calculate
Play demoTry asking
- Calculate 1234 times 5
- calculate 1234 times 5
- use the calculator to divide 100 by 8
- what's 2 to the power of 10
ghost skill run ghost/calculator calculate Fossify Calculator: evaluate arithmetic, read the display and history, convert units. No account, no network.
Before With skill · median tokens per task
3 more actions not benchmarked yet
maintainer-run, not independent · claude-haiku-5-5 · 2026-10-22. Every failure is counted (intention-to-treat).
12.3× fewer tokens, 2.5× faster with the skill. Task check: passed without the skill, passed with it.
| Run | Time | Tokens | Tool calls | Result |
|---|---|---|---|---|
| No skill: agent drives by hand | 29.4 s | 357,211 | 13 | passed |
With skill: find_skill, run_skill (calculate) | 12.0 s | 29,005 | 4 | passed |
Prompt: Calculate 1234 times 5
Final pair uses Fossify Math on BOTH sides; final foreground package org.fossify.math verified. BEFORE29.44s/357211 tokens/13 calls; AFTER11.99s/29005/4. Original stock-calculator pilot and earlier camera-shutdown takes superseded; all filmed token usage retained in proxy ledger.
One recorded pair on samsung-demo, claude-haiku-5-5; tokens count every input, cache and output token once per assistant message. Not a signed run: it adds no run evidence.
Demo shows: convert_units (category=Temperature, value=100, from_unit=Fahrenheit, to_unit=Celsius)
Ghost in the Droid seed skill for Fossify Calculator (
org.fossify.math), a FOSS calculator from F-Droid. Offline, no account.
| Action | What it does | Params |
|---|---|---|
calculate | Key in an expression, press =, return the result | expression |
read_result | Return the number on the display | none |
get_history | Return recent calculations from the History dialog | none |
convert_units | Convert a value with the built-in unit converter | category, value, from_unit, to_unit |
Note: the app evaluates left to right (84-2*6 gives 492).
The app: F-Droid, or https://f-droid.org/packages/org.fossify.math/. The skill:
pip install "ghost-in-the-droid>=1.6.0" # 1.6.0 is not on PyPI yet: until it ships, git clone the repo and `pip install -e .`
ghost skill install ghost/calculator
ghost skill run ghost/calculator calculate --param expression="1234*5" --device emulator-5554
ghost skill run ghost/calculator convert_units --param category=Temperature --param value=100 \
--param from_unit=Fahrenheit --param to_unit=Celsius --device emulator-5554
run_skill(name="ghost/calculator", action="calculate", params={"expression": "100/8"})
# {"display": "12.5", "value": 12.5}
Android 14 emulator (google_apis x86_64, 1080x2400, 420 dpi), Fossify Calculator 1.4.0, en-US. Signed self-runs in runs/. License: MIT.
| App version | Android | Device | Skill version | Signed self-runs (pass/total) | Dates |
|---|---|---|---|---|---|
| 1.4.0 | 14 | Google/sdk_gphone64_x86_64 (emulator, 1080x2400, 420 dpi, en-US) | 1.0.0 | 14/14 | 2026-10-05 |
Android 13 was additionally replayed on a 720x1600 Samsung SM-A037F (300 dpi, en-GB) and a 1080x2400 emulator (420 dpi, en-US). See the cross-device report for measured outcomes and limitations; other profiles remain unverified.
Historical self-run artifacts are retained in runs/, but the current registry rejects the pre-digest evidence as unbound-helper. These old counts do not establish current-version passing coverage; fresh digest-bound signed runs are pending.
The historical runs above were green. Samsung failures and subsequent fixes are recorded in the cross-device report.
+ - * / ^ and multiplication/division aliases x, ×, ÷; % has a keypad mapping but is not accepted by the arithmetic validator.The manifest targets Android >=13 and Fossify Calculator >=1.4.0, with en-US and en-GB flows. Its authoring screen is 1080x2400 at 420 dpi; measured Samsung/emulator outcomes are in the cross-device report. calculate rereads the device display and compares the left-to-right result. read_result independently rereads the display; get_history leaves History open and rereads its visible rows; convert_units rereads input, output, both unit names and the symbol, comparing each with the returned data. Conversion also checks caller input/unit names; it does not independently recompute conversion mathematics.
SOFT.md is the narrative an agent reads when the exact steps no longer match your app version and it has to adapt the skill.
Package: org.fossify.math (Fossify Calculator, GPL-3.0, from F-Droid)
Historical device check: 2026-10-05 on a headless emulator pool, Android 14 google_apis x86_64, 1080x2400 @ 420 dpi, app 1.4.0, en-US.
Skill class: HARD. Offline, no account.
Main screen: display id/result, keypad buttons id/btn_0 to id/btn_9, btn_plus, btn_minus, btn_multiply (×), btn_divide (÷), btn_power (^), btn_percent, btn_root, btn_decimal, btn_clear (C; long-press clears all), btn_equals. Toolbar: content-desc History, Unit converter, More options.
The app evaluates immediately, left to right: 84-2×6 is (84-2)×6 = 492, not 72. Agents that need precedence must split the work.
Numbers may show thousands separators (1,234); readers strip commas.
calculate(expression)Long-press C, tap one button per character of the expression, tap =, read id/result. Success: the display equals the left-to-right value of the expression (absolute tolerance max(1e-6, abs(expected) * 1e-6)).
read_result()Read id/result on the main screen as a number. The validator rereads id/result and requires its text and parsed numeric value to equal the returned display and value.
get_history()Tap History; a dialog lists id/item_formula and id/item_result pairs, newest first. Read them and leave the dialog open for validation. Only the entries visible in the dialog are returned. The validator rereads both row lists and requires exact equality with the returned lists, equal list lengths and the History dialog title. An empty history passes only when the open dialog has no rows. The next action closes the dialog during startup.
convert_units(category, value, from_unit, to_unit)Tap Unit converter, tap the category tile (id/unit_label text), then tap id/top_unit_name and pick the source unit from the popup (items read like Fahrenheit (°F)), same for id/bottom_unit_name. Long-press C, key the value (and +/- for negatives). Read id/bottom_unit_text. The validator rereads source/target unit names, input value, converted value and symbol from the device and compares them with the returned data. Both unit names and input must also match the request. This validates the device result; it does not independently recompute conversion mathematics.
+ - * / ^ and multiplication/division aliases x, ×, ÷; % has a keypad mapping but is not accepted by the arithmetic validator.Try a sample prompt or run an action by name. Replace any parameters in braces with your own values.
calculateTry asking
ghost skill run ghost/calculator calculate read_resultTry asking
ghost skill run ghost/calculator read_result get_historyTry asking
ghost skill run ghost/calculator get_history convert_unitsTry asking
ghost skill run ghost/calculator convert_units Each release is an immutable git tag. The tree hash pins the exact skill directory the installer verifies.
| Version | Released | App versions | Tree | Size | Provenance |
|---|---|---|---|---|---|
1.0.3latest | never | >=1.4.0 | pending | 23.5 KB | untagged |
maintainer-run, not independent
Model: claude-haiku-5-5 · Date: 2026-10-22
Same task across three arms. Skill replay runs without a model. Medians include every valid episode, including failures.
N=5 seeds · 2026-10-22
| Arm / provenance | Pass a/b · Wilson 95% CI | Tokens, median [q1–q3] | LLM calls, median | UI actions, median | Wall, median |
|---|---|---|---|---|---|
| Agent aloneclaude-haiku-5-5 · 2026-10-22 Skill 1.0.2 · SHA b37d4c36 | 5/5 · Wilson 95% CI 0.566–1.000 | 184,681[132,498–267,125] | 6 | 6 | 42 s |
| Skill replayno model · 2026-10-22 Skill 1.0.2 · SHA b37d4c36 | 5/5 · Wilson 95% CI 0.566–1.000 | 0[0–0] | 0 | 6 | 43 s |
| Agent + skillclaude-haiku-5-5 · 2026-10-22 Skill 1.0.2 · SHA b37d4c36 | 5/5 · Wilson 95% CI 0.566–1.000 | 61,837[61,677–111,942] | 3 | 1 | 60 s |
Agent + skill vs agent alone: 5 paired seeds; 0 fail → pass, 0 pass → fail; McNemar exact p = 1. Median tokens saved: 123,150.
Skill replay vs agent alone: 5 paired seeds; 0 fail → pass, 0 pass → fail; McNemar exact p = 1. Median tokens saved: 184,681.
| Task | Arm | Seed | Model | Pass | LLM calls | Tokens | UI actions | Wall | Trace |
|---|---|---|---|---|---|---|---|---|---|
| Calculator "3+3" | Agent + skill | 0 | claude-haiku-5-52026-10-22 · Skill 1.0.2 · SHA b37d4c36 | pass | 3 | 62k | 1 | 60 s | |
| Calculator "3+3" | Skill replay | 0 | replay2026-10-22 · Skill 1.0.2 · SHA b37d4c36 | pass | 0 | 0 | 6 | 43 s | |
| Calculator "3+3" | Agent alone | 0 | claude-haiku-5-52026-10-22 · Skill 1.0.2 · SHA b37d4c36 | pass | 8 | 185k | 6 | 55 s | |
| Calculator "3+3" | Agent + skill | 1 | claude-haiku-5-52026-10-22 · Skill 1.0.2 · SHA b37d4c36 | pass | 6 | 254k | 2 | 49 s | |
| Calculator "3+3" | Skill replay | 1 | replay2026-10-22 · Skill 1.0.2 · SHA b37d4c36 | pass | 0 | 0 | 6 | 41 s | |
| Calculator "3+3" | Agent alone | 1 | claude-haiku-5-52026-10-22 · Skill 1.0.2 · SHA b37d4c36 | pass | 6 | 132k | 6 | 46 s | |
| Calculator "3+3" | Agent + skill | 2 | claude-haiku-5-52026-10-22 · Skill 1.0.2 · SHA b37d4c36 | pass | 3 | 62k | 1 | 62 s | |
| Calculator "3+3" | Skill replay | 2 | replay2026-10-22 · Skill 1.0.2 · SHA b37d4c36 | pass | 0 | 0 | 6 | 43 s | |
| Calculator "3+3" | Agent alone | 2 | claude-haiku-5-52026-10-22 · Skill 1.0.2 · SHA b37d4c36 | pass | 6 | 115k | 6 | 42 s | |
| Calculator "3+3" | Agent + skill | 3 | claude-haiku-5-52026-10-22 · Skill 1.0.2 · SHA b37d4c36 | pass | 3 | 62k | 1 | 60 s | |
| Calculator "3+3" | Skill replay | 3 | replay2026-10-22 · Skill 1.0.2 · SHA b37d4c36 | pass | 0 | 0 | 6 | 41 s | |
| Calculator "3+3" | Agent alone | 3 | claude-haiku-5-52026-10-22 · Skill 1.0.2 · SHA b37d4c36 | pass | 8 | 288k | 5 | 37 s | |
| Calculator "3+3" | Agent + skill | 4 | claude-haiku-5-52026-10-22 · Skill 1.0.2 · SHA b37d4c36 | pass | 4 | 112k | 1 | 67 s | |
| Calculator "3+3" | Skill replay | 4 | replay2026-10-22 · Skill 1.0.2 · SHA b37d4c36 | pass | 0 | 0 | 6 | 43 s | |
| Calculator "3+3" | Agent alone | 4 | claude-haiku-5-52026-10-22 · Skill 1.0.2 · SHA b37d4c36 | pass | 6 | 267k | 5 | 37 s |
Display only: points come from signed run artifacts.