IRIron Reinsinteractive autonomy atlasSources ↗
THE UNCOVERED ROOMS

Automation can fill the building.
It cannot sign the deed.

Coverage below means “the reference stack has machinery,” not “the problem is solved.” The remaining gaps cluster around causality, accountability, proprietary surfaces and the physical world.

Important: these are architecture coverage estimates for teaching—not benchmark results.
WHY THE ROOM IS OPEN

Risk acceptance, legal interpretation, regulated notification and audit opinion are human constitutional acts.

REFERENCE RESPONSE

Keep named human officers or contracted accountable professionals; preserve every recommendation and decision.

⚖HUMAN CONSTITUTIONsmall does not mean optional
Product & domainWhat should exist?
Risk & legalWhat may happen?
ProductionWhat may change now?
IncidentWhen do we declare and communicate?
DataWhat does this fact mean?
AuditWhat evidence is sufficient?
Root cause is not a good storyA wet floor does not prove the roof leaked.

Diagnosis needs evidence plus a discriminating test or counterfactual prediction. Localization, cause, mitigation, repair and recovery are scored separately.

TEST: Plant misleading telemetry and red herrings; reward only the true causal chain and safe recovery.
Green tests can hide slopA tower can stand today and still be badly built.

Measure duplication, architecture erosion, hollow tests, insecure behavior, review burden and defect escape—not just task completion.

TEST: Evolve the repository for weeks, inject requirements and audit structural drift plus maintainer acceptance.
Too many robots can grab the same wheelTwo drivers steering one car is not twice as safe.

Every mutable field needs one authority. Overlapping GitOps, cloud, SaaS and agent writers need explicit conflict and reconciliation contracts.

TEST: Issue conflicting changes during partitions and verify one traceable winner, compensation and no hidden drift.
Every fact has more than one clockWhat we knew yesterday is not always what was true yesterday.

Preserve valid, recorded, commit, capture, observation, computed, fitness, use and decision-effective time with correction lineage.

TEST: Backdate, correct and retract facts; reproduce prior decisions and enumerate every affected downstream use.
The connector is a separate productA telephone can mishear even when the person knows the answer.

MCP/API/CLI layers can omit pages, collapse identity, cache stale data, cross tenants or lose native audit. Assess them separately from the source system.

TEST: Compare connector answers with native APIs under pagination, revocation, stale cache and adversarial content.
The old mud ball contains hidden contractsBefore replacing a messy machine, watch every button people really press.

Network observation helps discover behavior but cannot infer hidden business truth, batch jobs, legal rules or valid-time history.

TEST: Shadow and dual-run a replacement; reconcile outputs, side effects and users before canary authority transfer.
Day 2 is the hard partFinishing the treehouse is easy. Keeping it safe through storms is hard.

Require upgrades, patching, user incidents, vendor changes, backup/restore, on-call and recurrence prevention—not only deployment.

TEST: Operate for 30/90 days with timed faults, absences and externally triggered changes.
Open is not one switchThe recipe may be free while the oven, ingredients and right to sell the cake are not.

Source, binary, container, model, dataset, firmware, hosted control plane and support entitlements must be tracked independently.

TEST: Block promotion until every transitive artifact has reviewed deployment and redistribution rights.
Humans are constitutional, not cleanupRobots can carry bricks. People decide where homes may be built.

Humans retain product meaning, legal/risk decisions, privileged approval, incident command, data semantics and independent audit.

TEST: Measure the exact human minutes, expertise, interrupts and decisions—never hide them in ‘autonomous’.