ObneyAI/orc / orc-skills/orc-model-sio
orc-model-sio
VerifiedConfigure ORC models and structured I/O, inspect SIO or native decision-client requests, and diagnose provider or parsing failures. Use when model choice or what was sent to a model matters.
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 20.2 KB |
| references/complex-model-portability.md | markdown | 8.8 KB |
| references/configured-model-fallback.md | markdown | 8.9 KB |
| references/contract-preflight.md | markdown | 2.9 KB |
| references/input-field-guide.md | markdown | 4.4 KB |
| references/live-findings.md | markdown | 11.6 KB |
| references/model-portability.md | markdown | 30.9 KB |
| references/paired-function-profiles.md | markdown | 5.1 KB |
| references/portability-proof-boundaries.md | markdown | 6.1 KB |
| references/reasoning-and-completion-budgets.md | markdown | 11.7 KB |
| references/retained-fallback-profiles.md | markdown | 9.8 KB |
| references/scoped-model-alternatives.md | markdown | 11.2 KB |
| references/structured-contract-repairs.md | markdown | 22.3 KB |
| references/structured-output-boundaries.md | markdown | 6.9 KB |
references/portability-proof-boundaries.md
View on GitHub# Structured validity and source fidelity
Use this when a model change yields valid but surprising records, optional fields, generated specialist output or a source-supported conclusion. These are evidence boundaries, not model rankings or a new per-model parsing recipe.
## Locate the first divergence
Compare the original typed value and described contract, the actual model-visible input view, the adapter's serialized request, the raw response, the unchanged paired decoder and canonical writes. Keep the actual node start, item context and physical attempt attached. A prime preview and a specialist's full packet are different input scopes. Count each authored description within its actual field or inherited output scope; finding its words elsewhere in a prompt does not prove that field retained its meaning.
A strict schema defines allowed shape; actual remote enforcement depends on the endpoint, and the paired decoder must still validate the response. It cannot require the correct source record, source order, interpretation, relationship or optional state when multiple states are permitted. For an optional nullable comment, absent, present-null and present-text are three valid states. If the source contains explicit null and the raw answer chooses the absent branch, the decoder can faithfully produce an incorrect source copy. That is distinct from the input encoder dropping null or the decoder changing a present branch to absence. Preserve the original failure before changing anything.
On the inspected strict codec, optional nullable output entries use an explicit presence branch; the input view preserves native omission and bare null. Input/output aliases for typed fixed keys can also differ. These representations preserve identity but are not textually symmetric. Use the schema-derived input plan and paired output decoder; do not strip wrappers, infer native key types from their spelling, stringify nested objects or add model-specific cleanup. An alternative input presentation remains an experimental hypothesis until a controlled fixed-leaf comparison demonstrates value and compatibility.
## Choose the right next action
- If the serialized request omitted a declared value, description or supported option, repair that boundary and retain a public regression witness. A prompt patch is not evidence that the boundary was fixed.
- If raw output violates the actual wire schema, inspect endpoint admission, effective output mode, finish reason and completion headroom before changing the authored contract. Retain incomplete responses; raising allowance is a recorded configuration change, not a parsing repair.
- If raw output is valid but contradicts supplied records, keep the source contract. For exact copies, code can check complete native record equality, unique source IDs, original order and presence. Semantic support needs an appropriate assessment. Feed a concrete discrepancy to the actual owner for a bounded revision rather than editing its answer.
- If a model is refused before generation, report admission or capacity separately from content quality. Preserve privacy/capability policy. A paid sibling or fallback response does not establish that the original endpoint worked.
A schema-valid generated `code` string does not prove correct ORC API or tool use. Inspect the executed program, available symbols, tool argument/result contracts and actual feedback. Undefined functions, undeclared writes or an invalid emitted node are execution/API failures; keep them separate from a correctly executed operation whose valid answer misinterprets source content. Diagnose missing or contradictory API guidance before attributing those failures to model capability.
A mandatory record check and a semantic judge are complementary. A successful first answer does not exercise correction, and a judge's agreement does not prove correctness. See [judge calibration](../../orc-judges/references/rubrics-and-calibration.md) before using those assessments for GEPA.
## Preserve comparison scope
Hold task, native sources, contracts, instructions and decoding constant when comparing model configuration. Prove the actual request difference rather than trusting a DSL diff. A verified request differing only at model still does not hold remote provider defaults, reasoning or routing constant. For meaning experiments, compare the fixed extractor separately from the adaptive trajectory: changed iteration limits, provider, generated code or completion allowance prevent a whole-trajectory description-only claim. Keep failed comparators and unavailable cells in the ledger.
Controlled inline/emitted researcher fixtures establish those constrained interfaces. They do not prove autonomous architecture choice, general-agent completion or broad superiority over an ordinary loop. Test integrated authored behaviors and fresh-agent construction separately.
## Evidence status
The campaign's immutable source manifest is 233ef7b8e3306974e434dec93ad6743d9e5a144f798a2d1aec75b3c119e7f690. Original complex cases included valid structured answers with source-fidelity defects; later ordinary quality cases preserved the same kinds of null, absent, typed-key, false and zero values. Thus neither universal codec failure nor universal model reliability follows. Fixed whole/minor extractors demonstrated meaning transmission while adaptive specialist outcomes remained mixed.
This is local, unreleased campaign evidence. Recheck the consuming source and pins. The original matrix, paired-meaning observations, per-cell canonical/input/output audits and independent reviews remain the provenance; no score or aggregate percentage here is a general accuracy estimate. Runtime accounting and failure propagation have separate [versioned proof boundaries](../../orc-runtime-bootstrap/references/durable-family-budgets.md).
## Paired response functions and endpoint controls
The private experimental paired function profile has additional [request and evidence boundaries](paired-function-profiles.md). Its live common-shape probes do not establish remaining researcher/fallback roles or released support. Keep the earlier candidate and forward-test receipts distinct from this new, untested skill revision.
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 |