Meta-n report verification

Type: agentic-system-analyses/types/agentic-system-verification.md

Verification

The current reports support a bounded implementation account. The previous invocation and helper-provenance blockers are repaired; no material blocker remains. I judged the current reports and correction answers afresh against SRC-1, not against the previous verdict as authority. The checkout HEAD matches the boundary, and its tracked working tree is clean. No target code, tests, examples, providers or experiments were executed. The supplied structural report has no findings. Implementation establishes capacity and wiring, not successful operation or causal benefit.

Correction dispositions

The runtime and workflow now display meta-n --tasks <json> --use-archive and place archive-enabled configuration before baseline reachability. Independent reading of SRC-1 meta_n/main.py confirms that the startup guard returns 1 before model-client construction and task loading if archive mode is disabled. RT-RTE-native-solve and the workflow account preserve the supported solving mechanism after that condition. Both invocation answers are supported.

Independent tracing of SRC-1 meta_n/core/verified_code.py and meta_n/core/evolutionary_orchestrator.py confirms last-line stdout JSON parsing without a preliminary nonzero-return-code rejection, a status-derived sandbox bit, and name stamping on pass plus that bit. RT-RTE-helper-verification now describes check-result metadata, not authenticated helper-call evidence. Its immediate return, selection predicate, evidence limits, explanatory passage and load-bearing guarantee retain that qualification. MEM-OBJ-python-library, MEM-RTE-archive-resume and the memory integration issue likewise preserve the trust condition through storage, ranking and reload. Both provenance answers are supported. Real-harness filtering remains implemented; correcting metadata strength does not erase it or establish an exploit.

The workflow also incorporates the earlier launcher limit in WF-CLM-delivery-boundary. Its new passage agrees with SRC-1 scripts/run_agentic_stage2_cobench.sh. The limit remains below because correct CLI finalization does not establish correct completion detection by all launchers.

Records, source support and evidence strength

I reviewed every declaration and annotation, its source anchor and evidence layer, status, consumer scope, operative-part relation and dependent account or ledger finding. Independent source tracing tested the consequential conditions and ordering in these groups:

  • Inference and native effects. RT-CMP-llm-client, RT-CMP-inner-llm and RT-CMP-external-backend identify separate invocation responsibilities, not necessarily separate weights. SRC-1 meta_n/core/llm_client.py, meta_n/core/llm_helpers.py and the shipped SDK runner seams support configured provider requests and lazy callback-client construction. A model name is not an immutable weight pin; no inspected Meta-n training path establishes parameter change. SRC-1 meta_n/core/solver.py, meta_n/core/base_executor.py and meta_n/core/meta_layer.py support RT-RTE-native-solve: rendered text goes to the model, extracted source goes to the executor, and local verification runs only after successful script exit. Nested layers normally return source; the outer layer executes and owns its retry loop, retaining the best score.
  • Injection objects and authority. RT-OBJ-pre-process, RT-OBJ-code-library, RT-OBJ-rationale, RT-BAP-prompt-guidance and RT-BAP-executable-injection distinguish executable rules, emitted prose, helper advertisement and retained explanation. SRC-1 meta_n/core/meta_layer.py, meta_n/core/omega.py and meta_n/core/external_agents/injection.py support those consumers. Python/bash parts and EPI-OBJ-additional-context preserve materially different checks and force. Advisory supply does not prove compliance; staged source does not prove invocation. RT-CLM-host-boundary correctly rejects whole-system sandboxing: preprocessing executes in a host daemon thread, timeout abandons rather than kills it, and failed deep copy shares the task. Local process-group killing and bounded pipe draining do not prevent host filesystem or network effects.
  • Proposal, gate and repair. RT-RTE-chain-admission, MEM-RTE-omega-context, EPI-RTE-gate-check and EPI-RTE-gate-disposition match SRC-1 meta_n/core/evolutionary_orchestrator.py, meta_n/core/archive.py, meta_n/core/omega.py and meta_n/utils/context_manager.py. Parent eligibility and elite/weighted selection differ from retention. The gate admits on a clearing sample without a floor; an enabled floor scans for sampled vetoes. Missing baseline or disabled margin leaves liveness-only clearing. Gate traces can be reused in full evaluation. Sampling, output truncation and oldest-layer eviction alter Ω's view, not the full executable chain. Rationale has a subsequent Ω formatter reader. RT-RTE-within-refine re-gates a distinct repaired candidate attached to the archived parent. Its matched-subset event comparison is not its archive veto. RT-RTE-downstream-repair replaces an intermediate injection, evaluates and archives nonempty variants even when their comparison is non-improving. Higher-layer feedback is not clean causal blame evidence. These paths support improvement-directed retained organization change, not demonstrated criticism, learning benefit or exercised self-improvement. The independent theory conditions remain appropriately qualified.
  • Helper checking and downstream selection. EPI-RTE-helper-check and EPI-RTE-helper-disposition separate measurement from admission within the runtime helper route. SRC-1 meta_n/integrations/co_bench.py supplies a real verifier only for crew-only selection, using the test split, the fixed comparator and feasibility of at least 50%. The runner evaluates source under upstream config and normalization; outcome warrant remains conditional on that contract and truthful runner output. Stub retention supplies no value evidence. Dependency closure can restore failed sibling entry points without newly stamping them verified. Override-aware live-name accounting and syntactic call evidence control per-task-best bars, not candidate history or mean. Adapter demotion separates retention from deployment. Reusing the selection split later does not provide an untouched holdout for helper-selected candidates.
  • Persistence and alternate consumers. MEM-OBJ-trace, MEM-OBJ-checkpoint, MEM-RTE-archive-resume, MEM-RTE-seed-import and MEM-RTE-test-readback distinguish evaluated content, control metadata, reconstruction, operator-selected import and downstream evaluation. SRC-1 meta_n/core/run_persistence.py, meta_n/core/archive.py, meta_n/utils/atomic_io.py, meta_n/main.py and the orchestrator test branches support these distinctions. Candidate files are not one transaction; checkpoint/index replacement is individually atomic with best-effort directory fsync. Reload rejects incomplete injection counts, unions checkpoint and candidate bars, and reconstructs child counts. Raw Ω sidecars are retained but omitted by ordinary rebuilding; rationale and source restore. Import selection is not evidence of authorship or per-function review. RT-OBJ-task-solution-map and RT-RTE-frozen-routing match inherited-trace and deterministic-assembly branches. Matching native tasks bypass prompting, prepend and retries. Assembly's stored score is not fresh execution; test branches execute stored scripts or reconstruct a solver. Exact replay, unseen-task transfer and external-map parity are not established.
  • Task frames and external solving. RT-RTE-agentic-loop, MEM-OBJ-prior-summary, MEM-RTE-context-compaction and the epistemic summary counterparts agree with SRC-1 meta_n/core/agentic_solver.py. The summarizer receives removed middle messages; later turns receive first message, summary and four-message tail without a fidelity check or raw-turn reload. RT-RTE-external-spine matches SRC-1 meta_n/core/external_agents/solver.py and meta_n/core/external_agents/injection.py: plan construction, precheck, lease, after-wait headroom check, provision, staging, SDK execution, accounting, extraction, scoring and best-effort log copy have separate owners and failure boundaries. Cancellation propagates and telemetry creation precedes guarded plan construction. SRC-1 scripts/oh_runner.py, scripts/t2_runner.py and meta_n/core/external_agents/backends/openhands.py confirm OpenHands' suffix/tool grant, Terminus' prefix fold and differing host/container envelopes. None establishes deployed isolation or complete transcript preservation. RT-CLM-budget-bound appropriately limits admission checks and proven spend: concurrent or opaque runs need not obey an exact dollar ceiling.
  • Epistemic outcomes, reporting and delivery. EPI-OBJ-predictions, EPI-OBJ-eval-result, EPI-OBJ-eval-feedback and EPI-RTE-classification-check distinguish asserted labels, comparison results and interpreted feedback. SRC-1 meta_n/integrations/text_classification.py supports normalized match fractions and partial-success semantics, not medical label truth. The concrete deductions are source deductions, not observed runs. MEM-RTE-postrun-analysis, EPI-OBJ-arm-report and EPI-RTE-arm-report match SRC-1 meta_n/analysis/best_of_k.py and the emergent-role loader in meta_n/analysis/emergent_roles.py: shared-task comparison, cross-run maxima and advisory statistics do not enforce compute parity or establish causal superiority. The loader's sidecar comment conflicts with the writer, but the reports describe the actual omission correctly. WF-CLM-delivery-boundary matches CLI finalization and individual final artifact writes in SRC-1 meta_n/main.py and meta_n/core/run_persistence.py. Hard budget exceptions skip finalization; returned aborted results are not completed searches.

EPI-CLM-recursive-improvement remains attributed README doctrine/reporting. EPI-CLM-verified-channel retains the verifier module's stronger claims while comparing them with counterpaths. Neither is converted into empirical support. No evidenced-absence record is declared. The epistemic ledger separates checking, disposition, retention and operational use; its route-local lack of accepted explanation integration is not a system-wide negative. Unsupplied generation and summary instances justify indeterminate semantic relations rather than invented ampliative or faithful-reshaping classifications.

Reconciliation and conflicts

The four amendments in reconciliation establish identity for the Python helper part, bash helper part, prior summary and summary-replacement route. Matching source fields, slices and consumers support them. They preserve the epistemic findings rather than importing memory status as warrant. The library container remains valid beside parts. Selected Ω input, gate check/disposition, helper check/disposition and compaction are supported parts of their enclosing routes; emitted prose is an output part of the preprocessing container. Numeric results and feedback may precede traces, and neither is identical to the complete trace. Maps, checkpoints, import, resume, test reads and diagnostic output retain distinct referents. Diagnostic loading and full statistical reporting overlap without coinciding. No missing part necessary to the emitted bounded conclusions was established.

The previous two conflicts no longer describe opposing current findings. Memory's issue 1 and the epistemic reference to older runtime shorthand are stale references, explicitly disposed by reconciliation; they do not override the corrected entry condition or substantive provenance qualification. They are not material unresolved defects. Atomic single-file protection versus multi-file non-atomicity, test-split selection, helper demotion, frozen assembly and repair-event acceptance remain additive qualifications. The reconciler has not selected a stronger unsupported status or concealed an unresolved source-settled challenge.

Omission questions and stopping decision

Omission selection used the frozen top-level tree, alternate dispatch branches, shipped launchers and producer/consumer dependencies, not only submitted anchors. I investigated three consequential questions:

  1. Can retained source reach later work without ordinary proposal admission? Seed import, resume, frozen maps, test evaluation and external staging establish alternate conditions. The reports already preserve these distinctions and emit no universal gate guarantee.
  2. Does a stamped helper name establish trustworthy execution or value evidence independently of the runner channel? The runner/parser/predicate/stamping/override/bar chain confirms the corrected conditional account. No identified stronger provenance assertion remains. Further security inspection is unnecessary to judge that account; no exploit or universal safety conclusion is claimed.
  3. Do other shipped consumers change the delivery or knowledge-production conclusions? SRC-1 scripts/run_agentic_stage2_cobench.sh confirms the retained completion-detection limit. Inspection of the entry seams in meta_n/analysis/depth_attribution.py and scripts/generate_synthetic_sr.py confirms additional diagnostic-verdict and data-writing responsibilities. The epistemic coverage table already withholds a system-complete knowledge-production negative. These seams do not establish their full warrant or provenance, and no such comparison is accepted here.

Further optional expansion stopped because the selected challenges have supported dispositions and no identified material contradiction remains without one. Detailed unassessed evaluator, diagnostic, embedding, SDK-transcript and cache behavior still prevent complete warrant, memory-inventory, freshness or backend-policy comparisons. They do not invalidate the supported trace-to-code, trace-to-summary, persistence, selection and selective CLI workflow account. Missing campaign evidence independently prevents exercised-operation and causal-benefit conclusions.

Blockers

none

Limits

  • The selected workflow account and WF-CLM-delivery-boundary establish CLI finalization, not uniform launcher completion detection or automatic recovery. SRC-1 scripts/run_agentic_stage2_cobench.sh skips on summary-file existence, while SRC-1 meta_n/core/run_persistence.py writes mid-run summaries, consistent with MEM-RTE-archive-resume. A mid-run artifact can therefore suppress this launcher's rerun. Withhold those broader launcher/recovery comparisons; selective delivery and operator-directed resume remain supported. The current workflow explicitly carries this limit and its source passage.