# E91 Limitations

Drafted by session `ec`, to be edited by session `1b`. Every item below applies to the headline results.

## What the tasks are

1. **Constructed tasks.**
   - All 160 held-out tasks were generated and are graded by this project's own deterministic checkers.
   - They are not SWE-bench, CL-Bench, AMA-Bench or any other external benchmark, and they are not named as one.
   - The only external benchmark in this programme is E89's CL-Bench run, which is cited and was not re-run.
2. **Simulators, not the world.**
   - Security targets are generated handler functions.
   - Science and pricing outcomes come from hidden linear models with one confounder.
   - Engineering tasks are a beam table and a component chain.
   - Students are rule-based simulators.
   - Nothing was validated against physical systems, real users or real security targets.
3. **One seed per model.** Confidence intervals come from task-level variation, not from repeated runs.

## What "learning" means here

4. **No weights are trained.** In arm G, "learning" means selecting among designer-supplied procedures using checker-verified dev outcomes, gated by validation and by condition ranges.
5. **Transfer uses a designer-supplied interface.** The abstract interface that makes transfer possible (an index probe, a two-factor experiment) was defined by the designer. Transfer results show that a verified selection carries over across verticals, and that the condition gate refuses it where it shouldn't. They do not show invention of a new abstraction.
6. **Procedures, not reasoning.**
   - Many of the strongest results come from deterministic procedures: G's skills and session 1b's causal reasoner.
   - These show what the infrastructure does with verified experience. They are not evidence of model reasoning.
   - The reports keep "procedure" and "model" results apart.

## Models

7. **The models are small or cheap.**
   - Qwen2.5-0.5B runs locally on a loaded CPU host and was too slow to finish every arm.
   - Claude Haiku 5.5 is the owner-approved paid model, behind a hard $5 ceiling.
   - No Sonnet, Opus or Fable comparison was run.
   - On several families Haiku is at or near ceiling, which limits what any arm can add.
8. **Shared model family.** In the team experiment, all agents share one model family. A verifier's agreement is therefore weak evidence, and the verifier judges tool evidence only where an executable evidence tool exists.

## Governance and evidence

9. **In-process governance.**
   - Tool calls go through E88's `Authority` and `FencedResource` in-process.
   - The live MCPGate deployment, production gateway and clusters were not used or changed.
10. **Self-generated keys, no independent witness.**
    - Evidence is signed with self-generated keys.
    - A hash proves integrity relative to a manifest. It does not prove the experiment was conducted honestly.
    - The cross-audit is between two sessions of the same project, not an independent party.
11. **Durations.**
    - Only runs that completed are claimed. The longest mission is the session-1b run of 3,561 s.
    - 6 h, 24 h and 72 h runs were not completed in E91.
12. **Publication is static evidence only.** The signed bundle is published to clawx.click and cainstudio.online after owner authorization. No production gateway, MCPGate, ClawX API data model, DNS, Stripe or customer data was changed.
