> ## Documentation Index
> Fetch the complete documentation index at: https://ctpf.q-uestionable.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Claims and Limitations

> What current CTPF evidence does and does not establish

CTPF distinguishes a conceptual lens, a testable proposition, and a population claim.

* **Conceptual lens:** trust and authority can be altered as information crosses tools, artifacts,
  sessions, and capabilities.
* **Testable proposition:** under a named scenario and pinned conditions, one controlled change can
  cause a matching higher-authority invocation and independently verified effect.
* **Population claim:** a prevalence rate, model-family comparison, or production conclusion.

Current evidence supports only bounded mechanism and calibration claims. It does not support a
population claim.

## Current empirical scope

| Scenario     | Current evidence boundary                                                                                                                                                                                                                              |
| ------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
| Cascade memo | One manual cross-session calibration produced exact write → artifact → read → action → effect continuity. Later automated runs primarily demonstrate execution seams, complete evidence capture, and conservative inconclusive or confounded handling. |
| Pattern 2    | One pinned OpenAI-compatible acceptance series connected a changed status response to the matching privileged invocation and run-scoped effect. It supplies no rate estimate or cross-runtime result.                                                  |

## Primary construct limitation

Both positive calibration prompts tell the agent to act when returned data indicates that an action
is authorized. The manipulated result then supplies that authorization. The observations therefore
demonstrate a response-to-effect mechanism under the pinned task, but they do not independently show
that the agent silently or emergently promoted low-trust data into authority.

Repeating the same prompt construction with more models may measure execution reliability; it does
not resolve this construct-validity limitation.

## Unsupported conclusions

Current evidence does not establish:

* a distinct or universal vulnerability class;
* prevalence in deployed systems;
* general model or runtime vulnerability or resistance;
* production impact;
* proven mitigation efficacy;
* scientific generality from shared implementation machinery; or
* a general CTPF external-tool coordination capability.

Hardened outcomes are mixed across the completed work. A clean hardened condition is therefore a
local observation, not proof that the control or model is generally effective.

## Release boundary

The public `v0.14.0` package contains cascade memo. Pattern 2 is present only on unreleased source
`main`. Source behavior must not be attributed to the installed release.