feat: MNCS-native RAVEL core — five verified MNCS-language modules - #14
Conversation
…rvice integration Reconstructs RAVEL's reasoning core in the MNCS language (Source Profile 0.5) as one self-contained module per concern: - ravel.core: status dominance lattice over the full finite domain, evidence combination, gate-based disposition, advisory confidence; - ravel.loop: hypothesis/prediction/attribution with commit eligibility requiring clean gates AND confirmed prediction; negative retention; - ravel.checkpoint: immutable checkpoints, parent-bound candidates, structural commit/rollback isolation, orphan refusal; - ravel.memory: promotion lifecycle, transfer tests, counterexample immutability under later success, protected strategy selection; - ravel.task: task contexts bound to Language Service snapshot identities with typed refusals for stale/authority/budget drift. All modules execute through the language experiment flow on research bytecode with layered body/SSA/backend agreement PASS and expected values met (17+22+5+11+5 cases); overall statuses are honest (UNKNOWN where checked-arithmetic exact-cost obligations remain unresolved). Corpora carry nested record expectations; generators included. Adds Forge workflow 'mncs-experiments' (tools/ravel_mncs_check.py) verifying every module end to end against the sibling mncs-language checkout, plus architecture and evidence documentation including the backend capability matrix, legacy differential mapping table, and language-friction log. Development used mncs-language-service resident semantics throughout; its Phase 4 candidate analysis was driven by this conversion.
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 838418748b
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| } | ||
| ) | ||
| ) | ||
| return 0 |
There was a problem hiding this comment.
When the sibling checkout is absent or its CLI build fails, the registered mncs-experiments workflow prints BLOCKED but exits successfully; this was reproduced in the inspected environment. Any Forge or CI caller relying on the command's exit status therefore treats a run that executed none of the five experiments as successful, so this path should return a nonzero status.
Useful? React with 👍 / 👎.
| validations = [v.get("judgement") for v in document.get("translation_validations", [])] | ||
| status = document.get("status") | ||
| ok = met == total and all(v == "PASS" for v in validations) |
There was a problem hiding this comment.
If the CLI omits translation_validations—for example because an incompatible version stops producing the configured layered-agreement validator—validations is empty and all(...) evaluates to true. The workflow can consequently report PASS without the body/SSA/backend agreement that it claims to require; verify that the expected validator is present as well as that every returned judgement is PASS.
Useful? React with 👍 / 👎.
| if principle.id == strategy.principle_id { | ||
| if task_domain == strategy.scope_domain { | ||
| return TransferOutcome.TRANSFERS; |
There was a problem hiding this comment.
When a strategy references a promotable principle by ID but supplies a different scope_domain, a task in the strategy's scope is reported as TRANSFERS even though it is outside principle.scope_domain. Because the records are independently supplied, this allows a strategy to relabel a principle for another domain; transfer eligibility should also bind the strategy and task to the principle's declared scope.
Useful? React with 👍 / 👎.
| fn affordable(context: TaskContext, requested_steps: i64) -> (result: bool) { | ||
| return requested_steps <= context.budget_steps; |
There was a problem hiding this comment.
Because both fields are signed i64, any negative requested_steps satisfies this comparison against a normal nonnegative budget. plan_request() then emits ISSUE_REQUEST with a negative max_steps, allowing malformed requests past the stated budget boundary and potentially causing wrapping or rejection at an unsigned host boundary; affordability should require a nonnegative request as well.
Useful? React with 👍 / 👎.
| if primary_in_scope { | ||
| if primary_clean { | ||
| return Selection.PRIMARY_STRATEGY; |
There was a problem hiding this comment.
When both strategies are clean and in scope, this immediate return always selects the primary even if fallback.rank is higher; in fact, neither rank field is read anywhere in select(). This makes the advertised ranking and the comment that equal ranks prefer primary ineffective for non-ties, so the clean in-scope candidates should be compared before selecting.
Useful? React with 👍 / 👎.
1 participant
Summary
First full conversion cycle of RAVEL's reasoning core into the MNCS language itself, under
mncs/:mncs-experimentsverifying all modules against the sibling mncs-language checkout;Ecosystem changes driven by this conversion (merged upstream)
Non-claims
Bounded local development evidence only. No independent evaluation, promotion, conformance, or production claims. No executable differential harness against legacy RAVEL yet (mapped contracts documented instead).