OUTSEED / AI-assisted engineering
OUTSEED ChangeProof
Transformez une intention de changement en chemins d’impact et en dossier de preuves manquantes.
Nouveau workflow OUTSEED en aperçu. Aucune revendication de première mondiale ni de simulation réelle.
Un compte ingénieur suffit
Les MCP et ChangeProof sont réservés aux comptes ingénieurs connectés. L’approbation pour travailler sur les missions OUTSEED n’est pas nécessaire.
Vérification du compte…
How AI clients use ChangeProof
Ask your MCP-enabled AI client to collect permitted platform inventory and help organize an evidence-backed dependency graph. Review every inferred edge before passing it to change.rehearse. The tool analyzes only the supplied graph; it does not discover topology or predict hidden failures.
Use MIGRATE, RESTART, NETWORK_CHANGE, STORAGE_CHANGE or UPGRADE to select evidence gates. The output traces potentially affected dependents, reported failures, missing/old proofs and next evidence actions. Gate evaluation is deterministic; natural-language interpretation is performed by your chosen AI client, not an undisclosed OUTSEED model.
Evidence debt rather than a made-up confidence score
A PASS requires a non-empty reference and timezone-aware observation time inside the chosen evidence window (72 hours by default). Missing references, future timestamps and stale observations cannot become fresh proof. Supplied references are not independently authenticated.
Changed assets require action-specific baseline/backup/rollback/owner/application and technical checks. Dependent assets need owner/application evidence. Cluster assets also require a quorum review. Known FAIL is BLOCKED; incomplete proof is EVIDENCE_REQUIRED. Even EVIDENCE_GATES_MET_ON_SUPPLIED_GRAPH is not approval to execute.
Limits and data handling
Up to 200 assets, 500 dependency edges, 20 proposed changes and 1,000 checks; hosted arguments are additionally limited to 32 KiB. Dependency paths are bounded to eight assets and one reported path per asset. Cycles/depth limits and edges without references require review. External dependencies may be absent.
The website demo is local browser computation. The downloadable 0.2.0 MCP runs the same class of evidence rules locally; hosted use passes arguments/results transiently through your paired connector. A 0.1.0 connector must be replaced with a fresh 0.2.0 pairing before it advertises the new tool. No migration, restart, configuration edit, real simulation or rollback is executed.