32 KiB

Raw Blame History

11 - Architecture Turnaround Package

Purpose

This folder is the execution-oriented continuation of the baseline note:

11 - unified_project_architecture_and_reference_update_plan_2026-04-15.md

That baseline note answers:

what the project is today;
where the main architectural fragility sits;
what direction is safe.

This package answers the next question:

how the team should design the architectural turnaround without breaking the current exact-data baseline.

Package Contents

Current Status Snapshot (2026-05-01)

This package is no longer planning-only.

Status canon for planning:

Post-F Semantic Integrity Hardening is operationally closed at 99% and should now be used as a regression gate, not as the active module denominator.
Planner Autonomy Consolidation is closed at 100% for the declared phase83 planner-brain slice.
The active next module is now Open-World Bounded Autonomy Breadth over unfamiliar 1C asks, with Post-F and phase83 retained as semantic canaries.
The first active slice is Business Overview Evidence Fusion: broad company-analysis wording now produces a richer evidence-grounded business overview from confirmed MCP/session facts instead of a thin generic summary.
The current follow-up slice is Business Overview Catalog Route Fabric: business_overview is now a reviewed catalog/data-need/planner chain, while fresh multi-probe execution remains honestly bounded behind an unsupported runtime scope until the next bridge is implemented.
The short source of truth for status wording is 21 - current_status_canon_2026-05-01.md.

It now documents a turnaround that is already operational in code, already materially past the acute regression breakpoint, and already moved through bounded MCP autonomy, Post-F hardening, inventory breadth proof, and the declared Planner Autonomy slice:

route, transition, boundary, meta, memory, and provider policy owners exist as separate modules;
exact-lane truth and coverage/evidence contracts exist as explicit runtime artifacts;
scenario acceptance writes machine-readable scenario_acceptance_matrix.json and pack_state.json;
AGENT semantic packs and source catalogs already exist for mixed domain/meta validation.
the reset toward MCP-first bounded autonomy is now formalized;
Big Block A/B/C of that reset are now closed in runtime code and replay-backed;
Big Block D/E/F are now also materially closed in runtime code and replay-backed:
- Question -> Data Need Graph
- dynamic schema traversal and primitive search
- multi-hop evidence loop with bounded clarification recovery
the Post-F semantic integrity layer is now a closed regression-gate layer:
- protect grounded subject integrity against stale scope contamination
- protect exact and planner-selected pivots from metadata/discovery drift
- keep temporal continuity and repeated lane switches semantically stable
- recover already-supported questions that still look broken to a human user
the Post-F module is now operationally closed as a hardening slice:
- code fix commit: 739e8b8 Post-F: закрыть ручные провалы НДС, выручки и item-flow
- runtime artifact tail commit: 837e1fe Post-F: сохранить хвосты ручных runtime-прогонов
- live map sync: 18 - post_f_code_documentation_sync_2026-04-24.md
the next bounded breadth slice is now replay-backed under Post-F gates:
- inventory stock/provenance/sale-trace pack: inventory_stock_open_world_breadth_rerun_semantic_integrity_20260501_fix5, accepted
- live map sync: 19 - inventory_stock_open_world_breadth_proof_2026-05-01.md
Planner Autonomy Consolidation is now closed for the declared phase83 slice beyond the initial catalog-template cut:
- MCP catalog now carries reusable chain templates in addition to primitive contracts;
- planner route-fabric selection has started moving from local recipe branches toward catalog-instantiated chains;
- value-flow, value-flow comparison, value-flow ranking, lifecycle, metadata, movement, document, and entity fallback branches now expose catalog-template instantiation reason codes in planner output;
- explicit document/movement data-need now scores over ambiguous carried metadata surfaces without forcing neutral follow-ups into a lane;
- lifecycle now behaves as a bounded activity-window inference chain with an explicit legal-fact boundary instead of an unqualified age answer;
- current-turn value-flow aggregate questions can override narrower supported exact routes when the user asks for totals/net/payment amounts;
- broad business evaluation remains in the deterministic living-chat bridge instead of being displaced by generic metadata discovery;
- inventory stock snapshot, supplier overlap, purchase provenance, and sale trace are now reviewed catalog chain templates; generic free-form inventory execution remains forbidden, and evidence must pass through reviewed exact recipe bridges;
- runtime bridge and answer adapter now keep unsupported inventory route templates behind an explicit user-facing boundary instead of letting template planning look like confirmed stock/supplier/purchase/sale evidence;
- inventory catalog templates now bridge through existing exact inventory recipes (41.01 scoped stock, supplier overlap, purchase provenance, and sale trace) inside the bounded MCP discovery pilot, while missing selected-item anchors still clarify instead of guessing;
- unambiguous metadata surfaces can now infer the next reviewed lane from Document.*, Register.*, or Catalog.* objects even before upstream labels downstream_route_family, while mixed surfaces still do not guess;
- catalog index now scores reviewed chain templates directly from fact/action/axis/comparison/ranking needs, and planner/runtime/debug surfaces expose ranked catalog chain matches through the structured catalog_chain_template_matches contract path instead of relying only on reason-code strings;
- planner/runtime/debug surfaces now expose catalog_chain_template_alignment, so semantic replay can see whether selected chains match the catalog top match, fall back to a lower-ranked template, or bypass catalog search;
- planner reason codes now also emit stable catalog-alignment telemetry, so automated replay review can filter top-match, lower-rank, outside-match, and unscored selected-chain states without hand-parsing debug JSON;
- catalog-alignment now carries a single alignment_status verdict through planner/runtime/debug, making replay divergence detection explicit instead of reconstructing it from booleans;
- truth-harness and scenario acceptance artifacts now preserve catalog-alignment status/top-match fields, so AGENT replay review can spot planner-vs-catalog divergence directly in truth_review.md and scenario_acceptance_matrix.json;
- truth-harness now emits a warning finding when selected chains fall below or outside the reviewed catalog top match, unless a spec explicitly allows that divergence;
- scenario acceptance now exposes catalog_alignment_ok, so planner-vs-catalog divergence is a first-class acceptance invariant instead of an ungrouped warning;
- truth-harness specs can now assert expected catalog-alignment status/top-match/top-flag per step, so AGENT packs can validate the planner brain's selected chain against the reviewed catalog route fabric;
- the phase66 open-scope money dialog spec now asserts expected catalog-chain top matches across value-flow totals, bidirectional comparison, and ranking follow-ups;
- the phase32 selected-counterparty chain spec now asserts expected catalog-chain top matches across entity grounding, incoming/outgoing/net value-flow, document evidence, and movement evidence follow-ups;
- AGENT semantic source catalog generation now preserves expected catalog-alignment fields and tags reusable steps as planner_catalog_alignment, so mixed pack construction can find planner-brain regression probes explicitly;
- phase83 planner-brain mixed replay spec is now generated from the AGENT source catalog and interleaves selected-counterparty catalog alignment, open-organization money flow/ranking, broad-evaluation continuity, metadata drilldown, and off-domain living-chat safety;
- phase83 live replay now accepts 20/20 under guarded MCP live-readiness and proves catalog-alignment, direct-answer, temporal honesty, selected-object continuity, truth-gate, human-answer-quality, and meta-context invariants together;
- checked-source failure answers now keep raw MCP transport/internal continuation errors out of the user-facing layer while preserving those details in technical debug artifacts;
- explicit-counterparty incoming-vs-outgoing data-need graphs now select the reviewed value_flow_comparison chain instead of falling back to generic value_flow;
- confirmed metadata-surface follow-ups now promote the surface-grounded chain template to the catalog top match, so neutral catalog drilldowns no longer look like lower-rank planner/catalog divergence;
- live map sync: 20 - planner_autonomy_consolidation_2026-05-01.md

Current honest status:

turnaround implementation progress: ~99%
exit-from-danger-zone readiness: ~97%
pre-multidomain readiness: ~90%
bounded-autonomy foundation readiness: ~89%
open-world bounded-autonomy readiness: ~86%
active Open-World Bounded Autonomy Breadth progress: ~18%, with business-overview evidence fusion and the reviewed business_overview catalog/data-need/planner route-fabric slice locally tested; fresh multi-probe runtime execution is still pending
Post-F semantic integrity module progress: ~99% operationally closed, with remaining risk now treated as next-slice discovery rather than an open blocker inside the closed slice
active inventory-stock breadth slice progress: 100% for the declared scenario pack, not for arbitrary inventory questions
Planner Autonomy Consolidation progress: 100% for the declared module, with catalog-fabric, value-flow arbitration, lifecycle bounded inference, broad-evaluation bridge, inventory catalog templates, inventory runtime-boundary honesty, exact inventory recipe bridging, unambiguous metadata-surface lane inference, catalog chain-template scoring, structured chain-match contract exposure, runtime/debug propagation, subject-aware bidirectional comparison arbitration, structured catalog-alignment verdicts, representative alignment regression guard, catalog-alignment reason-code telemetry, explicit alignment_status propagation, truth-harness/acceptance-matrix surfacing, soft divergence warning, catalog_alignment_ok acceptance invariant, step-level expected catalog-alignment assertions, phase66 and phase32 spec alignment expectations, AGENT source-catalog surfacing, generated phase83 mixed planner-brain replay spec, checked-source user-facing error sanitation, surface-grounded catalog promotion, and guarded live phase83 acceptance validated. Broader unfamiliar 1C asks are now next-module breadth work rather than an open blocker inside this declared slice
graph snapshot after latest rebuild: 5977 nodes, 12983 edges, 137 communities
current regression-gate breakpoint:
- the validated hot paths are no longer structurally broken;
- flagship continuity collapse is no longer the primary risk;
- the main closed-slice regression risk is no longer "A/B/C or D/E/F do not exist", but "already-supported semantic chains can still be contaminated by stale scope, legacy focus state, or wrong post-pivot arbitration";
- pure wording polish remains secondary debt, but semantic integrity and explicit-subject protection are now first-class blockers;
- the practical product risk is no longer only "the route collapsed", but "the user can still occasionally see a semantically wrong answer on a question that the architecture should already support".
main remaining architectural pressure:
- open-world breadth is still narrower than the intended arbitrary 1C blast radius
- planner-selected chains are now real, but still not broad enough to cover unfamiliar 1C asks without additional primitive/search growth
- semantic integrity can still fail on stale carryover, repeated pivots, and mixed scope contamination if those seams are not replay-hardened
- central domain-intent pressure inside resolveAddressIntent()
- replay breadth is still below the future open-world autonomy surface

Latest live proof now includes:

address_truth_harness_phase24_metadata_lane_choice_loop_live_rerun14 accepted
address_truth_harness_phase32_planner_selected_chain_end_to_end_live_rerun2 accepted 6/6
address_truth_harness_phase42_catalog_metadata_drilldown_live_rerun2 accepted
address_truth_harness_phase45_multi_hop_open_total_clarification_loop_live_rerun2 accepted
address_truth_harness_phase67_svk_grounded_counterparty_integrity_live_rerun_vatfix accepted
address_truth_harness_phase68_referential_document_followup_integrity_live_rerun1 accepted
address_truth_harness_phase69_document_to_payments_pronoun_pivot_live_rerun3 accepted
address_truth_harness_phase72_document_to_contracts_year_switch_live_rerun3 accepted
address_truth_harness_phase80_payments_to_contracts_to_documents_all_time_live_rerun1 accepted
address_truth_harness_phase82_human_mixed_integrity_status_dialog_post_m23_rerun_documents_scope_bidirectional accepted 19/19
address_truth_harness_phase82_human_mixed_integrity_status_dialog_post_f_account_injection_guard_clean_scope accepted 19/19, with the Жуковке 51 numeric counterparty suffix kept as counterparty scope instead of leaking as account 51
address_truth_harness_post_f_cross_stage_canary_agent_20260424_live7 accepted 24/24, proving a saved cross-stage AGENT canary across VAT metadata, metadata-scoped organization/document pivots, numeric counterparty suffixes, open-organization value-flow clarification, ranked value-flow year switches, and SVK grounded reset; the saved autorun is AGENT | Post-F cross-stage semantic integrity canary (gen-ag04241406-abe4d8)
address_truth_harness_post_f_manual_failures_20260424_live3 accepted 11/11, proving the manual failure slice from assistant-stage1-9liEOh-7JP: VAT purchase-date, VAT February 2017, highest-value customer, and Chepurnov item-flow after stale inventory context; the saved autorun is AGENT | Post-F ручные провалы VAT revenue item-flow live3 (gen-ag04241710-bdb248)
address_truth_harness_phase11_manual_followup_meta_quality_live_rerun_vatfix accepted 10/10
address_truth_harness_phase20_continuity_stabilization_live_rerun_vatfix accepted 6/6
addressQueryRuntimeM23.test.ts full semantic/runtime slice accepted 403/403 after Post-F VAT/date-basis, scope-recovery, open value-flow organization clarification, document-vs-bank arbitration, and reply-shape hardening
inventory_stock_open_world_breadth_rerun_semantic_integrity_20260501_fix5 accepted all declared inventory stock scenarios: root snapshots, selected-item supplier provenance, supplier overlap, stock aging/unresolved supplier-link, sale trace, and purchase-to-sale chain
address_truth_harness_phase66_human_org_open_scope_dialog_planner_template_rerun2 accepted 7/7, proving catalog-instantiated value-flow still preserves open-organization clarification, year switches, and guarded candidate arbitration
address_truth_harness_phase52_metadata_movement_full_recovery_planner_metadata_scoring_rerun2 accepted 4/4, proving metadata-born movement continuation keeps lane choice, organization, and period recovery intact
address_truth_harness_phase54_metadata_document_full_recovery_planner_metadata_scoring_rerun2 accepted 4/4, proving metadata-born document continuation keeps lane choice, organization, and period recovery intact
MCP planner/catalog consolidation slice accepted locally: assistantMcpCatalogIndex.test.ts + assistantMcpDiscoveryPlanner.test.ts passed 47/47, full MCP-discovery slice passed 227/227 with 9 skipped
lifecycle/value-flow Planner Autonomy response gate accepted: address_truth_harness_phase19_mcp_discovery_response_gate_planner_lifecycle_rerun4 accepted 8/8, proving bounded lifecycle inference, current-turn value-flow aggregate arbitration, and sanitized evidence wording
broad-evaluation bridge continuity accepted: address_truth_harness_phase21_net_followup_after_broad_eval_planner_lifecycle_rerun2 accepted 3/3 and address_truth_harness_phase22_broad_business_evaluation_bridge_planner_lifecycle_rerun2 accepted 3/3
latest local Planner Autonomy slice accepted: full MCP-discovery suite passed 268/268 with 9 skipped; broad MCP/living-chat/route/meaning slice passed 305/305 with 9 skipped; build passed
business-overview route-fabric slice accepted locally: catalog/data-need/planner/pilot boundary slice passed 102/102; the pilot executor exposes business_overview_route_template_v1 as an explicit unsupported scope until the fresh multi-probe bridge exists
inventory template lift accepted locally: catalog/data-need/planner/turn-input slice passed 139/139 with 6 skipped; full MCP-discovery slice passed 276/276 with 9 skipped; build passed; graphify stayed at 5912 nodes, 12833 edges, 138 communities
inventory runtime-boundary hardening accepted locally: runtime-bridge/answer-adapter/pilot-executor slice passed 68/68 with 1 skipped; full MCP-discovery slice passed 277/277 with 9 skipped; build passed; graphify rebuilt to 5913 nodes, 12837 edges, 138 communities
inventory exact-runtime bridge accepted locally: runtime-bridge/answer-adapter/pilot-executor slice passed 70/70 with 1 skipped; full MCP-discovery slice passed 279/279 with 9 skipped; build passed; graphify rebuilt to 5930 nodes, 12884 edges, 135 communities
unambiguous metadata-surface lane inference accepted locally: planner slice passed 36/36; full MCP-discovery slice passed 281/281 with 9 skipped; build passed; graphify rebuilt to 5937 nodes, 12899 edges, 138 communities
live inventory exact-bridge rerun inventory_stock_exact_bridge_live_20260501_after_runtime_bridge is recorded as infrastructure-blocked, not accepted: route/intent/recipe/capability matched, but MCP calls aborted and direct get_metadata timed out while proxy health showed active_sessions_count=0 with pending commands
catalog chain-template scoring accepted locally: catalog/planner slice passed 54/54; full MCP-discovery slice passed 282/282 with 9 skipped; build passed; graphify rebuilt to 5938 nodes, 12903 edges, 139 communities
structured chain-template planner contract accepted locally: planner slice passed 36/36; full MCP-discovery slice passed 282/282 with 9 skipped; build passed; graphify rebuilt to 5939 nodes, 12906 edges, 138 communities
structured chain-template runtime/debug propagation accepted locally: runtime/debug slice passed 18/18; full MCP-discovery slice passed 282/282 with 9 skipped; build passed; graphify rebuilt to 5940 nodes, 12909 edges, 137 communities
subject-aware bidirectional comparison arbitration accepted locally: planner slice passed 36/36; full MCP-discovery slice passed 282/282 with 9 skipped; build passed; graphify rebuilt to 5940 nodes, 12909 edges, 137 communities
structured catalog-alignment verdict accepted locally: planner/runtime/debug slice passed 54/54; full MCP-discovery slice passed 282/282 with 9 skipped; build passed; graphify rebuilt to 5941 nodes, 12911 edges, 136 communities
representative catalog-alignment regression guard accepted locally: planner slice passed 37/37; full MCP-discovery slice passed 283/283 with 9 skipped; build passed; graphify rebuilt to 5942 nodes, 12912 edges, 140 communities
catalog-alignment reason-code telemetry accepted locally: planner/runtime slice passed 53/53; full MCP-discovery suite passed 283/283 with 9 skipped; build passed; graphify rebuilt to 5943 nodes, 12915 edges, 136 communities
catalog-alignment status verdict accepted locally: planner/runtime/debug slice passed 55/55; full MCP-discovery suite passed 283/283 with 9 skipped; build passed; graphify rebuilt to 5943 nodes, 12915 edges, 136 communities
catalog-alignment replay artifact surfacing accepted locally: Python truth-harness/acceptance tests passed 4/4; graphify rebuilt to 5946 nodes, 12918 edges, 136 communities
catalog-alignment divergence warning accepted locally: Python truth-harness/acceptance tests passed 5/5; graphify rebuilt to 5947 nodes, 12920 edges, 138 communities
catalog-alignment acceptance invariant accepted locally: Python truth-harness/acceptance tests passed 6/6; graphify rebuilt to 5949 nodes, 12923 edges, 136 communities
catalog-alignment spec assertions accepted locally: Python truth-harness/acceptance tests passed 7/7; graphify rebuilt to 5951 nodes, 12926 edges, 139 communities
phase66 planner-alignment spec hardening accepted locally: Python truth-harness/acceptance tests passed 7/7; load_truth_harness_spec confirmed expected top matches [value_flow, value_flow, value_flow, value_flow_comparison, value_flow_comparison, value_flow_ranking, value_flow_ranking]
phase32 selected-counterparty planner-alignment spec hardening and AGENT source-catalog surfacing accepted locally: Python replay-tooling tests passed 9/9; load_truth_harness_spec confirmed expected top matches [entity_resolution, value_flow, value_flow, value_flow_comparison, document_evidence, movement_evidence]; regenerated source catalog exposes planner_catalog_alignment as a reusable tag
phase83 mixed planner-brain spec generation accepted locally: Python replay-tooling tests passed 10/10; generated spec has 20 steps and 15 expected catalog top-match checks after the phase19/21/22 alignment hardening; regenerated source catalog exposes planner_catalog_alignment with 26 reusable entries; graphify rebuilt to 5952 nodes, 12927 edges, 138 communities
checked-source error sanitation accepted: targeted answer/candidate/policy tests passed 61/61 with 1 skipped; build passed; phase83_first2_sanity_live_20260501_errorfilter confirms the user-facing assistant section no longer exposes raw MCP fetch failed or Entity-resolution could not continue strings; graphify rebuilt to 5953 nodes, 12930 edges, 137 communities
live-readiness guard accepted: scripts/check_mcp_live_readiness.py --confirm-live --wait-for-polling-seconds 60 now confirms backend/proxy/direct read-only 1C evidence before expensive live replays;
guarded phase83 acceptance: phase83_planner_brain_alignment_live_20260501_readygate_rerun3 accepted 20/20, with 0 warnings, 0 failures, catalog_alignment_ok=true, direct_answer_ok=true, temporal_honesty_ok=true, selected_object_continuity_ok=true, truth_gate_ok=true, human_answer_quality_ok=true, and meta_context_integrity_ok=true;
surface-grounded catalog promotion accepted locally: targeted planner/response-policy/pilot/continuity slice passed 109/109, build passed, and graphify rebuilt to 5973 nodes, 12971 edges, 138 communities.
accepted phase83 is saved as an autorun canary: AGENT | Planner Autonomy phase83: мозг маршрутов, pivots и legacy continuity (gen-ag05011759-6f85fc).

Current architectural reading:

the system is already materially past the dangerous regression breakpoint;
it is now safe for continued architecture hardening and controlled domain-by-domain enablement under replay gates;
it is materially closer to pre-multidomain stability, but still not safe to declare broad low-risk expansion over arbitrary unfamiliar 1C questions.
the practical next target is no longer Post-F rescue itself; it is broader open-world bounded autonomy over 1C evidence while preserving the Post-F semantic-integrity invariants as regression gates.
from this point onward, readiness must be judged not only by route truth and replay pass rate, but also by whether already-supported questions stay semantically correct through stale memory, pivots, clarifications, and mixed scope resets.

For the detailed audit, current percentages, and remaining debt, read:

Architectural Objects Of Planning

This package makes five objects explicit:

state model
transition model
capability contract model
coverage / evidence / truth gate
assistantService extraction plan

These are the objects that should now drive refactoring discussions.

How To Use The Package

Read in this order:

baseline note in docs/ARCH/11 - unified_project_architecture_and_reference_update_plan_2026-04-15.md
01 - project_architecture_baseline_map.md
02 - state_and_transition_contracts.md
03 - capability_contract_spec.md
04 - coverage_evidence_truth_gate.md
05 - assistantService_extraction_map.md
06 - phase_acceptance_matrix.md
07 - external_reference_appendix.md
08 - current_status_audit_2026-04-17.md
09 - pre_expansion_cut_2026-04-17.md
10 - regression_breakpoint_analysis_2026-04-17.md
11 - continuity_stabilization_plan_2026-04-17.md
12 - manual_run_system_analysis_3NilqwT1G2_2026-04-18.md
13 - pre_multidomain_readiness_audit_2026-04-18.md
14 - semantic_dialog_authority_recovery_plan_2026-04-19.md
15 - mcp_bounded_autonomy_reset_plan_2026-04-21.md
16 - data_need_graph_and_open_world_mcp_plan_2026-04-22.md
17 - post_f_semantic_integrity_hardening_2026-04-23.md
18 - post_f_code_documentation_sync_2026-04-24.md
19 - inventory_stock_open_world_breadth_proof_2026-05-01.md
20 - planner_autonomy_consolidation_2026-05-01.md
21 - current_status_canon_2026-05-01.md
22 - open_world_bounded_autonomy_breadth_2026-05-01.md

Planning Rules

Do not treat this package as a rewrite plan.
Do not dissolve AddressQueryService into generic chat logic.
Do not move state back into transcript-only memory.
Do not let answer wording substitute for policy/runtime fixes.
Use scenario-based acceptance as the primary gate for all phases.

Expected Outcome

When this package is fully operational, the project should stop being described as:

"a big custom assistant service with many heuristics"

and start being described as:

"a stateful exact-data assistant with explicit transition contracts and isolated truth gating."

As of 2026-05-01, the project is already materially closer to the target description and is no longer in the same acute collapse state. The remaining blocker is no longer the original continuity failure itself, no longer the A/B/C or D/E/F build-out, no longer the first Post-F rescue slice, and no longer the declared phase83 Planner Autonomy consolidation slice. The active blocker is now the combination of:

unfinished convergence from reviewed bounded MCP chains toward broader open-world autonomy;
continued use of Post-F semantic integrity invariants as regression gates while that breadth grows.

The biggest remaining blockers are:

broader open-world primitive search is still narrower than the future arbitrary 1C blast radius;
dynamic schema traversal is still not broad enough for many unfamiliar 1C asks outside the repaired families;
new stale-scope or post-pivot seams may still appear in future breadth work and must be treated as regression-gated semantic defects, not as wording polish;
residual assistantService overload;
central intent pressure in resolveAddressIntent();
semantic robustness gaps may still appear where already-supported questions meet new wording, typo pressure, short follow-up retargets, or human-answer mismatch.

32 KiB Raw Blame History