Comments
Login to post comments.
wenaus · Apr 22 16:05
· re https://github.com/eic/epic/pull/1081
added to sysprompt:
The lesson fits as a new Procedure step 5, in the same terse-bold-lead voice as the existing four:
5. When a comment looks wrong, check history before theorizing. A comment that disagrees with surrounding code is more often stale than wrong-intent. Use GitHub MCP commit-history operations (e.g. list_commits filtered by path, or get_file_contents at older refs) on the code lines the comment describes — not on the comment — to find what changed and when. Do not draw conclusions about the file's purpose or supported scope from a comment whose underlying code has moved underneath it.
Worded in terms of GitHub MCP rather than git blame because the agent has no shell — git blame is not actually one of its tools, even though that's the mental model.
kalinkin · Apr 22 15:39
· re https://github.com/eic/epic/pull/1081
· on codoc-pr-review
The issue about `connected` is identified correctly, it was already seen from commit history. One could reasonably conclude that it got set again to true during debugging and got committed by accident.
The second issue, I believe, is real. Due to the small crossing angle it's a small effect. It could have been explained more clearly if agent stated that the extrusion is placed relative to (0,0,0) of the Union's reference frame, which has an offset relative to the global coordinates and so in those coordinates the IP is at (0, 0, -(zmax+zmin)/2) instead.
The last remark identified an outdated comment correctly, but "the file's headline purpose is the new tracking_volume for TV, not the legacy TC constants" is not correct – both "TC" and "TV" modes of operation are meant to be supported by the file and the comments in <definition> explain so. A helpful context could have been obtained by looking at `git blame main -- compact/definitions.xml`, from which one can find that the asymmetrical region was introduced as a two line change to zmin/zmax definitions in https://github.com/eic/epic/commit/7f224217d8a247ba8255f70d2cbf1602b3dbb0c4. That happened without updating the currently outdated comment.
wenaus · Apr 6 21:00
· re Summarize IRT-based PID reconstruction for DRICH and PFRICH implemented across E
· on codoc - OPUS v9
Finally complete
wenaus · Apr 6 18:50
· re Summarize IRT-based PID reconstruction for DRICH and PFRICH implemented across E
· on codoc - OPUS
Will rerun after possible fix of 'permission issue'
wenaus · Apr 6 18:23
· re Summarize the calorimeter algorithms in EICrecon.
· on codoc - GEMMA4
No improvement from deep think mode
wenaus · Apr 6 17:01
· re Summarize the calorimeter algorithms in EICrecon.
· on codoc - GEMMA4
Improved the sysprompt to leverage MCP tools but not significantly improved apart from LXR links
wenaus · Apr 6 16:49
· re Summarize IRT-based PID reconstruction for DRICH and PFRICH implemented across E
· on codoc - OPUS
Interesting. It overwrote most of its report with an addendum. I will tweak the sysprompt and try it again.
wenaus · Apr 6 12:57
· re Summarize the calorimeter algorithms in EICrecon.
· on codoc - GEMMA4
Gemma 4 added with two options:
Gemma4 - gemma4:31b (google/gemma-4-31B-it, 31.3B dense, Q4_K_M, 256K context, 19.9 GB)
Gemma4-fast - google/gemma-4-E4B-it, 8B total / ~4.5B effective, Q4_K_M, 128K context, 9.6 GB
Result of Gemma4 on the calo algorithm summary is (as expected) underwhelming compared to sonnet or opus.
Gemma4 runs on Torre's desktop Mac Studio, so free.
wenaus · Apr 2 16:39
· re Summarize the tracking algorithms in EICrecon.
· on codoc-SONNET
Comment on the doc
wenaus · Apr 2 11:43
· re Summarize the tracking algorithms in EICrecon.
Important part of this app: assessment by experts, discussion of the docs generated by the models. Discussions sit here at the prompt level, can discuss performance of different prompt runs with different models and settings.