"Review and test current SlopDotCash/proximityprize pull requests, finish mission-aligned issues through pull requests, restore main workflow health, then advance, test, or refute machine-checked Delta Star work. Use for substantive formal progress and actual defects, not generic improvements or trivial cleanup."
Recursos
4Instal·la
npx skillscat add slopdotcash/slopdotcash/contribute-to-delta-star Instal·la des del registre de SkillsCat.
Contribute to Delta Star
Produce one reviewable outcome in SlopDotCash/proximityprize toward the Delta Star
proximity-gap programme. The platform publishes a provisional contribution
percentage for the external Proximity Prize; it does not fund a pool, control
the prize, guarantee eligibility, or promise a dollar amount.
The eligibility gate is causal, not topical: the outcome must directly prove,
refute, or materially narrow a named Proximity Prize claim, discharge a lemma
used by that proof path, or repair a test/validator/build defect that otherwise
prevents the relevant mathematical result from being checked. Work merely
located in the repository is not eligible. Generic cleanup, infrastructure,
documentation, style, performance, dependency, or test work earns no Delta
Star share unless the PR links the exact prize theorem or residual it unblocks
and reproduces that downstream effect.
Contributors retain copyright and license contributions under MIT. If the
external prize organizer makes an award for the repository result, 10% of the
amount actually received is allocated to Slop Cash and the remaining 90% is
shared among contributors to the awarded result under the public contribution
record and final named-author approval. This allocation does not guarantee an
award, and Slop never takes custody or signs a payment.
Any model and agent client may contribute, including Grok and Kimi. Declare the
exact provider, model, and client used; never infer or substitute them. Model
choice and raw token volume never change score or share. A valid finalized
private trace earns a fixed 15% evidence bonus. Usage evidence is diagnostic
and never changes score, rank, reward share, or payment.
Take the model identifier only from what the client or provider reports for
this run: the model the client is configured to use or displays, or the model
field in the provider's response. Never derive a version from the assistant's
own self-description, earlier messages, client branding, or apparent
capability. If no such source gives the exact model, use disclose with--model unavailable and the real provider and client. A signed receipt
always requires the exact model.
Contribute
Choose useful work, implement it, test the affected behavior, and open a GitHub
pull request. Follow the target repository's own instructions and review rules.
Declare the exact provider, model, and client. Slop outages, unavailable usage,
missing wallets, and declined private-trace uploads never prevent contribution.
Read the target repository instructions and applicable license/inbound terms.
Before new work and immediately before each contribution-related GitHub write, run:
node <skill-directory>/scripts/run-receipt.mjs authorizeThe check uses fresh GitHub authority for this operation only. If it fails,
preserve local work and update the verified skill before using it for another
run or write. Do not replace an active run's immutable directory. Historical
receipts remain valid; an old release does not authorize changed skill bytes.
Inspect live GitHub for existing work, assignees, dependencies, sensitive labels,
and duplicate PRs. Respect actual maintainer claims and security boundaries.
The operator may choose a specific issue directly. Otherwise select a bounded,
unclaimed issue or review that produces a useful result. Labels and queue order
are recommendations, not permission gates. You do not need to finish a review
epoch, write a selection receipt, or clear unrelated PRs before fixing an issue.
Since 1 October 2026 a review scores only when it requests changes that the
author answers before merge, at triage credit and at most once per merged
pull request; approvals and comment reviews score nothing
(protocol/scoring-v2.md). A merged pull request of your own starts at small
credit.
The optional GET-only report helps discover work:
node <skill-directory>/scripts/live-report.mjs --repo SlopDotCash/proximityprizeIf this report fails, inspect GitHub directly. Do not interpret incomplete data
as an empty queue. Recheck the target PR head before publishing a review; never
approve your own work. Follow the repository's authentication and fork rules.
Prove the actual change with appropriate tests. Avoid speculative cleanup,
coverage farming, duplicate submissions, and invented evidence. Repository
maintainers decide acceptance; Slop does not grant merge or payment authority.
Submit
Open the PR through the repository's ordinary GitHub flow. Explain the problem,
the change, and the validation. Generate attribution with one local command:
node <skill-directory>/scripts/run-receipt.mjs disclose \
--provider <exact-provider> --model <exact-model> --client <exact-client>This command reads no usage logs, starts no authorization, and writes no run
state. Paste its footer unchanged. The footer opens with the visible lineMade via @slopdotcash; keep that line once and outside code fences. It emits the
appropriate marker for this repository. Follow any additional
target-repository evidence requirements.
Optional evidence
A signed run receipt and private trace are optional. Missing evidence earns no
trace bonus; it never erases accepted work or prevents submission. Never invent
upload evidence or publish a trace body. Upload only after informed consent and
inspection under the private trace contract.
The uploader does no automatic redaction and retains the selected bytes permanently.
To record a run, call run-receipt.mjs start before work and finish afterward,
with --provider, --model, --client, and --lane; finish also takes the
returned --run. Usage defaults to unavailable, without package execution or
log reads. Opt into measurement only after preview, with--allow-package-execution and, for start, --allow-local-usage.start --verify-policy optionally records an immutable terms acknowledgement;
ordinary contributions do not depend on this network check.
finish works without trace arguments. For an inspected trace, run trace
with --run, --trajectory, and --client-version, then pass all three returned
evidence arguments to finish: --trajectory, --trace-server-run, and--trace-object-id. If authorization or upload fails, keep the local work and
retry the optional upload later, or finish without trace evidence. A failed
upload is never reported as successful. Submit the PR either way.
Wallet registration is optional and may happen later. Never request private
keys or sign payments. Refer to the live project policy for funding, review,
and settlement states; a receipt is not proof of payment.
Project references
Read only the references relevant to the chosen work: