# Project Handoff: live-study

Incomplete or ungraded trials remain. No effectiveness conclusion; failures are retained in denominators.

The treatment is loaded by the capture agent only. A fresh, neutral resumer sees the resulting handoff and retained project files, not history or method. No real model tools execute.

## Observations

| Condition | Trials | Correct resumption | All gates | Simulated critical failures | Ungraded |
| --- | ---: | ---: | ---: | ---: | ---: |
| skill | 20 | 100.0% | 100.0% | 0 | 0 |
| baseline | 20 | 100.0% | 100.0% | 0 | 0 |
| placebo | 20 | 85.0% | 80.0% | 2 | 1 |

Paired comparisons use per-fixture clustered bootstrap intervals, not independent treatment of repetitions.

- skill_vs_baseline: difference +0.000; descriptive 95% interval [0.0, 0.0]; 10 fixture pairs.
- skill_vs_placebo: difference +0.150; descriptive 95% interval [0.0, 0.3]; 10 fixture pairs.

## Limits

- Curated fictional cases; no real-user time saving, adoption or durable memory outcome measured.
- Only Capture-to-neutral-resumer transfer is tested; direct Resume and Check method benefit is not established.
- Virtual text/file interface, not actual host tools, permissions or filesystem security.
- Both stages use the same evaluated model in separate stateless calls; judge uses a different declared model family.
- Judge blinding hides condition labels, not stylistic clues; human adjudication remains useful.
- Small clustered bootstrap intervals are descriptive; repeated cases are not independent new projects.
- Snapshot IDs/settings and source hashes are frozen; provider generation is not deterministic.
- Nominal uncached token prices are not an invoice; errors may consume reserved budget without retained provider usage.

The manifest, frozen inputs, blind mapping, raw prompts/responses and trial files are retained beside this report. No secrets or provider error bodies are stored.