Gemma pointer pilot: does one line make a small model read past decisions?
Project / context
Does one line in the prompt make a small model read a past design decision? Tested on Gemma 4 E2B on one laptop. Agents start each session blind to what the last one decided. With the line: read the decision file 18 of 18 times. Without: 0 of 18. But the task was too hard for th
Images from this run
- Files changed
- 7
- Elapsed
- 1h 46m
- Typed turns
- 4
Activity along the route
Session traceRecorded details and sources
Recorded tokens
Text the model read and wrote.
- 33.5M
Tool calls
Actions the agent took.
- 122
Elapsed
Time between the first and last captured event, including waiting.
- 1h 46m
About these measurements
Recorded tokens. 33,406,835 input + 130,558 output. Cached input is included; only calls with recorded usage are counted.
Model. Model names from the capture.
You typed. Recorded human messages, not words or keystrokes.
Tool calls. Actions the agent took.
Elapsed. Time between the first and last captured event, including waiting.
Est. dollars. No stored per-model usage and price provenance. No dollar estimate is shown.
Tool-active time · proxy. No stored tool-call bin provenance. No active-time proxy is shown.
Models and usage
- Platform / harness
- Claude Code
- Recorded models
- Claude Opus 5.5
- Recorded tokens
- 33,537,393
Where these numbers come from
Model names and usage come from the imported capture. STRIVE has not verified them with the provider. Source: claude-message-usage. Recorded as claude-opus-5-5.
33,406,835 input + 130,558 output tokens. Cached input and reasoning tokens are already included, and are not added again. Only calls with recorded usage are included; missing usage remains unknown. This is not a bill or a measure of wasted work.
Recorded activity
Human messages: 4
Messages typed or queued by a person; excludes tool results and injected context.
Tool requests: 122
Requests the agent made to tools. A request does not prove a successful action.
Files present after writes: 7
Files the capture says were written and existed when checked. This does not prove they work.
Source and limits
Imported · client-reported. The capture supplied these numbers. STRIVE has not independently verified the source or outcome.
Recorded session span; may include idle time. Not human effort. Session date supplied by the capture.
The measurement reference identifies this submitted capture. A hash is not proof the counters are true.
Photos and descriptions are chosen by the author. They do not attest to when a session happened or what succeeded.