| Requirement | Owner / actual path | Boundary |
|---|---|---|
| Selective trigger; small fixes stay lightweight | Five Skill descriptions, SessionStart route, global template | Model routing is guidance; host-trigger evidence is recorded separately |
| Current local/online evidence and code reuse | rigor-evidence, research-routing, papers; evidence/fetch CLI | Receipt proves retrieval; Lead verifies primary claim, version and license |
| Complete causal implementation and architecture | rigor-task, implementation-debugging/design | No canned architecture or mandatory paper count |
| ML train/infer/data/mask/NaN/checkpoint/metric correctness | ml-experiments, project template, declared run metrics/artifacts | Project declares real split/runtime/DDP/thresholds; generic tests do not prove model learning |
| Resource preflight, naming, quiet long jobs | compute-resources/naming, inspect_resources/check_run_id helpers | External job ownership and GPU availability still need observation |
| Real acceptance and stale-result rejection | tasks.py run/finish; frozen source/argv/artifact/metric contract | Source/data/environment omitted from declared scope cannot be inferred |
| Zero-agent default; substantial audit/review only | delegation, policy.delegate, agent plan | Written gain is not a quantitative token-savings measurement |
| Models, effort, no-parent-history | role templates, native_request, agent run | Native receipt may not expose actual model; no claim of zero system/project context |
| Cap, overlapping writers, reuse/fresh boundary | SQLite assignments/task scope, bind/close, fresh exec | Arbitrary native calls outside CLI are not centrally enforced |
| Exact cleanup, obsolete weights, preserve unknowns | artifacts reserve/register/retention/plan/apply | Existing files require separate exact user-authorized review; external writers are not sandboxed |
| Resumption and interruption | external SQLite, task resume, recover-run | Crash before child identity needs manual inspection; no fake success recovery |
| Selective project memory, one canonical home | rigor-continuity + memory/retrieval/state/consolidation/policy | No automatic global memory edits or per-turn document touch |
| Obsidian/Serena navigation | memory helpers and references | Derived pointers, structural audit not semantic truth |
| Library retrieval/bootstrap/ingest/link-code/verify | papers/catalog/converter references; installed PaperMeld CLI | No copied parser; unavailable CLI blocks mutations, read-only catalog remains possible |
| Git checkpoint and preservation | Git recovery + next_checkpoint helper | Task-owned local commit only, no automatic push |
| Short answers without hiding failures | global template, skill completion guidance | Caveman-style brevity cannot erase material evidence or limitations |
| One plugin, no duplicate engines | shared rigor package/scripts/references; five entry Skills | External connectors reused only when relevant; no new MCP daemon |
| Global vs deep-learning instructions | templates/global.md, templates/deep-learning.md, setup responsibility table | Templates supplied, never silently installed |
Original inputs: global and DualCMP AGENTS, code-implementation (9 reference topics/4 scripts), memory-maintenance (4 reference topics/3 scripts), papermeld (catalog/converter and CLI workflows), user transcript and current corrections. Engineering and memory references were reused with Rigor-relative paths; delegation was rewritten around the user's cost boundary. Shared helper implementations remain single copies. No private library configuration is distributed.
Integration references: official Codex plugin/skill/hook/subagent mechanisms; Ponytail's event-driven entry inspired short lifecycle routing rather than copying its orchestration. GitHub and Exa support evidence discovery when needed; no meaningless tool calls on ordinary local fixes.