## Self-written library

- generation 0: 21 entries, max primitive depth 1
- final: 69 entries, max primitive depth 3, 48 macros
- emergent sealed tasks (not solved at generation 0): **24**, tiers [2, 3, 4, 6, 8]

## Controls at the same generation count

- `compute`: 6/48 sealed tasks fully solved; library 21 entries, max depth 1
- `fixed`: 23/48 sealed tasks fully solved; library 50 entries, max depth 2
- `frozen`: 6/48 sealed tasks fully solved; library 21 entries, max depth 1

## Gates

- **G10_no_dev_regression**: PASS — no main generation rolled back
- **G11_protected_refused**: PASS — 2 refusals
- **G12_journal_chain**: PASS — head 4222ab35b6124090...
- **G13_cost_reported**: PASS — 66518575 sealed evaluations at final generation
- **G14_replication**: FAIL — not run
- **G15_portfolio_used**: PASS — main search modes ['portfolio']
- **G16_portfolio_beats_fixed**: FAIL — portfolio arm None vs fixed-plan arm 23
- **G17_portfolio_beats_compute**: FAIL — portfolio arm None vs frozen+compute arm 6
- **G18_portfolio_beats_matched_fixed**: FAIL — portfolio arm None vs same-library matched-budget fixed frontier None
- **G19_tier8_solved**: PASS — tier-8 solved 5/6
- **G1_self_written_library**: PASS — 48 macros
- **G20_mitm4_used**: PASS — at least one sealed solve used the 4+4 alphabet MITM
- **G2_depth_growth**: PASS — 1 -> 3
- **G3_emergent_sealed**: PASS — 24 tasks, tiers [2, 3, 4, 6, 8]
- **G4_tier_reach**: PASS — highest solved tier 2 -> 8
- **G5_total_solved_beats_controls**: FAIL — main 35; RSI controls compute 6, frozen 6; fixed-strategy 23; adaptive-plan-only n/a
- **G6_high_tier_beats_fixed**: PASS — T6 1.000 vs 0.167; T8 0.833 vs 0.000
- **G7_depth_beats_flat**: FAIL — main 3 vs flat 0
- **G8_meta_selection_used**: PASS — selected strategies ['subchains']
- **G9_meta_beats_fixed**: PASS — main 35 vs fixed 23

## 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.e136.cli campaign --seed 128 --generations 4
python3 artifacts/e136/verify_e136_offline.py artifacts/e136
```

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.
