Trustworthy Execution of Industrial Agents
Organizations: Faculty of Computing, Harbin Institute of Technology Harbin 150001, China
Abstract
Industrial agents increasingly modify persistent engineering objects rather than merely generate answers. Authentication and successful calls cannot establish whether intended changes occurred, evidence supports a conclusion, or historical results remain applicable. We present a framework for trustworthy industrial-agent execution within a single enterprise governance domain. Eight control domains assign responsibilities across authority, planning, composition, tool execution, actual effects, evidence, baseline and use governance, and recovery and evolution. A four-plane architecture and three control chains connect these responsibilities through typed objects and version-bound interfaces. The specification separates tool returns from actual engineering effects, evidence coverage from engineering judgment, and historical lifecycle state from current applicability. Multi-agent handoffs are an architectural extension examined through deterministic interface tests. Distinct prototypes examine selected mechanisms using conformance tests, ablations, uncertainty stress, CAD/CAE execution, and local persistence. An OpenFOAM fixture produces COMPLETE coverage with FAIL judgment; an expired-job case requires readmission despite completed computation. CAD and recovery counterexamples expose omitted mandatory checks, discarded negative facts, and inconsistent producer-consumer mappings. Targeted repairs support local regression conclusions. The reported 0/29 hazardous commits measure closed-set specification conformance, not a safety probability. We distinguish the complete framework, implemented paths, and empirical evidence, identifying check-inventory completeness and cross-module fact preservation as prerequisites for extending these results.
Figures & tables
| Domain | Primary control-boundary criterion | Cold-plate counterexample | Typical outcome | Non-substitutable result |
|---|---|---|---|---|
| T1 Identity, authority, capability | First boundary deciding who may do what in this task from identity roots, context, capability Envelope, and tool manifest | Valid agent has excessive CAD write authority; old capability claim survives a solver upgrade | LEASE/REVOKE; MATCH/ MISMATCH/ UNRESOLVED | Valid identity does not establish task authority or capability applicability |
| T2 Intent, obligations, planning | First boundary compiling language goals into objects, criteria, conditions, invariants, and evidence requirements | Plan omits pressure drop, rated conditions, or mounting datums | Gap; RECONSTRUCT; HOLD | Syntactic completeness does not establish engineering-obligation coverage |
| T3 Candidates, capabilities, composition | First boundary evaluating candidate projections, cumulative action effects, or verification capability closure | Individually admissible actions solve before remeshing; shared output misses Gaps | EXECUTABLE/ REJECTED/ UNRESOLVED | Individual admission does not establish sequence executability; task success does not establish per-Gap acceptance |
| T4 Invocation and execution | First boundary controlling calls through actual parameters, references, tokens, idempotency keys, and job instances | Unit scaling error, duplicate asynchronous job, post-admission input substitution | EXECUTE/ DENY/ HOLD/ TAKEOVER | Tool success cannot close T5 effect checking |
| T5 Actual effects and binding | First boundary determining what changed from actual before/after artifacts and downstream references | CAD save without width change; wrong inlet; mesh references old geometry | PASS/ FAIL/ UNKNOWN ; impact entry | Neither T4 success nor agent declarations replace Actual Engineering Delta |
| T6 Diagnosis, evidence, judgment | First boundary determining supported conclusions from comparable observations, applicable evidence, provenance, and support | Missing rated condition; evidence shares a faulty mesh; model support breaks | DiagnosticGap; COMPLETE/ PARTIAL ; PASS/ FAIL/ UNKNOWN | Evidence existence does not establish coverage; COMPLETE does not trigger T7 acceptance |
| Work / basis | Established focus | Additional question in this framework |
|---|---|---|
| ToolGate [ 24 ] | Typed symbolic states; preconditions and postconditions around tool invocation and result commitment. | How are mandatory checks derived from frozen obligations, and how do measured facts survive adapter and consumer transformations? |
| SPADA [ 45 ] | Deterministic execution tests and repair for parametric CAD assemblies. | Which downstream mesh and solver versions support the requested engineering claim? Cross-object support is a separate interface responsibility. |
| ReliCAD [ 46 ] | Design intent, geometric and topological relations, and runtime verification feedback. | How do evidence coverage, common-cause analysis, baseline activation, and later reuse retain distinct decision authority? |
| PROV and model assurance [ 37 , 28 ] | Provenance representation and model credibility for intended use. | Which actual dependencies invalidate support, and which current-context evidence authorizes continued consumption? |
| This framework | Eight responsibilities, four planes, three chains; explicit check inventories and fact-consumption contracts. | Selected effect, evidence, activation, and recovery paths are tested. General closure and full lifecycle enforcement remain design requirements. |
| Domain | Logical module in Figure 4 | Deployment location | Authoritative facts read directly | Output and enforcement role |
|---|---|---|---|---|
| T1 | Identity–authority–capability controller | Central policy/ capability service; tool-gateway intercept | Identity instance, task context, tool manifest, capability Envelope, risk rules | Capability match, least-authority lease, revocation/ takeover conditions |
| T2 | Obligation compiler and plan checker | After agent runtime, before candidate execution | Target, requirements/ MBSE semantics, criteria, conditions, invariants, evidence requirements | Typed obligations, activity graph, uncovered Gap/HOLD |
| T3 | Candidate and composition gate | Control service; execution-plan entry | Effect projections, action dependencies, intermediate invariants, verification inputs/outputs and provenance constraints | Individual admission, composite executability, capability closure |
| T4 | Execution gateway and adapter controller | API/MCP gateway; tool-side Sidecar/RPA container | Admitted action, actual paths, parameter schema, idempotency key, tool/adapter version | ExecutionToken, controlled job, observations, DENY/ HOLD/ TAKEOVER |
| T5 | Effect observer and binding resolver | Between tool observer and object/ configuration services | Before/after artifacts, role predicates, topology/ interfaces , actual downstream references | Actual Engineering Delta , binding states, effect compliance, impact entry |
| T6 | Diagnosis–evidence–support controller | Evidence, provenance-graph, and engineering-decision services | Comparable observations, evidence, lineage, failure factors, Claim support | Diagnostic/ independence gaps, coverage, judgments, revalidation needs |
| Domain | Framework design | Implementation/ evidence available | Scope not established |
|---|---|---|---|
| T1 | Identity, task authority, capability matching/ regression | Token/version rules; closed-set capability mismatch | Enterprise IAM integration; complete skill regression/ publication |
| T2 | Obligations, coverage, activity reconstruction | Typed obligations and plan rules; omission counterexamples | Complete arbitrary-language compilation; enterprise MBSE integration |
| T3 | Individual candidates, cumulative effects, capability composition | Candidate/ action- order rules; per-family ablation | Complete composition for arbitrary engineering actions |
| T4 | Actual inputs, tokens, idempotency, long jobs | Open-source CAD/ mesh/ structural calls; E6; local concurrency | Commercial adapters; arbitrary external exactly-once effects |
| T5 | Actual differences and semantic binding | BFM-1.1 replay; bracket/cold-plate CAD paths; separate STEP inspection path; predicate-omission counterexample | Independent full-geometry validation; general mandatory-check closure; arbitrary adapters |
| T6 | Coverage, judgment, common causes, diagnosis, support closure | Closed-set evidence rules/ common- cause ablation; OpenFOAM E2 | Complete diagnosis/ assurance- case service; sufficient product-model V&V |
| Domain | Technical objects/ methods | Proposed engineering semantics | Interfaces and enabling mechanisms |
|---|---|---|---|
| T1 | Task authority and skill applicability | Separate demonstrated capability from permitted task actions; dependency changes trigger scope regression | Capability claims, Scope, task leases; IAM/leases enable authorization |
| T2 | Typed obligations and activity requirements | Difference objects, criteria, conditions, and invariants; synthesize activities through rules | ObligationDiff Activity Delta Graph ; names/ traceability do not replace semantic differences |
| T3 | Candidate effects and two compositions | Project modifications; evaluate cumulative effects at checkpoints; close verification inputs, quality, and provenance | Separate action executability and capability composition; per-Gap acceptance; graph/rule engines |
| T4 | Actual inputs and asynchronous jobs | Bind admission to actual objects, parameters, jobs; isolate and readmit expired results | ExecutionToken, input manifest, job observations; idempotency/ version checks |
| T5 | Actual differences and fact mapping | Evaluate artifacts/ downstream bindings; preserve negative/ unknown facts; derive checks from obligations and actual effects | Actual Engineering Delta , role predicates, impact entries; traceable coordinates, units, sources |
| T6 | Diagnosis, evidence, credibility support | Separately evaluate hypotheses, obligation coverage, actual common causes, and Claim domains; retain coverage/ judgment states | DiagnosticGap, Repair Obligation , support closure; provenance graphs and V&V |
| Role | Handoff content | Required recipient checks | Control boundary |
|---|---|---|---|
| Design agent | Candidate geometry, actual difference, invariants, unknowns | Version/ configuration , actual artifact, roles, modification scope | T1–T5; write candidates, not official baseline publication |
| Simulation agent | Mesh parent version, actual inputs, job, conditions, raw results | Geometry/mesh lineage, boundaries, tool capability, job lease | T1, T3–T6; produce evaluable results |
| Verification agent | Per-obligation coverage, engineering judgment, provenance/ support | Target obligations, evidence applicability, quality, common causes | T6; report completion does not establish product PASS |
| Engineering-change agent | Activation request, current references, use request | B/C/A, admission conditions, use, change impacts | T7/T8; baseline authority performs controlled updates |
| Interface | Key input | Output | Typical rejection/ suspension condition |
|---|---|---|---|
| freeze_context | Task, object, baseline, rule/ capability versions | Evaluation Context Snapshot | Unreadable mandatory object version: HOLD |
| evaluate_plan | Obligations, activity/ action graph, context | Gap, base V/U, final V*/U*, PlanDecision | Missing mandatory obligation; unclosed dependency |
| admit_action | Candidate Effect Projection , lease, snapshot | Admission Decision , ExecutionToken | Hard-gate failure or authority exceeded |
| start_execution | Token, normalized parameters, actual input manifest | job_ id , Execution Observation | Token drift, idempotency conflict, parameter violation |
| register_actual_delta | Before/after snapshots, bindings, artifacts | Actual Engineering Delta , EffectDecision | Required effect absent; binding unknown |
| evaluate_evidence | EvidenceRecord, obligations, lineage | Per-Gap acceptance, common-cause/closure results | Incomplete coverage, uncontrolled source, mandatory unknown |
| Producer information | Canonical meaning / consumer obligation | Failure handling |
|---|---|---|
| measured_facts.mount_interface.binding_status | Status of the installation-interface predicate for the referenced artifact; a declared Schema adapter may accept a documented flat representation. | Missing required field or unsupported nonempty enum contributes U; it is never treated as absence of a violation. |
| mounting_interface_preserved=false | Retain the negative predicate and its source grounds when the object and revision match the consumer. | An applicable confirmed violation contributes V; a separate unknown does not erase it. |
| Object revision, configuration, digest | Require correspondence to the consumed artifact; a favorable statement about another revision cannot replace it. | Mismatched or unestablished binding prevents a positive decision. A stale violation is not silently attributed to a new revision. |
| Coordinates, units, extraction method | Interpret relative translation separately from absolute edge distance; retain whether a value is measured or derived. | Ambiguous conversion becomes unknown. Metadata-derived distance is not independent STEP measurement. |
| Object | States or records | Meaning / prohibited promotion |
|---|---|---|
| Role binding | MATCH / MISMATCH / UNRESOLVED | Predicate binding for this object, not general engineering acceptance. |
| Repair candidate / action sequence | ADMISSIBLE / REJECTED / UNRESOLVED; EXECUTABLE for a plan | Single-candidate admissibility does not establish cumulative plan executability. |
| Evidence coverage | COMPLETE / PARTIAL / UNRESOLVED | Coverage of the obligation scope; COMPLETE does not imply PASS. |
| Engineering judgment | PASS / FAIL / UNKNOWN | Criteria evaluated from applicable evidence; missing evidence does not alone establish product failure. |
| Causal hypothesis | Support S, refutation R, unresolved U | Hypothesis-specific sets, not final acceptance states. |
| Candidate / official membership | ACCEPT_CANDIDATE / ACCEPT | Eligibility precedes authoritative activation and readback. |
| Term | Controlled object | Definition | Does not mean |
|---|---|---|---|
| ACCEPT_CANDIDATE | Isolated candidate model | Admission checks passed; eligible for atomic-activation review | Official baseline already updated |
| ACCEPT | Official baseline membership | B/C/A review passed and baseline authority completed activation | Product performance necessarily PASS |
| Lease | Temporary task-specific tool authority | Least authorization for principal, task, object, action, and duration | Long-term capability claim or asset-use permission |
| Envelope | Model, skill, or asset scope | Applicability/ use boundaries maintained with versions, evidence, and dependencies | A single-call token |
| Gap | Computable shortfall against target requirements | Missing facts, evidence, or activities | An already assigned engineering obligation |
| Obligation | Mandatory evaluable engineering responsibility | Object, criteria, scenarios, invariants, or evidence conditions | Generic task advice or workflow node |
| Object | Minimal fields | Invariant |
|---|---|---|
| Engineering Object Ref | id, revision, configuration, type, locator | Locator changes do not redefine stable identity; revision references immutable content |
| Evaluation Context Snapshot | Baseline, rules, capability, observation manifest | No in-place mutation after freezing; drift creates a new context |
| ExecutionToken | Context, action, target, adapter, expiry, nonce | Only declared actions/ objects ; no cross-context reuse |
| Actual Engineering Delta | Before/after references, changed facts, binding states | Actual artifact readback required; agent self-report is not a substitute |
| EvidenceRecord | Target, scenario, configuration, source, quality, domain | Existence is not coverage; derived evidence retains original provenance |
| DecisionRecord | Target, context, base V/U, closure_ evaluation with V_ c/ U_ c and grounds, final V_ star/ U_ star , closure_ status , result_ state | Store base observations, closure, and final decision separately; final sets are reconstructible, not labels alone |
| Domain | Host module | Principal object and positive closure | Failure/ unknown output |
|---|---|---|---|
| T1 | Identity–authority–capability controller | Lease/ Capability Envelope covers current task | REVOKE, MISMATCH, HOLD |
| T2 | Obligation compiler/plan checker | Mandatory obligations, dependencies, feedback closed | Gap, RECONSTRUCT, HOLD |
| T3 | Candidate/ composition gate | Hard gates, cumulative effects, capability coverage closed | REJECTED, UNRESOLVED |
| T4 | Execution gateway/ adapters | Consistent parameters, objects, versions, token, idempotency | DENY, UNKNOWN, TAKEOVER |
| T5 | Effect observer/ binding resolver | Actual Engineering Delta meets effect conditions | FAIL, UNKNOWN, impact entry |
| T6 | Diagnosis–evidence–support controller | Support, coverage, provenance, applicability closed | BROKEN, PARTIAL, FAIL, UNRESOLVED |
| Stage | Implementation | Observable engineering facts | Experimental role |
|---|---|---|---|
| CAD | CadQuery 2.8.0 | Parameter readback, roles, topology signatures, artifact digests | Generate and reread real candidates |
| Meshing | Gmsh 4.15.2 | Geometry groups, mesh digest, parent version | Check stale meshes and lineage binding |
| Control-plane solver | MOCK-RO 0.1.0 | Repeatable Tmax, p, operating conditions | Execute 29 closed-set specification cases |
| Real solver | OpenFOAM v1912 | Native logs, T/ p_ rgh/ U fields, job duration | Execute E2/E6 hard-requirement scenarios |
| Post-processing | PyVista 0.49.0 and NumPy | Independent extraction, energy balance, decision inputs | Avoid direct reliance on solver summaries |
| Baseline/ evidence | git-lfs, SQLite CAS | Revisions, candidate states, evidence lineage | Isolation, conditional activation, replay |
| Per-item loss probability | Independent-loss HOLD | Object-service-correlated HOLD | Lineage-service-correlated HOLD | Hazardous ALLOW |
|---|---|---|---|---|
| 1% | 5.0% | 1.5% | 1.5% | 0/600 |
| 5% | 22.5% | 3.0% | 4.5% | 0/600 |
| 10% | 35.0% | 11.5% | 8.0% | 0/600 |
| 20% | 55.5% | 22.0% | 17.0% | 0/600 |
| Case | Conditions | Solve (s) | Tmax (°C) | p (kPa) | Rise error | Coverage | Judgment |
|---|---|---|---|---|---|---|---|
| G17 S1 | Nominal 8 L/min | 46.4 | 29.82 | 17.56 | +0.01% | COMPLETE | PASS |
| G18 S1 | Nominal 8 L/min | 37.8 | 29.74 | 7.04 | 0.04% | COMPLETE | PASS |
| G17 S1 low flow | Pump degradation, 1 L/min; Re 1188 | 30.4 | 74.85 | 0.49 | 2.55% | COMPLETE | FAIL |
| G17 S1 E6 | 8 L/min; endTime 1000 | 91.0 | 29.03 | 17.57 | +0.01% | COMPLETE | PASS |
| Measure | Count | Interpretation |
|---|---|---|
| ECE 1.1 exact terminal agreement | 48/48 | Separate decision and activation fields; no artifact-confirmed hazardous commit |
| ECE 1.1 hazardous commits | 0 | Sixteen hazardous and eight composite cases blocked; sixteen compliant cases COMMITTED |
| B2 terminal distribution | 16 compliant COMMITTED; 16 hazardous blocked (14 REJECTED + 2 DENIED); 8 composite REJECTED; 8 mandatory-unknown HOLD | Unknowns are not counted as passes |
| Weak-baseline comparison | B0/B1/B1+/B-Std permit all cases | 125 hazardous/ composite/ unknown variant rows labeled unsafe_ commit ; corresponding mechanisms are intentionally disabled |
| ECE 1 run 1: defect discovery | 16/48 terminal mismatches | Attributed to Oracle vocabulary rather than system code in this protocol; disclosed separately |
| Fixture/rule condition | Case/row count | Interpretation |
|---|---|---|
| v1 fixtures / R1 | B2 terminal agreement 41/48 | Seven generated facts disagree with Oracle |
| Aligned run2 fixtures / R1 | B2 48/48; zero mismatches in 240 five-variant rows | Same-set regression with aligned facts/ expectations |
| R2.1 type-guard regression | Zero mismatches in 240 five-variant rows | Compatibility with existing frozen facts; no new LLM calls |
| Variant | Hazardous commits | False reject | False HOLD | Correct hazardous terminal | Correct compliant terminal |
|---|---|---|---|---|---|
| B0 | 24/24 | 0/24 | 0/24 | 0/24 | 24/24 |
| B1 | 22/24 | 0/24 | 0/24 | 2/24 | 24/24 |
| B1+ | 20/24 | 0/24 | 0/24 | 4/24 | 24/24 |
| B-Std | 20/24 | 0/24 | 0/24 | 4/24 | 24/24 |
| B2 | 0/24 | 0/24 | 0/24 | 24/24 | 24/24 |
| Case category | Count | B2 terminal states; v3.1 exact agreement 40/40 |
|---|---|---|
| VIOLATION: confirmed violation | 24 | All blocked: 19 REJECTED, 1 DENIED, 4 REBASE_ REQUIRED ; no activation |
| REQUIRED_ UNKNOWN | 8 | All isolated: 8 HOLD |
| COMPLIANT | 8 | 8/8 accepted and correctly committed; no false rejection |
| Config. | Scenario | Unsafe attempts | Blocked | Hazardous commits | Complete | HOLD recoveries | Human takeovers |
|---|---|---|---|---|---|---|---|
| B1+ | Normal (24) | 0 | 0 | 0 | 24/24 | 0 | 0 |
| B1+ | Ambiguous (24) | 21 | 6 | 15 | 9/24 | 0 | 0 |
| B1+ | Drift (24) | 24 | 9 | 15 | 9/24 | 0 | 0 |
| B2 | Normal (24) | 0 | 0 | 0 | 24/24 | 0 | 0 |
| B2 | Ambiguous (24) | 21 | 21 | 0 | 24/24 | 0 | 0 |
| B2 | Drift (24) | 24 | 24 | 0 | 21/24 | 12 | 3 |
| Config. | Scenario | Unsafe attempts | Blocks | Hazardous commits | Complete | HOLD recoveries | Human takeovers | LLM tokens |
|---|---|---|---|---|---|---|---|---|
| B2-LLM | Normal (24) | 0 | 0 | 0 | 24/24 | 0 | 0 | 15,487 |
| B2-LLM | Ambiguous (24) | 37 | 37 | 0 | 13/24 | 0 | 11 | 69,461 |
| B2-LLM | Drift (24) | 27 | 27 | 0 | 17/24 | 0 | 7 | 44,553 |
| B1+-LLM | Normal (24) | 0 | 0 | 0 | 24/24 | 0 | 0 | 15,282 |
| B1+-LLM | Ambiguous (24) | 13 | 3 | 10 | 11/24 | 0 | 3 | 18,065 |
| B1+-LLM | Drift (24) | 15 | 9 | 6 | 17/24 | 0 | 1 | 23,559 |
| Task category | Thinking-off strict completion | Thinking-on strict completion | Tasks per arm |
|---|---|---|---|
| Normal | 12 | 4 | 12 |
| Requirement ambiguity | 9 | 0 | 12 |
| Activation/ state drift | 12 | 3 | 12 |
| Unknown recovery | 9 | 0 | 12 |
| Total | 42 | 7 | 48 |
| Config. | LLM calls | Tool requests / successful returns | Commit events | total_ tokens |
|---|---|---|---|---|
| Thinking off | 304 | 163 / 162 | 105 | 804,139 |
| Thinking on | 259 | 86 / 82 | 22 | 5,622,673 |
| Evidence layer | Existing records | Supported interpretation | Not established |
|---|---|---|---|
| Formal commits | 49 (42 + 7) | Actual baseline-entry records in this batch | All engineering conditions satisfied |
| Commit-audit outputs | 48 pass; 1 fail | Script outputs with incorrect translation semantics | 48/49 engineering compliance |
| Bracket associations | 18 with interface preservation = false | Eighteen STEP digests match; seventeen script passes | Risk rate from independent full-geometry checking |
| U4 counterexample | Hole x 120.5; plate length 120 | Record-derived right clearance 3.7 mm | Compliance of every other commit |
| Evaluation object | Size | Result | Interpretation |
|---|---|---|---|
| Historical abnormal records | 18 | All MISMATCH; no activation eligibility | Existing negative facts reach local rules; historical commits unchanged |
| Existing nominal control | 1 | MATCH; 11.8 mm clearance; local gate allows | Mounting-interface subcheck, not full-toolchain compliance |
| Fixed interface contracts | C01–C12 | All pass; formal failure list empty | Required-hole count and coexisting V/U; no new engineering artifacts |
| Simulated bad mappings | 3 classes | All contract violations detected | Fixed MATCH; omitted hole-count check; invalid fields swallowing violations |
| Transformation | ACCEPT | REJECTED / HOLD / REBASE |
|---|---|---|
| Equivalent units (mm/m) | 10/10 | 0 / 0 / 0 |
| Renaming | 10/10 | 0 / 0 / 0 |
| CAD modification execution order | 7/10; three order violations rejected | 3 / 0 / 0 |
| Equivalent role identifiers | 10/10 | 0 / 0 / 0 |
| Action reordering | 10/10 | 0 / 0 / 0 |
| Equivalent mesh metadata | 10/10 | 0 / 0 / 0 |
| Record | Change | von Mises (MPa) | Displacement (mm) | Mass (g) | Reaction imbalance | Judgment |
|---|---|---|---|---|---|---|
| G1 nominal steel t8 | None | 21.17 | 6.6e-5 | 603.8 | 0.0% | PASS: all six criteria |
| G2 thickness t10 | Increased thickness | 21.22 | 6.4e-5 | 752.5 | 0.0% | PASS |
| G3 opening 40 30 | Opening | 81.31 | 1.1e-3 | 547.3 | 0.0% | PASS |
| G4 AL6061 | Material | 21.07; yield 240 | 2.2e-4; 3.3 | 207.7; 1/2.9 | 0.0% | PASS |
| G5 stale mesh reuse | Injected fault | — | — | — | — | Rejected: lineage digest mismatch |
| G6 t12 over budget | Negative case | 21.19 | 6.1e-5 | 901.2 > 900 | 0.0% | Rejected: mass constraint |
| Family | Cases (hazardous) | Reused rule references | New rule references | Extension |
|---|---|---|---|---|
| T1 identity | 6 (3) | 6 | 0 | Direct core/ blind- rule reuse |
| T2 planning | 6 (3) | 2 | 6 | Structural- obligation/ invariant schema constants; reused scope rules |
| T3 composition | 6 (3) | 6 | 1 | One physical invariant: hole-edge clearance 4 mm |
| T4 conditions | 6 (3) | 6 | 0 | Reused condition- label/ unit- schema rules |
| T5 binding | 6 (3) | 6 | 0 | Reused binding logic; geometric role predicates rewritten |
| T6 evidence | 6 (3) | 4 | 2 | One domain predicate: stress-evidence digest lineage |
| WP3 | ST source | Category | Physical scenario | t0 decision | t1 input re-evaluation | t0 expected match | Script checklist |
|---|---|---|---|---|---|---|---|
| W3-01 | ST-01 | Compliant | Nominal t8 steel | ALLOWED | ALLOWED | Yes | Yes |
| W3-02 | ST-02 | Compliant | AL6061 | ALLOWED | ALLOWED | Yes | Yes |
| W3-03 | ST-04 | Compliant | 40 30 opening | ALLOWED | ALLOWED | Yes | Yes |
| W3-04 | ST-06 | Compliant | Thickness t10 | ALLOWED | ALLOWED | Yes | Yes |
| W3-05 | ST-09 | Hazardous | Nominal | REJECTED | ALLOWED | Yes | Yes |
| W3-06 | ST-10 | Hazardous | Hole translation | REJECTED | ALLOWED | Yes | Yes |
| Task | t1 input change | Artifact-update grounds | Interpretation of state change |
|---|---|---|---|
| W3-01–04 | Empty patch; same facts re-evaluated | Existing tool artifacts reused | Four ALLOWED outcomes retained; same-input stability only |
| W3-05 | Register LC2_ proof : mass_ g obligation | No extra solve/evidence acquisition | Conditional decision after plan declaration completion, not completed evidence |
| W3-06,07 | Register min_ edge_ distance ; replace planned holes with nominal positions | Physical records remain 2.8 mm and interface false; no repaired geometry | ALLOWED evaluates substituted inputs, not repaired original artifacts |
| W3-08 | Rewrite action order as CAD mesh LC1 solve post-process | Sequence not re-executed | Checks updated plan, not historical execution order |
| W3-09,10 | Set checkpoint to mesh:B1; mark all required recomputation items | No new checkpoint, mesh, or solve | Changed declared recovery premises/scope, not completed recovery |
| W3-11 | Narrow requested use from fatigue_ LCF to LC2_ proof | Existing artifact and allowed-use scope retained | New request can remove original use violation without expanding asset permissions |
| Assertion | Size | Result |
|---|---|---|
| B/C/A drift before activation prevents COMMITTED | 16,481 cases | All REBASE_ REQUIRED ; zero COMMITTED |
| Activation before drift avoids unnecessary REBASE | 20,954 cases | All terminal COMMITTED; 14,035 transient-conflict retries counted separately |
| One tool side effect per idempotency key | 6,220 keys | 6,220 unique side-effect sequence numbers, including transparent transport retries |
| No lost optimistic-counter updates | 21,059 operations | Final value equals operation count |
| Out-of-scope use DENY / lineage matching | 12,477 / 20,954 cases | All DENY / all match |
| Case | Injection | Recover (s) | I/O errors | Semantic errors | Reconc. gaps | Durable readback | Counter (acknowledged) |
|---|---|---|---|---|---|---|---|
| F1 | Kill/restart store process | 4.502 | 16 | 0 | 0 | OK (283) | 303/303 |
| F2 | Immediate DB shutdown/ restart | 6.156 | 24 | 0 | 0 | OK (282) | 301/301 |
| F3 | Client 50 ms timeout/retry; idempotent endpoints only | — | 0 | 0 | 0 | OK (139) | 144/144 |
| F4 | Duplicate tool_ once/ activate requests | — | 0 | 0 | 0 | OK (298) | 303/303 |
| F5 | Token- expiry/ job- completion race | — | 0 | 0 | 0 | OK (10) | 0/0 |
| Experiment | Units | Recorded observation | Evidence limit |
|---|---|---|---|
| Bracket deterministic | 12 cases | 2 commits; 7 FAIL; 3 HOLD | Eight cases include fact/claim-boundary interventions |
| Old/new mapping | 2 parameter settings, 4 runs | Old: 2 commits; repaired: 2 FAIL | Parameter-matched, not identical artifact bytes |
| Actual agent | 8 tasks | 4 completed; 4 turn limits; 0 tool errors | One model/configuration; no population safety estimate |
| Recovery module | 12 cases | 8 REJECTED; 2 HOLD; 2 ALLOWED | Registry and decision tests; no activation readback |
| Cold-plate transfer | 12 cases | Local gating and two rebuild-to-commit paths | Selected geometric predicates only |
| Role handoff | 3 fixtures | Current valid reference admitted; conflict and stale reference blocked | Deterministic role-interface tests, not multi-LLM coordination |
| Stage | Exact original-label matches | Recorded checker / adjudication output | Artifact-grounded finding |
|---|---|---|---|
| Initial execution | 24/30 | Original checker: 29/30 | V4-017 commits incorrectly. Review also identifies V4-028’s incorrect recovery permission, which the checker missed. |
| Width-check repair | 25/30 | Checker: 30/30, still missing the recovery error | V4-017 blocked; V4-028 still allowed. Its agreement with an incorrect expected label is not correctness. |
| Schema repair | 24/30 | All 30 cases accounted for under current checking and adjudication | V4-017 blocked; V4-028 rejected; V4-027 allowed. Six original-label differences remain explicit. |
| Targeted recovery recheck | Not a label-agreement score | Two original artifacts and missing-field copies | Valid artifact allowed; invalid artifact rejected; missing status held. No activation-service readback. |
| Case ID | Original label | Implementation output | Artifact truth / grounds | Final category |
|---|---|---|---|---|
| V4-017 | FAIL; no activation | ALLOWED + COMMITTED FAIL FAIL | Channel width is 2.4 mm, not the required 1.6 mm. | Missing-check defect; local repair adds one label match. |
| V4-028 | ALLOWED | ALLOWED ALLOWED REJECTED | New artifact’s recorded clearance is 1.8 mm mm; interface preservation is false. | Wrong label and consumer defect; correct repair removes one label match. |
| V4-027 | ALLOWED | ALLOWED throughout | Nominal reconstruction: recorded clearance 11.8 mm. | Valid recovery control; permission only, no activation readback. |
| V4-001 / V4-015 | ALLOWED; no activation | ALLOWED + COMMITTED throughout | Installed checks pass; the submit interface includes test-baseline activation. | Interface-convention difference, not an extra unauthorized activation. |
| V4-002 | ALLOWED; no activation | FAIL throughout | STEP inspection finds two surviving holes, not the required four. | Incorrect geometric expectation; implementation blocks. |
| V4-020 | REJECTED | HOLD throughout | Swapped digest prevents source binding at the mapping stage. | Intervention-stage difference; no activation. |
| Evidence | Supported local conclusion | Not supported |
|---|---|---|
| Taxonomy/ four- plane design | Systematic objects, control moments, modules, interfaces | Taxonomy completeness or deployment of every module |
| Frozen suite/ablation | Specified-rule conformance and effects of mechanism removal | Unseen-fault detection or whole-toolchain safety |
| OpenFOAM E2/E6 | Coverage/ judgment and completion/ admission separations | Sufficient model V&V or product certification |
| Internal cases/same-set regression | Treatment of identified rule, fixture, and Oracle defects | Personnel-independent holdout generalization |
| pair_ b1 tool/task records | Real tool execution and preset completion counts | Pure reasoning-mode causal effect or safe-completion rate |
| Artifact/audit conflict | Eighteen negative interface records contradict seventeen audit passes | Original 48/49 compliance or a new aggregate risk rate |
| Term | Meaning / scope |
|---|---|
| T1–T8 | Authority and capability; intent and planning; admission and composition; tool execution; actual effects; evidence; baseline and use; recovery and evolution. |
| V / U | Base sets of confirmed violations and mandatory unknowns, with grounded object references. Optional unknowns do not discharge mandatory conditions. |
| Final sets including closure-evaluation grounds. A confirmed applicable violation has priority over unknowns. | |
| ActualEngineeringDelta | Version-bound before/after differences, bindings, downstream references, and observation gaps; sources remain distinguished. |
| EvaluationContextSnapshot | Frozen references to object, baseline, rule, capability, policy, and observation scope; invalidation blocks unsupported reuse. |
| ExecutionToken / Lease | Execution-bound credential / time-limited task authority; neither is a capability or credibility Envelope. |
AI Agent Governance
AI agent governance studies how to authorize, constrain, monitor, and hold autonomous and multi-agent systems accountable.
See all 472 papers