ObneyAI/orc / orc-skills/orc-decision-models
orc-decision-models
VerifiedBuild bounded semantic decisions in ORC with llm-decision, boolean llm-condition gates, and registered native decision providers such as Jev. Use for described enum routing, finite tool arguments, applicability and completion checks, or decision-model integration and evidence.
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 9.1 KB |
| agents/openai.yaml | config | 286 B |
| references/contracts-and-observability.md | markdown | 5.6 KB |
| references/current-proofs-and-limits.md | markdown | 3.7 KB |
| references/jev.md | markdown | 5.6 KB |
| references/native-provider-selection.md | markdown | 11.9 KB |
| references/tree-patterns.md | markdown | 6.8 KB |
references/current-proofs-and-limits.md
View on GitHub# Current decision proofs and limits
At ORC `1851212396cced2fc4533fda30b824bd4b691390`, source regressions passed 131 tests / 453 assertions. Independent decision experiments passed 13 / 91 through public workflow builds/ticks, actual Grain and SQLite, with fixture responses only at the provider effect boundary. A separate RED public-build contract preserves the incorrectly rejected boolean `:abstain false` configuration.
Six root-owned real native/conversational decision calls exercised supported routing and predicate paths. Exact requests, returned identities, recorded finite answers and public projections were checked. Successful examples establish live integration and these observed answers, not accuracy or calibrated confidence over other tasks. No universal temperature or confidence threshold is prescribed.
Fresh-context router authors using revised guidance and the snapshotted older guidance produced independently passing deterministic projects: revised 12 tests / 79 assertions; older 12 / 87. Both could consult the same current source, so those counts do not establish a causal skill advantage. Author snapshots remain separate from later evidence-driven skill edits.
## Three narrow limits
- Boolean `:abstain false` fails public build-time membership validation even though false is offered. Valid false decisions without that option still succeed. Use described enum clarify/defer for an explicit abstention route; preserve this engine repair obligation.
- Conversational answers and Noul supply no Choice confidence. A floor abstains on absent confidence, not on a derived probability-certainty measure. Choice uses declared reporting resolution, so rounded ties may be valid routing evidence. The stricter unique-maximum score-only judge policy is separate.
- Native decision completions preserve the provider-resolved model, but ordinary conversational leaf completion events currently retain the configured/requested node model instead. Capture actual prediction/HTTP response model separately and keep both fields labeled. Do not call the event field an exact resolved identity for ordinary generation.
Shared budget admission prevents further registered model calls across delegated descendants. The cancellation lifecycle still has a reproduced trace/publication defect: an active child can be cancelled without its canonical trace, and interrupting its worker can strand root cancellation after durable append but before live delivery. Read [tool proof limits](../../orc-tools/references/current-proofs-and-limits.md) before claiming prompt family cancellation or complete family monitoring. Native Score and multilabel policy remain custom integrations; judge scheduling fixes are separate work.
## Decision stage and live eligibility
A fresh parallel consumer's native eligibility gate selected clarify before invoking tools for a replenishment request, producing an honest failure with no domain effect. The same unchanged tree completed when its runtime request explicitly described reading stock/price, computing the observed shortfall and keeping the task read-only. These are separate observed outcomes, not a calibration dataset or proof of the model's internal reason.
Describe the gate's stage precisely: whether a task is supported by the offered capabilities is different from whether sufficient observations have already been gathered. A gate before retrieval should evaluate whether the planned reads can answer the task; a post-retrieval check should evaluate the actual evidence. Exercise representative selected/defer paths through the real provider instead of treating a forced fixture answer as semantic proof. Keep an honest defer outcome; do not retry an unchanged decision until a desired choice appears.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| slack-gif-creatoranthropics/skillsKnowledge and utilities for creating animated GIFs optimized for Slack. Provides constraints, validation tools, and animation concepts. Use when users request animated GIFs for Slack like "make me a GIF of X doing Y for Slack." | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| ponytailDietrichGebert/ponytailLazy senior dev mode for any coding task: the smallest change that fully solves it, and a reply a busy human understands in one read. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailHarvest every shortcut: comment into one debt ledger, so deferrals get tracked instead of forgotten. One-shot report. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailList every `shortcut:` comment (and older `ponytail:` ones) as a debt ledger. One-shot report, changes nothing. Use for "ponytail debt", "what did ponytail defer", "list the shortcuts", /ponytail-debt. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |