Skip to content

Add Roadmap and Kill criteria to the README - #39

Merged
kishuxz merged 1 commit into
mainfrom
roadmap-kill-criteria
Aug 15, 2026
Merged

kishuxz merged 1 commit into
mainfrom
roadmap-kill-criteria

Conversation

@kishuxz

@kishuxz kishuxz commented Aug 15, 2026

Copy link
Copy Markdown
Owner

Two short sections between Limitations and License. Nothing else added, nothing trimmed.

Roadmap

Labelled as intent, not work done, so a reader cannot mistake v2 and v3 for things that exist.

  • v1, current. 200 threads, one annotator, two configurations scored on the dev split.
  • v2. Inter-annotator agreement measured and kappa published, especially on superseded. A submission harness so another extractor can be scored against the same corpus. A test split held out by mechanism rather than by discipline.
  • v3. Calibration, meaning whether stated confidence predicts correctness, since that is the gate on any system acting without asking. Longitudinal loops that live across days and threads rather than within one. Resolution that arrives on a different channel from the commitment.

Kill criteria

Stated plainly as conditions for abandoning parts of the benchmark rather than defending them.

  • Agreement on superseded below roughly 0.6 means the label is ill-defined, and the headline metric needs redesigning rather than more data, because more data labeled against an unclear rule buys nothing.
  • PII redaction costing less than a couple of points once the run completes means the tension this was built around does not exist, and the framing narrows to supersession alone.

Nothing else changed

Section Before After
Corpus separability 82 lines 82 lines
Limitations 66 lines 66 lines
Cost-weighted error table 5 weights 5 weights

README goes from 308 lines to 339, so both sections together cost 31 lines.

pnpm check passes with 229 tests. pnpm validate passes with 200 threads, 273 loops and 510 spans all resolving. No em or en dashes introduced.

🤖 Generated with Claude Code

Two short sections between Limitations and License.

Roadmap is labelled as intent rather than work done, so a reader cannot
mistake v2 and v3 for things that exist. v1 states what is actually
here. v2 is inter-annotator agreement with kappa published, a submission
harness, and a test split held out by mechanism rather than by
discipline. v3 is calibration, longitudinal loops, and resolution
arriving on a different channel from the commitment.

Kill criteria states plainly when to abandon rather than defend. If
agreement on superseded falls below roughly 0.6 the label is
ill-defined, and the headline metric needs redesigning rather than more
data. If PII redaction turns out to cost less than a couple of points
once the run completes, the tension the benchmark was built around does
not exist and the framing narrows to supersession alone.

Nothing else was added, and Limitations and Corpus separability are
untouched at 66 and 82 lines with the cost table still at five weights.

README goes from 308 lines to 339.
@vercel

vercel Bot commented Aug 15, 2026 •

Copy link
Copy Markdown

The latest updates on your projects. Learn more about Vercel for GitHub.

Project Deployment Actions Updated (UTC)
openloop-bench Ready Ready Preview Aug 15, 2026 8:42am

@kishuxz
kishuxz merged commit 58c12a1 into main Aug 15, 2026
4 checks passed

This branch was successfully deployed

1 active deployment
Preview — 82876dfa Deployed Aug 15, 2026 by vercel[bot]
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant