HUMAN + AGENT · HOSTED OR OFFLINE ENGINE
Explore what the declared evidence warrants.
Original declaration engine. This advanced path preserves the original rules, including declaration and prose-sensitive behavior. For checked quotations, contradiction handling and agent records, use the record workbench.
Your agent can prepare a structured claim packet and run the PFABC Python evaluator. Use the hosted evaluator when this page is served by the lab, or run the downloaded local runtime for an offline path. Follow a deliberately paired workspace when you want to share one bounded case with an agent.
This evaluates supplied declarations under PFABC’s rules. It does not collect measurements, authenticate evidence, or determine whether an AI is conscious.
Connect to an evaluator
A hosted lab connects to its same-origin Python Worker automatically when available. For the offline local path, run python3 native/server.py --port 8769 from the downloaded project folder, then open the local engine lab.
The hosted evaluator is stateless for ordinary evaluations. An optional shared workspace stores its latest successful packet and result on the server until you end it or its 30-minute expiry; packet limits are shown after connection. No provider login, model call, or API key is used.
Use the downloaded local runtime
Download the complete source and offline runtime · Archive SHA-256
The local server keeps one latest successful result in memory for clients holding its local capability token. It binds to loopback and does not make the hosted workspace available. Restarting it clears that local result.
Start with python3 native/cli.py --help. Enable local following below to see evaluations submitted by your local agent. Connecting your agent to the lab does not give PFABC access to the agent’s private history or hidden model state.
Optional shared workspace
Create an isolated 30-minute server workspace only when you want an agent to submit into the same case. Creation requires your consent because the latest successful packet text is retained on the server until deletion or expiry.
No shared workspace. Ordinary evaluation remains available.
Closing this tab does not guarantee deletion. Use End shared workspace, or let the 30-minute expiry clear it.
1. Prepare the packet
Ask your agent to make the claim, system boundary, evidence, theory choice, and missing observations explicit. Inspect its draft before evaluating it.
A saved draft is available in this tab. It has not been submitted.
A field saying a control passed is a declaration by the packet’s author. The evaluator does not establish that the control was performed. Keep fictional inputs labeled as fictional. Raw engine heuristic annotations can fire on bridge identifiers or quotation presence; they are disclosed engine output, not observed evidence.
Make the operating assumptions visible
The target_realist mode keeps protected “presence-for” distinct from reports, behavior, and processing by assumption. This does not prove that protected target exists. The illusionist and theory_indexed modes frame the question differently. Theory choices and bridge thresholds are assumptions to inspect, not measurements.
Read the unchanged formal specification and the runtime limitations. You can use the conversation exercise without endorsing a metaphysical position.
2. Inspect the result
Evaluate a packet to begin. No evaluation has been performed on this page.
Requested claim, conditional on the packet
Declared evidence credits
Bridge states
Suggested next checks
Full input and evaluator output
The result includes the exact structured input and its consistency digest. The digest does not authenticate an observation or a person.
3. Explain what you learned
A useful explanation credits the bounded observation, states which stronger inference remains open, and proposes a relevant next check. Distinguish an unmeasured bridge from a failed one; conflicting declared measurements remain contested.
Share the input and the assumptions alongside the result. A software result alone does not validate a theory or an empirical claim.
Try the short worked exercise →