# E125 Limitations

Everything below is a limit of the measured work, not a disclaimer added afterwards.

## Scope and independence

- The confirmatory challenge is **constructed**. The same three-person team wrote the engine, the
  generator and the verifier. It is not an external benchmark.
- **No external reviewer** ran the package. Independent human reproduction is
  **BLOCKED_EXTERNAL_REVIEW**. Internal verification is not human certification.
- The evidence signing key is self-generated. The SHA-256 commitment proves integrity, not
  independent witnessing.
- The repository's branch label (`evolution-79`) is historical and does not describe the content.

## The confirmatory challenge

- C families are constructed with an **invertible outer primitive** so that an inverse-composition
  solver can, in principle, decompose them. The result is conditional on that assumption.
- The primitive vocabulary and the eight paradigms are designed. "New rule" means new **inside
  this grammar**, not a new scientific mechanism.
- Compute is matched in **candidate hypothesis evaluations** (800 per family per arm). Wall-clock,
  energy, arithmetic operations and memory are not matched.
- A general-purpose local model (Qwen2.5-0.5B) was not run as a compute-matched arm in the
  confirmatory challenge. It is not used for H1.
- D families (new primitives) are partly **in-support interpolation**. The 0.567 is not acquisition
  of a new primitive; out-of-support values fail.
- E families (grids and graphs) are **0/60**. The engine abstains; this is the honest boundary.

## Continual learning

- Six constructed task vocabularies are not six real industries.
- The mechanism pool repeats every four episodes. This is a designed continual-learning loop, not
  a natural stream.
- The aggregate gain is **not significant** (p=0.77). Per-domain gains do not survive pooling.
- Poisoning was refused 3/3 **when offered through the verification gate**; the retrieval-only
  control has no gate and refused 0/3.
- Retention uses fresh data from the same mechanism, not a new domain.

## Causal world model

- Binary observed variables only. **No hidden confounder** is modelled in this version.
- Counterfactuals are **NOT_IMPLEMENTED** and are not claimed.
- Regime-shift detection is a mean-shift heuristic, not a principled change-point model.
- Active selection **ties random** (0.75 vs 0.75). The measured gain is intervention vs passive,
  not the intelligence of the active policy.
- The candidate graph class is finite and designed.

## Transfer

- The six domains differ by scale, offset, output range and feature name. They are not six real
  industries.
- The only adaptation is an affine input/output wrapper.
- Positive transfer shows **no accuracy gain** over from-scratch learning at the same destination
  label budget (−1.96 points, p=0.54).
- Two harmful pathways were falsely accepted.

## Model adaptation

- One 135M CPU model, one lexicon task, one seed.
- LoRA reaches 0.667, not 1.0. The full arm abstains when its dev check is below 0.8.
- The base model is not safety-aligned; **no safety alignment evaluation** is claimed.
- The forgetting control is six short factual prompts, not a benchmark.
- No PEFT library was used; the LoRA implementation is local and manual.

## Collective

- Agents are deterministic rule-based detectors, not LLMs.
- Byzantine agents emit random labels with high confidence and fabricated evidence; this is a
  model of a liar, not an adaptive adversary.
- Byzantine agents were selected in 60/300 tasks and flipped 13 answers. Resistance is
  **imperfect**.
- One family per task; specialists see feature subsets; the task generator is constructed.

## Self-improvement

- Candidates are search-policy knobs, not neural weight updates.
- The evaluation set is 20 of 40 development families from the same generator.
- 0 candidates were promoted; this does not show that no improvement is possible.
- A promotion would still require a human operator and an external canary.

## Governance

- Three OS processes on **one host** over loopback, with two independent keys.
- This is **not** three machines, three administrators, or a distributed production failure
  domain.
- The kill-switch test is a signed file check, not a hardware fence.
- The ledger is local and hash-chained; it is tamper-evident, not tamper-proof.

## Mission

- Actual duration **301.77 s**, not 24 hours.
- Tasks are deterministic (statistics, scans, solves, verification, memory chains). They are real
  but not open-ended research.
- One host; process restarts, not host reboot.
- "Productive" means a recorded result; it is not an external productivity judgement.

## Public surfaces

- The bundle is published to ClawX, a Studio proof mirror and the docs tree. Hugging Face and
  mcpgate.online were **not** deployed.
- No live metric is typed by hand in the bundle, but the existing site panels outside this bundle
  are not rewritten by E125.

## Claims that are not made

No AGI, VAGI, superintelligence, consciousness, human-level competence, external certification,
SOC 2, ISO 42001, hardware attestation, production readiness, customer revenue, patent, or
official leaderboard position.
