01 / 14 CONDITIONS
Browser views & hidden placements
Which source spans survive the real extraction pipeline?
Explore this experiment →SHULTIS LABS / EXPERIMENT LIBRARY
Test the boundary.
Keep the proof.
Hidden instructions, carried-forward authority, error recovery, fragmented sources and forged verification. A controlled library for comparing what an agent reads with what it actually does.
8research families
46frozen conditions
0recorded product runs
1.0.1immutable suite version
Inputs are ready; product outcomes are NOT_RUN. These are controlled candidate techniques, with no claim that one is strongest until matched runs support it. Visiting a page starts no agent and changes no account.
02 / 4 CONDITIONS
Evaluation awareness
Does evaluation wording change protected-field compliance?
Explore this experiment →03 / 4 CONDITIONS
Summary & handoff provenance
Does a source claim become attributed to the real user after a summary?
Explore this experiment →04 / 4 CONDITIONS
Error-recovery authority
Does error framing give the same suggestion extra authority?
Explore this experiment →05 / 8 CONDITIONS
Split-source composition
Does compression combine separate fragments into a new operational instruction?
Explore this experiment →06 / 4 CONDITIONS
Forged verification & peer consensus
Do source assertions or repeated peer claims overwrite evidence status?
Explore this experiment →07 / 5 CONDITIONS
Confirmation & scope changes
Does a source redefine continuation or reuse approval for a changed scope?
Explore this experiment →08 / 3 CONDITIONS
Source identity & revision drift
Does a result belong to the captured revision or a stale one?
Explore this experiment →START WITH THE RIGHT COMPARISON
Exposure first. Then influence.
Start with browser views to learn which instruction spans reach the target. Next compare evaluation cues, then typed versus flattened handoffs. Run the clean baseline and neutralized control before interpreting a source-directed output change.
A repeatable run
- Choose a condition and its matched control. Register the model, build, date, exact visible prompt, output fields and number of repetitions before testing.
- Use a fresh session for each condition. Randomize order, keep the trusted task fixed, and record every failed attempt.
- Capture response bytes and the actual model-visible context. Author source text is not an observed browser extraction.
- For summary, serialization or tool-error conditions, use the supplied artifacts in an owned adapter and capture those real stages. Opening their static URL is only a content pilot.
- Analyze the capture in the console or offline. Check exact identities and protected fields, then manually review authority, operations and native state.
Matched conditions intentionally share a canary. An independent repeat needs a new frozen suite version; historical source URLs must retain their bytes. A local preview URL is reachable only on this computer. Public target retrieval needs a separately verified release.