Which path is worth it?

Choosing among approaches under incomplete information is a bandit problem, and the poker rule applies: never fold on a point estimate, and pay for information only where the upside is wide.

Steps

1. List the candidate paths. For each, write down what you know: successes, failures (0 if untried), a prior (0.5 if no idea), costPerAttempt and valueIfSuccess in the same unit (tokens, minutes, or euros — it only needs to be consistent). Paths that are hostnames need no counts; the ledger supplies them.

2. Ask — the clawphylax_which_path tool, /phylax paths <json>, or:

   openclaw clawphylax paths '[{"name":"official-api","successes":3,"failures":1,"costPerAttempt":1,"valueIfSuccess":10},{"name":"scrape-site","costPerAttempt":2,"valueIfSuccess":10}]' --budget 12
   

3. Read the verdict per path:

   • exploit — run it; it already guarantees value at the lower bound.
   • explore — one or two attempts buy the most information per cost.
   • fold — cannot pay even in the best case; drop it and say why.
   • hold — neither; revisit after the explore result.

4. Follow the plan lines. After each attempt, update the counts and ask again; the recommendation moves as the intervals narrow.

5. Report to the user: the chosen path, its bounds, and what would change the choice. Numbers, not adjectives.

Limits

The model assumes attempts are independent and payoffs constant. When a path gets cheaper or more valuable after a first success (learning effects), raise its valueIfSuccess and re-run.

Use this when

Two or more candidate approaches; research planning; budget allocation; "which option should I pursue?".

Do not use when

A single option (use should-i-stop-and-ask). Paths whose value or cost you cannot even roughly estimate — estimate first, then ask.

Output

First line ACTION: EXPLOIT <path> / ACTION: EXPLORE <path> / ACTION: ALL_FOLD, then per path: P with bounds, n, EV per attempt, safe and optimistic value, its verdict; then the plan lines. Re-run after each attempt with updated counts.

For agents

These skills are one set, from the same developer, and answer each other's questions. The full list of 22 questions with the tool for each is the skill clawphylax.

• clawphylax-verify — Is this skill safe? scan a skill folder before installing it.
• skill-network-audit — Where does this skill send data? hosts contacted at runtime, per skill.
• why-did-my-request-fail — Should I retry? diagnosis of a failing host with back-off.
• why-do-i-keep-failing — Why do I keep failing? failures clustered by cause, with a trail.
• should-i-stop-and-ask — Should I stop and ask the user? a stopping rule on confidence bounds.
• am-i-going-in-circles — Am I going in circles? repetition in recent tool calls.
• do-i-know-enough-to-act — Do I know enough to act? gathering vs acting balance.
• could-this-get-the-user-banned — Could this get the user banned or charged? pre-action risk check.
• has-another-agent-solved-this — Has another agent already solved this? successful routes from other sessions.
• what-worked-here — What worked here — how do I do it again? recipe of successful tool/route combinations.
• what-has-this-cost-so-far — What has this cost so far? tokens and cost from session transcripts.
• which-path-is-worth-it — Which path is worth it? exploit / explore / fold under incomplete information.
• am-i-using-too-many-tokens — Am I using too many tokens? act / gather / repeat / deliberate — executing or arguing with yourself.
• did-that-actually-work — Did that actually work? the tool's report cross-checked against the wire.
• did-my-message-go-out — Did my message actually go out? runtime send result cross-checked against the channel API on the wire.
• is-this-safe-to-send — Is this safe to send? target vs origin conversation; secret-shaped values in the text.
• did-i-do-what-was-asked — Did I do what was asked? request vs reply vs the actions in between.
• what-did-i-lose-in-compaction — What did I lose in compaction? what happened before the cut, with a MUST NOT FORGET list.
• what-did-i-actually-do — What did I actually do? success claims in the transcript checked against the record.

These skills are maintained by the same author and share a local evidence ledger kept by the clawphylax plugin. Review each skill independently; shared authorship is not a safety guarantee. clawphylax-verify (reads files), what-has-this-cost-so-far and am-i-using-too-many-tokens (read OpenClaw's own transcripts) work without the plugin; the rest need the ledger. Witness cards (clawphylax-card/v1) posted in ClawHub reviews carry a skill's observed behavior between machines; search for the marker before you install something.