# Ten Arms

> A search plan for one lost everyday item - keys, glasses, an umbrella, a remote. You give the item, the last moment you surely had it, what you did since and up to 9 spots it could be. A metered reading goes through your story spot by spot: the words of yours that bear on each spot, a score, how easily a careful look would show the item there, why it could be there and how to search it, plus the steps of your story that no spot covers. The page turns that into chances, keeps a share for somewhere not listed, and re-orders the search with Bayes' rule - for free - every time you mark a spot searched. It draws a squid whose two long tentacles reach the two best next looks.

## What it does

- Free, in the browser: the form, a rough free plan (every spot weighted alike), search marks (not yet / quick look / searched well), "found it", the squid drawing (.svg/.png), Markdown, CSV and JSON downloads, saved searches.
- Metered (one task, `search-plan`, on gpt-terra): the reading of your story for each spot.
- The model never computes chances. The page does: somewhere not listed = 10 + 10 per uncovered step (up to 50); the rest is shared in parts of score + 1; look-ease easy 0.9, fair 0.6, hard 0.3; a missed look keeps chance x (1 - look-ease x f), f = 1 for searched well and 0.5 for a quick look, then everything is renormalised.
- Order: chance now x look-ease, then the easier spot, then your own order.
- Verdict, first match wins: widen the search (somewhere not listed holds 40% or more), go straight there (the best spot is at least twice as good a look as the runner-up), otherwise sweep the top three.

## The reply contract

One JSON object: `summary` (where the trail goes cold), `spots[]` (one per listed spot, in order: `n`, `score` 0-3, `spot_it` easy/fair/hard, `evidence` copied from the person's words, `why`, `how_to_look`), `elsewhere` (`uncovered` steps, `why`, `retrace`) and `closing`. The page checks every quote against what the person typed, the spot order, word limits, invented numbers, times and ranking words, and unsafe tips.

## Safety

No climbing, no reaching into appliances or drains, and ask before searching anyone else's things or space. If a wallet, phone, ID or keys with an address may be lost outside, contact the bank, provider or venue too, and never post personal details publicly. Chances are rough estimates from the story, not a promise.

## Sources

- Built on the SkillSafe skill @davila7/pymc-bayesian-modeling (https://skillsafe.ai/skill/@davila7/pymc-bayesian-modeling) - Bayesian priors and posterior updating - narrowed to Bayesian search for one lost item.
- The squid drawing nods to the giant squid, the classic thing that hides well in a large, dark search space.

## Pages

- https://ten-arms.skillsafe.ai/ - the app
- https://ten-arms.skillsafe.ai/api.html - the HTTP API with worked examples in eight languages