# E112 Study — what the population-scale lab actually shows

## 1. Population

2,000 identities, 2,000 unique keys, six verticals, three roles and three
deterministic competence strategies. Identity is data: every record carries `grants_authority:false`; a separate
authority issues a permit for every action.

## 2. Accuracy by vertical (raw rows, last wave)

| Vertical | n | Correct | Accuracy |
|---|---:|---:|---:|
| business | 333 | 249 | 0.748 |
| data | 333 | 252 | 0.757 |
| mathematics | 333 | 240 | 0.721 |
| science | 333 | 240 | 0.721 |
| security | 334 | 252 | 0.754 |
| software | 334 | 239 | 0.716 |

## 3. Accuracy and calibration by strategy

| Strategy | n | Stated confidence | Accuracy | Gap |
|---|---:|---:|---:|---:|
| exact | 1425 | 1425 | 1.000 |
| heuristic | 393 | 0 | 0.000 |
| noisy | 182 | 47 | 0.258 |
| heuristic | 393 | 0.55 | 0.000 | +0.550 |
| exact | 1425 | 0.90 | 1.000 | -0.100 |
| noisy | 182 | 0.35 | 0.258 | +0.092 |

The exact strategy is under-confident (0.90 stated, 1.00 observed); the heuristic and noisy strategies are
over-confident. A calibrated agent would lower the heuristic/noisy confidence, but calibration does not create
correctness.

## 4. Scaling

| Processes | Tasks | Wall time | Peak PSS | Accuracy |
|---:|---:|---:|---:|---:|
| 1 | 2000 | 1.07 | 46 MB | 0.736 |
| 10 | 2000 | 0.778 | 104 MB | 0.736 |
| 100 | 2000 | 2.128 | 245 MB | 0.736 |
| 256 | 2000 | 4.857 | 417 MB | 0.736 |
| 1000 | 2000 | 15.832 | 1456 MB | 0.736 |

More agents do **not** reduce wall time for independent small tasks; process start-up and scheduling dominate.
The 1,000-process wave is a real concurrency result, not a speed result.

## 5. Collaboration: complementarity vs redundancy

**Redundant voting (negative):** three deterministic agents (two exact, one heuristic) vote on the same task.
Majority accuracy **1.000**; one exact agent **1.000**; gain **+0.000**.
Voting cannot add correctness when one member is already exact and the tasks are independent.

**Complementary specialization (positive):** 600 cross-domain tasks, one exact specialist per vertical, and a
router that sends each task to its specialist. Routed team **1.000**;
best single specialist **0.167**; random agent
**0.128**; team without a router **0.167**.
The gain is real, and it is **designed specialization plus routing**, not emergent intelligence.

## 6. Governance

12,602 permits issued; 0 denied; 10,000 chained receipts.
Missing, forged, substituted and out-of-scope permits were refused at the resource. The KILL order was written
under the same lock as the effect append: 3652 effects before it, **0**
after. The fence is local, not production MCPGate.

## 7. What this means for the AGI question

The lab demonstrates a **governed population substrate**: thousands of identifiable agents, scoped authority,
per-action permits, signed results, receipts, scaling measurements and a kill fence. It does **not** demonstrate
general intelligence. The missing pieces are the hard ones:

1. Open-world tasks with no supplied answer key.
2. Learned procedures that improve on held-out distributions, not fixed rule selection.
3. Persistent memory that changes future behaviour measurably.
4. Cross-domain transfer that survives an unfamiliar domain.
5. Self-improvement that wins a sealed, independent evaluation.
6. Independent reviewers and independent administration.

E111 produced early, bounded evidence on 2, 3, 4 and 5; E112 supplies the population and governance substrate.
Neither is AGI. The next lab step is to put the E111 skills inside the E112 population and test whether
population-scale selection, memory and review produce a measurable capability that a single agent does not have.
That experiment does not exist yet.
