Review a Kapso build: the Solution and the record

A build returns a Solution, and writes the same thing to <output>/.kapso/solution.json every time it returns: finished, paused for you, or out of budget. From any process, later or elsewhere:

from kapso import Solution

solution = Solution.open("campaign")
print(solution.explain())
print(solution.diff())

Before the build has returned there is no solution yet; open() says so and names kapso watch.

What the Solution holds

Field	What it is
`brief`	The brief the build ran, as a `Brief`
`campaign`	The campaign directory, checked out at the delivered branch
`branch`, `commit`	The delivered candidate; both `None` when no candidate succeeded
`base_commit`	The seed commit, `diff()`’s base
`score`	The delivered candidate’s measurement under the registered evaluator
`goal_achieved`	The judge’s verdict that the objective was met
`stopped_reason`, `stop_detail`	`goal_achieved`, `max_iterations`, `budget_exhausted` (with the detail: `time_budget`, `cost_budget`, `attempt_would_not_fit`, `finishing_cut`) or `waiting_for_user`
`experiments`, `experiments_total`	Experiments this run, and across every resume
`cost_usd`	Model spend
`requests`	What you must answer, when `stopped_reason` is `waiting_for_user`
`feedback`	The last judgment, in full
`evaluator`	The evaluator version the score was measured under; `None` unless the evaluation maintainer is configured (see Evaluation)
`harness`	The harness the build ran with: `name`, `version`, `hash`, `source`; `None` without one
`verification`	The harness’s verdict on the delivered candidate: `verdict`, `summary`, `findings`; `None` when no verification ran
`rejected_candidates`	Candidates rejected for touching the evaluator, or voided by the harness’s verification
`callback_metrics`, `callback_primary_metric`	Your own evaluator’s numbers, when you passed an `iteration_evaluator`
`experience`	What the build was served: one entry per loaded source with its `reference`, `kind`, `resolved` location, pinned `head`, `provenance` and `scope`; empty when nothing was loaded
`session_failures`	The sessions’ runs that ended `error`, `timeout` or `refused`: each run’s `label`, `module`, `host`, `exit` and `error`; empty when every run ended well

explain() renders these as a block; diff() is git diff from the seed to the delivered commit, empty when nothing succeeded.

Solution for:    Raise the churn model's out-of-sample ROC AUC above 0.85
Campaign:        /home/me/churn/campaign
Candidate:       generic_exp_3 @ 4f1c0e2a…
Score:           0.8614
Goal achieved:   True
Stopped:         goal_achieved
Experiments:     3 this run, 3 in all
Cost:            $1.934

Feedback:
The AUC of 0.861 on the held-out rows clears the 0.85 target …

A build on a harness adds two rows before the feedback: Harness: with the name, version and hash, and Verified: with the verdict and its summary. A build whose sessions failed adds Sessions failed: after the experiments line, each kind with its count and the error’s first line (implementation on codex ×2 (Quota exceeded …)).

Three ways a build ends

The objective was met. goal_achieved is true, stopped_reason is goal_achieved. Deploy it, or learn from it.
The budget ran out. stopped_reason is max_iterations or budget_exhausted; branch is the best candidate so far and score its measurement. Resume with a bigger budget, or change the target. See Resume and budgets.
The campaign is waiting on you. stopped_reason is waiting_for_user and requests says what for. Answer through the inbox; the campaign continues the session that asked.

When branch is None, no candidate produced a valid measurement; the experiment branches are still in the directory, and kapso trace campaign shows what each one did.

The search, branch by branch

git -C campaign branch         # generic_exp_0, generic_exp_1, …, main
git -C campaign log --oneline generic_exp_2
kapso trace campaign           # each attempt: its sessions and runs, its score, its verdict, the cards it read

Every experiment stays checked in; the delivered one is checked out. Read the record walks the trace section by section.

What next

software = kapso.deploy(solution)      # running software; see Deploy
lesson = kapso.learn(solution)         # bank what the campaign taught; see Learn

kapso deploy campaign --target local

Resume and budgets

Give a stopped build more

Deploy

From a Solution to running software

Learn

What a finished campaign teaches

Python API

Every field, typed

Related pages: Resume and budgets · Deploy · Learn · Python API Kapso is an open-source framework by Leeroo that builds software toward measurable goals through experiment campaigns. Source code: github.com/Leeroo-AI/kapso · Install: pip install leeroo-kapso · Every page as plain text: llms.txt.

Start

Build

Harnesses

Experience

Deploy

Learn

How it works

Benchmarks

Review a Kapso build: the Solution and the record

What the Solution holds

Three ways a build ends

The search, branch by branch

What next

Resume and budgets

Deploy

Learn

Python API

​What the Solution holds

​Three ways a build ends

​The search, branch by branch

​What next

​Related

Resume and budgets

Deploy

Learn

Python API

What the Solution holds

Three ways a build ends

The search, branch by branch

What next

Related