Advance RAVEL 0.6 development foundations

README.md

@@ -6,7 +6,7 @@ RAVEL — the **Recursive Adaptive Vector Execution Lattice** — is an experime
 
 RAVEL operates beneath the technical authority of the Machine-Native Complexity Standard (MNCS) and the Machine-Native Complexity Development Standard (MNCDS). It is not intended to replace a language model, compiler, static analyzer, test framework, or the MNCS Forge. Its role is to decide what evidence should be gathered, what action should follow, and what experience should be retained for later use without redefining the governing status of that evidence.
 
-> **Project status:** RAVEL is research software. Historical RAVEL 0.4 and 0.5 results remain development `FAIL`; RAVEL 0.6 is preregistered but not finally evaluated or promoted. Formal MNCS/MNCDS conformance, independent attestation, protected custody, production safety, and general recursive self-improvement remain `UNKNOWN`.
+> **Project status:** RAVEL is research software. Historical RAVEL 0.4 and 0.5 results remain development `FAIL`; RAVEL 0.6 candidate-001 now has a reproducible development build and behavioral harness, but has not been evaluated, frozen, selected, or promoted. Formal MNCS/MNCDS conformance, independent attestation, protected custody, production safety, and general recursive self-improvement remain `UNKNOWN`.
 
 ## Place in the MNCS ecosystem
 
@@ -102,6 +102,12 @@ The recommended extraction is history-preserving rather than a manual file copy.
 5. Update MNCS to reference RAVEL as a sibling project without deleting historical evidence prematurely.
 6. Develop the Forge interface and MNCS/MNCDS-governed evidence and knowledge schemas before adding broad autonomy.
 
+The current 0.6 implementation status is recorded in
+[`ravel_versions/0.6/RAVEL_0_6_IMPLEMENTATION_STATUS.md`](ravel_versions/0.6/RAVEL_0_6_IMPLEMENTATION_STATUS.md).
+Optional provider/resource protocols in `src/ravel/providers.py` and
+`src/ravel/resources.py` record scoped execution observations without claiming
+algorithmic superiority.
+
 ## Non-goals
 
 RAVEL is not currently claiming:

docs/ARCHITECTURE.md

@@ -125,6 +125,28 @@ A bounded RAVEL episode should follow this sequence:
 11. **Extract provisional knowledge.** Create principles or strategies only when attribution and scope support them.
 12. **Test transfer.** Reuse remains restricted until separate contexts support the declared scope.
 
+## Implemented reusable foundations
+
+The repository now contains small, dependency-free foundations beneath the
+0.6 experiment:
+
+- `ravel.adaptation` performs copy-before-evaluate retention transactions and
+  emits stable rejection reason codes without treating raw observations as
+  authority.
+- `ravel.providers` defines a narrow Forge boundary with capability discovery,
+  immutable raw evidence, and fail-closed `UNKNOWN` receipts. It is an adapter,
+  not a Forge reimplementation.
+- `ravel.resources` separates future provider placement policy from the
+  preregistered 0.6 mechanism. Fake backends exercise CPU fallback, CUDA kernel
+  probes, VRAM headroom, sequential offload, and bounded OOM recovery.
+- `ravel.memory` retains versioned source identities, explicit scope contracts,
+  negative-aware full-text retrieval, append-only proposal lifecycle events,
+  atomic batches, and rebuildable relation projections.
+
+These modules are tested scaffolding and are not claims that RAVEL 0.6 has been
+evaluated, selected, independently evaluated, certified, promoted, or made
+production-safe.
+
 ## Separation of statuses
 
 RAVEL must not collapse different state spaces into one score.

docs/PROJECT_MAP.md

@@ -15,6 +15,11 @@ remain unchanged and are resolved by the versioned compatibility runners.
 | 0.5 | `ravel_versions/0.5/ravel_0_5.c` | `ravel_versions/0.5/RAVEL_0_5_CONTRACT.md` |
 | 0.6 | Candidate source is derived by `tools/ravel_0_6_seed_candidate.py` | `ravel_versions/0.6/RAVEL_0_6_SCOPE.md` and preregistration |
 
+Candidate-001 development builds are orchestrated by
+`tools/ravel_0_6_build.py`; generated source and binaries are temporary
+development material, not maintained source or final evidence. Behavioral
+facts are produced by `tools/ravel_0_6_behavioral_fixtures.py`.
+
 Build binaries are local outputs and are removed by `make clean`.
 
 ## Protocol, evidence, and identity
@@ -45,6 +50,9 @@ regenerating the frozen records.
 - [Architecture gaps](ARCHITECTURE_GAPS.md) records the early design gaps.
 - [`../tools/README.md`](../tools/README.md) documents evaluators, digest tools,
   mutation checks, runtime capture, and 0.6 candidate derivation.
+- `src/ravel/adaptation.py` provides the tested retention-transaction boundary.
+- `src/ravel/providers.py` and `src/ravel/resources.py` provide optional,
+  replaceable Forge/runtime protocols without heavy ML dependencies.
 - [`MIGRATION.md`](MIGRATION.md) records the standalone extraction provenance.
 
 ## Build and verification entry points

docs/PROVIDER_RUNTIME.md

@@ -0,0 +1,34 @@
+# Provider and resource runtime boundary
+
+RAVEL can request evidence from a Forge-compatible provider and retain a
+scoped execution experience, but Forge remains the evidence executor and MNCS /
+MNCDS remains authoritative. `src/ravel/providers.py` is intentionally small:
+it exposes capability discovery, a typed evidence request, immutable raw
+evidence, and a separate fail-closed receipt. Missing capabilities, provider
+crashes, malformed statuses, and identity mismatches become `UNKNOWN`.
+
+`src/ravel/resources.py` is an optional policy layer for future learned
+providers. It has no PyTorch, CUDA, Transformers, or model-weight dependency.
+It records observations such as:
+
+- `auto`, `cpu`, or `cuda` device policy;
+- no offload or sequential CPU offload;
+- actual CUDA-kernel and dtype probes supplied by the backend;
+- free VRAM, reserve/headroom, maximum VRAM budget, CPU availability, and RSS;
+- model parameter, workspace, and peak-module estimates;
+- placement reason, peak VRAM reservation, parameter residency, and timing or
+  failure observations supplied by the backend.
+
+Sequential CPU offload keeps primary parameters in system RAM between module
+executions. It trades VRAM for system RAM; it does not reduce total memory
+requirements. A CUDA out-of-memory signal permits at most one recovery decision,
+after which the backend either executes the selected fallback or fails closed.
+
+These measurements are scoped empirical experience, not evidence of
+algorithmic superiority. A future experience record should bind provider,
+model, hardware, task class, device, placement, dtype, budget, runtime, failure
+status, evidence quality, and verifier outcome before any strategy reuse.
+
+Standalone RAVEL tests use fake providers and resource backends. Forge is not a
+mandatory dependency for the core package, and lack of a GPU does not make the
+normal test suite fail.

docs/SEMANTIC_CONSOLIDATION.md

@@ -2,7 +2,12 @@
 
 ## Status
 
-This document specifies a development prototype. It does not establish MNCS or MNCDS conformance, protected custody, production safety, or validated recursive self-improvement.
+This document specifies a development prototype. The prototype now has named
+scope-compatibility contracts, versioned evidence/experience identity fields,
+deterministic source full-text retrieval, append-only proposal lifecycle events,
+atomic source batches, and rebuildable relation projections. It does not
+establish MNCS or MNCDS conformance, protected custody, production safety, or
+validated recursive self-improvement.
 
 ## Motivation
 
@@ -58,7 +63,9 @@ The source store remains authoritative. Consolidation proposals and layout plans
 
 ### 1. Candidate partitioning
 
-Records are first partitioned by memory class and exact declared scope. The prototype does not infer that two scopes are compatible. Future scope adapters may implement version-aware compatibility rules, but those rules must be explicit and testable.
+Records are first partitioned by memory class and a named scope-compatibility
+contract. The default contract remains exact scope equality; an alternate
+contract may be used only when its compatibility rule is explicit and tested.
 
 ### 2. Similarity grouping
 
@@ -159,11 +166,8 @@ A faster result is not sufficient if it hides negative evidence, increases false
 
 ## Next steps
 
-1. Bind records to the versioned RAVEL evidence and experience schemas.
-2. Add explicit scope-compatibility contracts instead of exact scope equality.
-3. Add full-text retrieval and a benchmark corpus before embeddings.
-4. Record query events in the experience store with privacy and retention controls.
-5. Implement proposal review, acceptance, challenge, and supersession records.
-6. Add graph projection rebuilds from the append-only source stream.
+1. Bind records to the complete versioned RAVEL evidence and experience schemas.
+2. Record query events in the experience store with privacy and retention controls.
+3. Add benchmark corpus measurements for retrieval quality, negative recall, and cost.
 7. Benchmark physical page, cache, and shard layouts on the planned local and distributed RAVEL environments.
 8. Add embedding-assisted candidate generation only after the deterministic baseline is measured.

ravel_versions/0.6/Makefile

@@ -0,0 +1,11 @@
+CC ?= cc
+CFLAGS ?= -std=c11 -O3 -Wall -Wextra -Werror -pedantic
+
+.PHONY: build behavioral-test
+
+build:
+	@set -e; out=$$(mktemp -d); python3 tools/ravel_0_6_build.py build --output-dir "$$out"; \
+	printf 'RAVEL 0.6 development build directory: %s\n' "$$out"
+
+behavioral-test:
+	python3 tools/ravel_0_6_behavioral_fixtures.py

ravel_versions/0.6/RAVEL_0_6_IMPLEMENTATION_STATUS.md

@@ -0,0 +1,36 @@
+# RAVEL 0.6 implementation status
+
+This is a development status record, not RAVEL 0.6 evaluation evidence.
+
+## Implemented and tested
+
+- **R6-01 foundation:** candidate `ravel-0.6-candidate-001` is generated from
+  the exact frozen 0.5 source identity. The temporary build record binds the
+  frozen source, generator, generated source, compiler/version/argv, selected
+  environment-key identities, worktree status, stdout, stderr, and exit
+  status. Generated source is explicitly development-only.
+- **R6-02 transaction foundation:** the Python adaptation surface evaluates
+  independent hard constraints over a copied checkpoint and restores the
+  preceding bytes on any rejection. Rejection codes cover objective epsilon,
+  retention floors, prediction degradation, transition support, topology,
+  replay, update, and compute budgets. This surface is tested with one
+  negative fixture per constraint; the generated C candidate has not yet
+  consumed this generic interface as a final evaluator.
+- **R6-03 behavioral fixtures:** a compiled harness reaches a goal only through
+  transition slot one and checks that a born child has only spawning-event
+  support. The candidate passes; frozen 0.5 and each reverted correction fail.
+  Outputs are integer facts plus a stable checksum.
+
+## Not yet implemented
+
+- R6-02 is not a completed 0.6 trial/evaluator pipeline. No development,
+  selection, or final material evaluation has been executed.
+- R6-04 has not been claimed: the large generated C mechanism has not yet been
+  split into independent mechanism, provider, planning, checkpoint, and
+  observation units.
+- R6-05 candidate ledgers and selection partition consumption are not present.
+- R6-06 external final custody/evaluation remains unavailable and `UNKNOWN`.
+
+The 0.6 candidate remains unfrozen, unselected, unpromoted, and unauthorized
+to alter evaluator identity, thresholds, partitions, evidence custody, or
+formal MNCS/MNCDS status.

ravel_versions/0.6/RAVEL_0_6_NEXT_STEPS.md

@@ -43,6 +43,17 @@ checking, and strict C11 compilation.
 These corrections produce development source only. Candidate 001 has not been
 integrated into the 0.6 evidence pipeline, selected, frozen, or evaluated.
 
+## Current bounded implementation status
+
+Candidate-001 is now integrated as a reproducible temporary development build
+by `tools/ravel_0_6_build.py`. A compiled behavioral harness verifies both
+corrections and fails against frozen 0.5 behavior and bounded reverted
+mutations. A separate Python transaction foundation tests all declared hard
+constraint classes with byte-identical rollback. These are implementation and
+mutation observations only; the C candidate is not yet a complete 0.6 trial
+evaluator, and no selection or final material has been consumed. See
+`RAVEL_0_6_IMPLEMENTATION_STATUS.md` for the exact boundary.
+
 ## Codex next steps
 
 Perform the following tasks in order. Treat each material post-evaluation change

ravel_versions/0.6/ravel-0.6-development-record.json

@@ -6,7 +6,10 @@
   "recorded_at": "2026-08-01T00:00:00Z",
   "record_class": "operator_controlled_preregistration_state",
   "candidate": {
-    "implementation_status": "NOT_IMPLEMENTED",
+    "implementation_status": "DEVELOPMENT_CANDIDATE_INTEGRATED",
+    "candidate_id": "ravel-0.6-candidate-001",
+    "generated_source_sha256": "b71faca6aff74be89139f8d0c9d9eadfc1485c1e81d12b6ccd945f98e8a81c9a",
+    "generator": "tools/ravel_0_6_build.py",
     "selected_candidate_id": null,
     "frozen": false,
     "manifest_identity": null

schemas/consolidation-proposal.schema.json

@@ -18,6 +18,7 @@
     "retrieval_keys",
     "clustering_confidence",
     "status",
+    "scope_contract_id",
     "limitations"
   ],
   "properties": {
@@ -68,6 +69,7 @@
       "maximum": 1
     },
     "status": {"const": "proposed"},
+    "scope_contract_id": {"type": "string", "minLength": 1},
     "limitations": {
       "type": "array",
       "minItems": 1,