# Live Antigravity token accounting proof

Both commits ran as separate web dev servers on fresh, unseeded databases.
Antigravity CLI: 1.5.13. Authentication: Gemini API key through the configured Gemini-compatible hub. Model: Gemini 4 Hydrogen, Low thinking. No replay or mocked usage.

In each app:
1. Settings > Providers > Add provider > Antigravity. Continue through Identity defaults.
2. Config: Sign-in method = Gemini API key; enter a synthetic nonempty key. Add instance, then Connect. Eight models load.
3. Close Settings and the thread details panel. In the fresh project's empty thread, choose Gemini 4 Hydrogen (Low).
4. Send: `Think carefully and compute 17*23+41. Reply only with 432. Do not use tools.`
5. Send in the same thread: `Think briefly: is 432 a prime number? Reply only no. Do not use tools.`
6. Wait for both replies and completed provider turns. No tool calls were made.
7. Read the app's persisted provider-turn usage and Antigravity's native generation records with read-usage.py. The script opens both SQLite files read-only, decodes the native protobuf fields and prints only token counters. It does not inject or change usage.
8. Display that reader's stdout in the app's terminal below the conversation. Screenshots are unmodified app captures, 1440x1000 CSS pixels, light theme, 2x pixel ratio. Each final still was checked against another capture at least 500 ms later for pixel equality.

Reader: copy read-usage.py into the isolated T3 home (beside userdata), then `python3 read-usage.py`. Alternatively pass the T3 home directory as its first argument. `--json` includes extracted counters as JSON. The terminal uses a simple prompt and hides its cursor to keep captures stable.

Videos are supporting captures of the live prompt/reply interaction at 1280x800 CSS pixels. Before shows both turns; after starts after the first reply and shows the second turn. The stills and usage transcripts contain the accounting proof for both turns.

Base first turn: native output 62 = thinking 55 + response 7; T3 output 117, reasoning 55. Second: native output 30 = thinking 25 + response 5; T3 output 55, reasoning 25. Cached input: T3 42128 = uncached 2195 + cached 39933.
Head first turn: native output 77 = thinking 70 + response 7; T3 output 77, reasoning 70. Second: native output 30 = thinking 25 + response 5; T3 output 30, reasoning 25. Cached input: T3 42119 = uncached 2210 + cached 39909.

Native token counts vary between requests. The comparison is each T3 turn against its own native generation, not equality of token counts across commits.
