ReCoAlign — closed research line and frozen negative-evidence resource

ReCoAlign is scientifically closed as of PIVOT_EXP_A3P.

Program decision: TERMINATE_CURRENT_PROGRAM
Current method line: RETIRED
New experiments: not authorized
Model development: not authorized
Claim-bearing paper writing: not authorized
Archive and release: authorized

The repository preserves a complete hypothesis lifecycle rather than a success narrative. The original Structured Reasoning Interface Gap was falsified. Its successor, Semantic–Structural Integration Failure, remained unresolved and was not identified as a single causal mechanism. M1 and M2 both failed the preregistered semantic-rescue gates, so no semantic manipulation and no method target were selected.

Frozen evidence through PIVOT_EXP_A3P

Study	Frozen result	Governed interpretation
EXP001	Graph < Caption; NO-GO	Graph is not a privileged interface over information-equivalent text.
EXP002	Correct > corrupt/partial/random; GO	Correct supplied relations matter in the frozen LLaVA-1.5 scope only.
EXP003	INCONCLUSIVE	Token-length integrity failed; no OOD claim is valid.
EXP004	Low SAS; NO-GO	High semantic availability was not supported; no all-layer absence claim follows.
PIVOT_EXP_A	NO-GO	Multiple mechanisms were supported; causal primacy was unresolved.
PIVOT_EXP_A2	INCONCLUSIVE	Semantic sufficiency and parsing integrity failed.
PIVOT_EXP_A3	RUNTIME_BLOCKED_PREINFERENCE	The full-field free-generation contract failed development smoke.
PIVOT_EXP_A3R	SECONDARY_CONTRACT_RUNTIME_FAILURE	127/160 valid; 33 entity-ID continuations; instrument retired.
PIVOT_EXP_A3P	SEMANTIC_MANIPULATION_FAILURE	27,000/27,000 valid forced-choice predictions; M1/M2 failed Gates C, D, E.

PIVOT_EXP_A3P measurement integrity passed for the scorer and prediction inventory only. It did not verify semantic comprehension. Its frozen assets are under research/construct_validity/PIVOT_EXP_A3P/results/.

Final pivot gate

PH005 Arbitrary Referential Binding Limitation, PH006 Serialization-Conditioned Evidence Usability, and PH007 Cross-Modal Evidence Competition were evaluated against fixed evidence, identifiability, novelty, feasibility, and post-hoc-risk gates. None passed every gate. STOP was therefore selected. No final-pivot preregistration or power analysis was created.

The adjudication is in reports/final_research_line_adjudication.md, the literature audit in research/final_pivot_literature_audit.md, and the machine-readable claim state in docs/evidence_map.yaml.

Research boundary

• Frozen predictions, metrics, figures, decisions, and manifests are immutable.
• EXP002 supports relation-evidence sensitivity only in the frozen LLaVA-1.5-7B setting.
• Free-generation instruments are retired; no prompt or parser repair is authorized.
• Training, reasoning-interface, alignment, comprehensive-matrix, and ablation modules are retained for provenance but are scientifically inactive.
• Retrieval is a capability-preservation control, not the new scientific contribution; benchmark infrastructure is likewise provenance only.
• A future independent decision would be required to authorize any experiment, backbone replication, model work, or claim-bearing manuscript.

See INACTIVE_RESEARCH_BOUNDARY.md, docs/research_identity.md, and research/current_line_closeout.yaml.

Repository use

The authorized use is provenance inspection, reproducibility checking, negative-results release, internal technical reporting, and research-governance study. Historical commands remain available to validate already frozen records; they are not a roadmap and do not authorize inference.

python -m recoalign validate-research
python -m recoalign validate-paper-package
pytest
ruff check .

Repository map

docs/                         evidence map, claim boundary, lifecycle, and closed roadmap
research/                     frozen protocols, results, closeout, literature audit, archives
reports/                      readiness and final adjudication reports
reports/paper_evidence/       immutable EXP001–EXP004 claim-evidence package
training/                     inactive historical implementation
models/reasoning_interface/   inactive historical implementation
models/alignment/             inactive historical implementation
configs/training/             inactive historical configs
configs/benchmarks/           comprehensive matrix retained but inactive
configs/ablations/            inactive historical configs
src/recoalign/                validation and provenance infrastructure
tests/                        governance and engineering regression checks

No “train ReCoAlign,” “run comprehensive benchmark,” or “paper submission” milestone is pending. Those former roadmap entries are retired.