hamburg-digital-twin-development-log.md

Hamburg corridor digital-twin development log

This is the concise, continuously maintained engineering log for the Hamburg three-intersection corridor digital-twin work. It records decisions and evidence boundaries that may later become reusable Torii product design. It is not a substitute for machine-readable manifests.

Fixed objective

Build a branched corridor of about three intersections in SUMO, place virtual detectors at the official detector locations, replay official signal control when the same-window history is available, and reproduce the observed traffic counts.

Product target and supporting benchmarks

• The only product corridor is Am Sandtorkai 2349 -> 2394 -> 2403: Großer Grasbrook, Am Sandtorpark, and Osakaallee.
• The first deliverable is its reviewed OSM-derived topology. Signal control, detectors, demand reconstruction, and replay remain downstream stages.
• OSM supplies the continuous road skeleton. Hamburg MAP/OCIT-C/TLD, HH-SIB/HVS, and current official road objects correct or constrain bounded facts; they do not replace the complete road geometry with disconnected local cells.
• Ingolstadt is a human-cleaned teacher/regression benchmark for learning cleanup operations. 5 -> 61 -> 148 and 1923 -> 2363 -> 2150 are validation fixtures. None of these may be promoted as the Hamburg product corridor.
• MCP/NetEdit automation and the estimator-controller-feedback loop are control instruments. A successful tool run is not a topology acceptance result.

Canonical work hierarchy (2026-07-21 reset)

There is one product line, not several parallel experiments:

1. Product target: the Hamburg 2349 -> 2394 -> 2403 corridor digital twin.
2. First product stage -- topology: keep OSM as the continuous skeleton; estimate its defects; use Hamburg official evidence as authority and the same-bbox Ingolstadt human-cleaned network only as a cleanup-action prior; execute bounded edits through Torii/MCP; feed machine and NetEdit audit results back to the next estimator iteration.
3. Second product stage -- observation/control: bind official signal movements, signal history, and colocated real/SUMO detectors to the frozen topology.
4. Third product stage -- traffic state: reconstruct a Saturday two-hour demand, run SUMO, and calibrate against detector counts without hiding missing evidence.
5. Productization: retain only the general estimator/controller/feedback contracts that succeeded on Hamburg and add them to Torii.

Ingolstadt, MCP, NetEdit automation, and the feedback loop are subordinate tools inside the topology stage. They are never alternative deliverables or reasons to leave the Hamburg corridor unfinished.

The accepted road-topology scope now covers 2349, 2394 and 2403. At 2403, the preserve-split physical layout, scoped junction polygons and official-plan lane permissions pass the geometry, connection and runtime gates in iteration 024. Controller ownership and timing remain unresolved because no current MAP/OCIT/TLD package is published. Signal replay and demand validation stay downstream of that separate evidence gate.

Reusable feedback loop

flowchart LR
    A["Frozen OSM source"] --> B["Estimator: OSM snapshot + 3 IntersectionIRs + canonical SUMO snapshot"]
    B --> C["Findings and bounded hypotheses"]
    C --> D["Controller: reversible edit ledger"]
    D --> E["Materialized candidate .net.xml"]
    E --> F["Feedback: exact diff + geometry + SUMO + NetEdit + official evidence"]
    F -->|"improves without hard regression"| B
    F -->|"fails or regresses"| G["Rollback and re-plan"]
    G --> C

Loading

NetEdit loads only the exact candidate .net.xml. Its companion state is a bundle of Torii's existing contracts: one IntersectionIR per target junction, a corridor CanonicalNetworkSnapshot, a WorkflowState control summary, and a hash-closed ArtifactManifestV1 containing the edit ledger, rollback, audits, parent candidate, and promotion trace.

Estimator state bundle

iteration_N/
  candidate.net.xml                    # exact SUMO/NetEdit payload
  estimator/
    source-osm.snapshot.json           # immutable downloaded OSM facts
    osm-derivation.json                # source -> filtered OSM provenance
    sumo.snapshot.json                 # exact imported SUMO facts
    osm-sumo-lineage.json              # source IDs and confidence
    canonical-network.snapshot.json    # corridor-level semantic state
    intersections/<node>/intersection_ir.json
  controller/
    workflow-state.json                # concise control-console summary only
    promotion-trace.json
  feedback/
    exact-semantic-regression/         # source/candidate diff and safety gates
  edit-ledger.json                     # reversible proposed edits
  rollback.json
  artifact-manifest.json               # authoritative paths, hashes, and DAG

No single JSON file replaces the network. The .net.xml remains the executable payload; the estimator bundle makes its source, semantics, uncertainty, and change history explicit. Only hash-bound artifacts produced by known Torii adapters may influence promotion. WorkflowState is never promotion authority. When two files contain the same canonical snapshot bytes, the authoritative artifact DAG keeps one identity instead of treating the copy as new evidence.

Evidence and authority order

1. Hamburg MAP/OCIT-C/TLD: movement, signal-group, and signal-state authority.
2. Current Hamburg HH-SIB/HVS and official road objects: road-axis, direction, stationing, carriageway, and lane-inventory authority.
3. Hamburg cross-sections and detailed mapping: independent structured topology corroboration; their survey/validity date remains a mandatory gate.
4. OSM: acquisition and candidate-topology input only.
5. Google Maps/Street View, Hamburg DOP, and NetEdit screenshots: visual review only; they cannot authorize automatic promotion.
6. Torii/SUMO load, routeability, smoke traffic, and detector fit: diagnostics and runtime validation only; they do not prove real-world topology.

If structured official sources conflict, the candidate remains blocked. Missing evidence is never replaced with a guessed connection or signal program.

Progress log

Date	Step	What changed or was checked	Result	Reusable Torii design
2026-07-11 to 2026-07-19	Initial Am Sandtorkai corridor	Reused Torii OSM cleanup, compact-corridor extraction, MAP/KML/OCIT parsing, detector mapping, routeSampler, SUMO E1/E2, signal replay, NetEdit review, and geometry audits.	Flow reconstruction worked, but incomplete 2403 signal assets and several topology/geometry defects blocked a complete twin.	Keep source snapshots immutable; candidate networks are versioned; every stage fails closed and records hashes.
2026-07-19	Compound-junction handling	Classified complex junctions separately from their signal controller. In particular, one controller no longer implies one merged SUMO junction.	Removed the unsafe whole-junction merge assumption and large overlapping junction shapes.	Represent skeleton, channelization, physical owners, and controller domain as separate dimensions.
2026-07-19	Alternative corridor screen	Selected 1923 -> 2363 -> 2150 because all three nodes have official MAP/KML/OCIT assets and a complete Saturday detector window.	Suitable as a topology and replay validation corridor; it does not replace the named Am Sandtorkai product target.	Add an automatic corridor-selection gate over signal assets, counts, distance, and official road continuity.
2026-07-19	Official road/MAP splice	Reused Torii lane-axis stitch, splice planner/materializer, PlainXML, netconvert, surface-overlap audit, Connection Mode audit, and SUMO load checks.	Candidate v26 loads, has zero structural connection failures and zero surface overlap, but generic Connection Mode still reports 27 semantic findings.	PlainXML is the editable representation; final .net.xml is always regenerated by netconvert.
2026-07-19	Exact movement and TLS binding	Bound official MAP movements to physical SUMO connections and TLD streams using exact lane keys and signal groups.	19/19 movements are routeable; 19/19 signal bindings are available, including two explicitly audited same-signal-group projections.	Signal-group projection is opt-in, same-node/same-group only, and retains provenance.
2026-07-19	Detector-constrained replay	Reused routeSampler and detector writers. Added an explicit detector-boundary diagnostic mode and lane-positioned departures.	Validation corridor achieved a near-exact detector replay, but this is a local detector-section twin rather than full OD proof.	Distinguish boundary detectors from internal detectors; never convert internal lane counts into departure-lane facts.
2026-07-20	Topology-first reset	Reviewed the remaining topology findings before any further traffic calibration. Split 27 generic findings into official MAP core findings, expected splice findings, and 16 unresolved HH-SIB axis semantics.	Runtime success is no longer accepted as topology proof. The network remains diagnostic until official facts cover each required transition.	Add an independent official-topology promotion gate after generic Torii/SUMO gates.
2026-07-20	Official cross-section refresh	Re-downloaded the corridor cross-sections from Hamburg OGC API Features with complete pagination and EPSG:25832. Snapshot: 3,504/3,504 features, SHA-256 1cea3b2ec6f4c2357a90d2ecce34a1d764017542621835e8db51fb9a02b7a341.	Acquisition is complete, but the dataset describes generalized surfaces from the 2016 survey and cannot alone prove the 2026 layout.	A source adapter must separate acquisition validity from claim-time validity and preserve raw strip types/widths.
2026-07-20	Google Maps cross-check	Reviewed 1923 and 2363 Street View and 2150 satellite imagery.	Gross channelization and compound layouts are visually corroborated; exact lane-to-lane transitions remain unproved. The artifact is explicitly review_only.	Image sources must be a separate evidence class with promotion_authority=false.
2026-07-20	Gate design safety review	Prototyped an official-topology promotion gate, then audited whether its claims were derived from the candidate and from trusted source adapters.	The first prototype was rejected: it could compare caller-reported candidate_value and source metadata without independently deriving either fact. It is not a production gate and cannot promote a network.	Promotion facts must be extracted from the exact candidate by a known fact extractor and from schema-specific official adapters; hashes and allow-listed labels alone are insufficient.
2026-07-20	Real cross-section adapter run	Ran the new adapter against the frozen OAF response.	Acquisition passed with 3,504 features and 287 station profiles; claim status is review_required because no official 2026 validity interval is declared; promotion stays blocked.	Keep acquisition completeness, temporal validity, classification, and topology authorization as separate gates.
2026-07-20	Topology fact ledger	Bound all 27 v26 Connection Mode findings to the candidate SHA and separated MAP-core, official-merge, and HH-SIB-axis findings.	The ledger contains 6 MAP-core, 5 merge-splice, and 16 axis facts. None is currently independently promotable; 4 axis facts have historical support, 4 conflict with that inventory, and 8 lack exact lane-transition evidence.	Promotion should operate on explicit fact IDs rather than a network-level pass flag. Runtime and visual checks remain downstream diagnostics.
2026-07-20	Current official geometry search	Queried the current Hamburg OAF services for Kreuzungsskizzen and Feinkartierung Straße over the complete validation corridor, froze six paginated snapshots, and recorded hashes and per-object dates.	The live sketch service covers all three nodes, but its node drawings are dated 2011, 2011, and 2021. Fine-mapping contains newer 2023/2025 road-space objects, but mixed dates and no legal lane-transition semantics. Both remain review_required.	Reuse official OAF snapshots as independent geometry evidence, while evaluating validity per feature and keeping physical surfaces separate from legal movements.
2026-07-20	Focused regression after topology work	Ran the topology, signal-binding, splice, and source-adapter regression set plus Ruff and whitespace checks.	72 focused tests passed; the safety review still blocks topology promotion because passing tests included a now-rejected self-reported evidence path.	Test success validates implementation behavior, not the truth of field topology; promotion tests need adversarial candidate/source fixtures.
2026-07-20	v26 background NetEdit review attempt	Reused Torii's existing no-focus NetEdit screenshot workflow against the exact v26 candidate.	The workflow failed closed before capture because the foreground Windows keyboard layout was Chinese (0x0804). No old-candidate screenshots were reused and no system setting was changed.	GUI review must be candidate-hash-bound and environment-safe; a blocked GUI review remains a blocked review, not an excuse to reuse stale images.
2026-07-20	v26 English-layout NetEdit review and rejection	Re-ran the hash-bound background review after the keyboard was switched to English. All 12 views were captured without changing the candidate, but the images show bent approach bundles, abrupt joins, and oversized/overlapping junction surfaces.	v26 is rejected as a topology/geometry candidate. Earlier automated “zero overlap” and load checks were insufficient and must not promote it.	Add source-geometry parity and visual-shape gates; a SUMO-loadable network can still be geometrically wrong.
2026-07-20	OSM-baseline architecture reset	Traced the bad geometry to using local Hamburg MAP cells and HH-SIB axes as a replacement road skeleton.	Stop repairing v26. Rebuild from a cleaned, tightly cropped OSM corridor; use Hamburg official sources as authoritative semantic/correction overlays rather than as the whole-road geometry generator.	Preserve OSM road curvature and continuity outside bounded junction cells; official MAP/OCIT controls lane count, movement, TLS, and detector identity.
2026-07-20	Estimator/state-contract reuse	Audited Torii's existing state models before adding another format. Reused IntersectionIR for each raw OSM junction, CanonicalNetworkSnapshot for the exact SUMO/NetEdit network, WorkflowState for the control summary, and ArtifactManifestV1 plus the edit ledger for hashes, dependencies, rollback, and promotion.	Removed the duplicate experimental osm_feedback_state contract because arbitrary JSON could have influenced its quality result without schema/candidate-hash validation.	A network iteration is a bundle of existing contracts, not a new monolithic network language; .net.xml remains the NetEdit payload and the canonical snapshot is its stable semantic companion.
2026-07-20	OSM feedback iteration 000	Ran the existing OSM cleanup workflow on a frozen bbox source with all automatic topology repair disabled, then ran connectivity, topology, routeability, Connection Mode, surface, rendered-edge, and background NetEdit audits.	Baseline is construction-invalid: 30 suspicious topology clusters, 43/50 vehicles arrived, 6 teleports, 108 lane-to-non-owner-junction overlaps, 56 junction-to-junction overlaps, and 5 rendered-edge overlap warnings.	The estimator observes an immutable candidate first; no controller action is allowed before a complete baseline quality vector exists.
2026-07-20	OSM feedback iteration 001	The provenance estimator found that bbox filtering changed 40 retained OSM ways. Rebuilt from the same frozen source with complete source ways, using the existing clip_source_ways_to_bbox=False switch.	Retained ways increased 471→499 and changed retained ways fell 40→0. Overall quality did not improve consistently: topology clusters 30→34, arrivals 43→36, teleports 6→3, rendered overlaps stayed 5. Exact canonical regression remains blocked.	A controller may accept an action for one layer (source fidelity) while rejecting candidate promotion. A/B decisions require per-dimension before/after deltas, not one weighted score.
2026-07-20	Raw OSM intersection observation	Generated Torii-native, non-materializing IntersectionIR artifacts for OSM nodes 1923, 2363, and 2150.	OSM reports 3, 3, and 4 passenger approaches respectively, while support paths inflate total approaches to 8, 4, and 12; 1923/2150 retain 4/8 internal fragments. These are estimator findings, not edit authorization.	Keep motor-vehicle skeleton, support modes, physical fragments, and signal controller evidence as separate state dimensions before proposing joins.
2026-07-20	OSM feedback iteration 002	Reused Torii's compact-corridor extractor on the complete-way import, then compared the parent and compact candidates with one frozen route file.	External edges fell 636→105 while all 50 frozen routes completed in both networks with zero teleports/collisions. The compact network remains construction-invalid: 14 Connection Mode reviews, 21 lane-to-non-owner overlaps, 7 junction overlaps, and 2 rendered overlaps.	Scope extraction is a new hash-bound baseline, not a topology promotion. Later edits compare against this compact baseline.
2026-07-20	Control-anchor estimator correction	Ran Torii's existing composable archetype estimator separately on the 1923 signal anchor and its adjacent road junction, then checked Hamburg's official LSA type.	Signal anchor 343634175 is an A2 crossing candidate, adjacent junction 312938329 is A3, and official node 1923 is F-LSA. They must not be merged.	Model physical conflict nodes, stop-line/signal anchors, and controller domains as separate identities. One controller or official point is not one SUMO junction.
2026-07-20	Corridor type gate	Re-ran corridor screening with the product policy required_signal_types=[K-LSA].	Candidate 1923→2363→2150 is automatically blocked because its official sequence is F-LSA→K-LSA→F-LSA.	Candidate screening must verify official control type before topology reconstruction starts. Validation corridors may use other types only when declared explicitly.
2026-07-20	Safe 1923 controller action	Wrote a Torii edit ledger with one review_marker and a no-change rollback; no network mutation was materialized.	Current evidence only supports preserving the F-LSA anchors and the real side-road junction. A future shape-only probe must keep coordinates, edges, connections, TLS/linkIndex, and route semantics unchanged.	The controller may output a deliberate no-op when the estimator cannot authorize an edit; no-op is a valid feedback action, not a stalled workflow.
2026-07-20	Hash-closed iteration bundle	Reused ArtifactManifestV1 and added a public file-to-artifact identity constructor. Bound the OSM source, SUMO source/candidate, IntersectionIRs, archetype profiles, official node type, controller state, route A/B, geometry audits, edit ledger, rollback, and overlay.	Iteration 002 now validates as 35 artifacts and 36 closed dependencies; source mutation is false. Hard gates still fail/stop promotion.	The reusable state format is an artifact DAG around the exact .net.xml, not a second executable network format.
2026-07-20	Three-K-LSA rescreen	Reused the official LSA, MAP/KML/OCIT, count, and road-axis screen and selected 5 → 61 → 148 (Lübecker Straße/Lübeckertordamm/Steindamm).	All three are official K-LSA; current static signal assets bind at all three nodes and 52/52 count streams cover the selected Saturday window.	Corridor selection is an evidence gate before OSM topology work, not a late manual choice.
2026-07-20	New OSM estimator baseline	Downloaded one immutable OSM bbox, kept complete selected ways, disabled all automatic topology edits, and generated an exact SUMO/NetEdit candidate plus canonical and per-junction observations.	Baseline is correctly blocked: 18 suspicious clusters, 36 Connection Mode reviews, 47 junction overlaps, 78 lane/junction overlaps, 5 rendered-edge warnings, 23 teleports, and 2 collisions.	The estimator must describe defects before the controller proposes an edit; SUMO loadability alone is not a pass.
2026-07-20	Control-domain/physical-cell observation	Compared each official LSA point with nearby OSM branch nodes and ran the existing composable archetype estimator.	Nodes 5 and 148 expose four-arm compound candidates; node 61 exposes a five-approach/slip-link candidate. All remain review-only and no official point is treated as a physical OSM node.	Control identity, signal anchors, physical conflict cells, and SUMO junction owners stay separate in state.
2026-07-20	Manifest byte re-verification	Added a general ArtifactManifestV1.assert_artifact_files_unchanged() guard and tamper/missing-file tests.	The controller can no longer rely on stale manifest hashes when reading a prior iteration.	Every feedback iteration rehashes the full artifact DAG before using it as promotion evidence.
2026-07-20	Model-directed NetEdit merge decisions	Used the exact OSM candidate, official intersection classifications, and Torii's hash-bound background NetEdit Inspect/TLS/Connection views to decide the three physical cells directly. Materialized only the two evidence-backed join hypotheses with the existing junction aggregation code.	Node 5: the complete four-node core is logically one junction, but the Torii materializer introduces 18 target-surface findings and a Connection Mode regression. Node 61: preserve the channelized multi-node layout. Node 148: the two-node join passes geometry and connection gates, but remains topology-only because the repaired route diagnostic increased collisions from 4 to 6 before official K-LSA restoration.	Keep model/visual intent separate from execution quality. Background NetEdit review is valid for screenshots, not editing: the controller needs real input plus post-edit machine evidence before accepting any GUI candidate.
2026-07-20	NetEdit session MCP design	Compared common Blender, FreeCAD, and Unity MCP interaction patterns: bind one scene/document, observe object and viewport state, perform one atomic edit, then save or undo. Mapped that pattern to one focused sumo_netedit_session(open|observe|act|finalize|abort) tool rather than a general desktop-control API. Official NetEdit operations remain the vocabulary: I/C/T/M/S/D/E modes, F5 recompute, F7 join, Enter accept, Esc abort, Ctrl+S save, and Ctrl+Z/Y undo/redo.	Each action is window-targeted and bound to the caller's exact last recorded screenshot SHA plus a fresh live capture. open copies the immutable source to a separate candidate; observe records the client-coordinate viewport; act permits one allow-listed mouse/shortcut edit; finalize saves, rehashes, and runs SUMO/geometry/connection gates; abort closes without promoting.	Reuse Torii's launcher, screenshot, hashing, SUMO-load, surface, and Connection Mode audits. Preserve an action ledger and before/after screenshots; GUI success can never promote a network automatically.
2026-07-20	Real NetEdit input transport probe	Tested background Win32 SendMessage delivery and real foreground mouse/keyboard input against the same four selected node-5 nodes. SendMessage reported success but did not execute F7 or change the file. Real target-window input did execute F7 and save a changed candidate.	The genuine four-node join produced an oversized central junction polygon, so the candidate was rejected rather than promoted. This also proves that transport acknowledgement is not edit evidence; file hash, screenshot, and machine audits are required.	The MCP must use real target-window input, restrict control to the launched NetEdit process, and bind every action to observable state. It exposes no arbitrary OS-wide input operation to callers and accepts no message-only success criterion.
2026-07-20	Human-cleaned Ingolstadt comparison	Parsed the current TUM sumo_ingolstadt network as a methodology and geometry benchmark. Its NetEdit-saved network contains 912 cluster_* junctions: 619 join two source nodes and 75 join four. A direct ownership audit finds 123 active controllers: 121 control one physical SUMO junction and only two control two separated junctions, about 73 m and 89 m apart.	The dominant human-cleaned pattern is therefore to merge one physical conflict core, retain short channelized approach/storage edges outside that core, and use a shared controller only for genuinely separated conflict areas. The network also contains explicit NetEdit-created nodes/edges and hand-shaped lane/junction geometry, so a raw automatic join is not the method.	Treat node 148's two-node core as an Ingolstadt-style join candidate. Treat node 5's four-node core as one logical conflict core in principle, but reject the current F7 output: Torii must rebuild its bounded junction envelope and explicit lane connections without swallowing nearby approach fragments. Node 61 remains split because its storage/slip geometry describes separated physical cells.
2026-07-21	Same-location Ingolstadt OSM A/B	Downloaded current OSM at two published Ingolstadt junctions, a 2023-08-02 Overpass snapshot for the four-node case, and OSM API history for the two-node core way. Compared those inputs with the exact human-cleaned SUMO network at commit e0a95de.	The two-node OSM core (one internal short segment) becomes one cluster with six retained external edges, 20 internal edges, and 12 controlled via connections. The four-node OSM ring (four internal segments and eight outside neighbours) becomes one cluster with nine retained external edges, 41 internal edges, and 19 controlled via connections.	The reusable operation is not distance-only joining: identify the complete conflict core, absorb its OSM micro-edges, preserve the external approach skeleton at explicit cuts, then rebuild internal lane movements. Hamburg node 5's prior four-node F7 selection was incomplete; its official a1-a4 cuts enclose 16 OSM/SUMO micro-nodes, which must be tested as a new bounded cluster hypothesis.
2026-07-20	Public MCP real-input and audit proof	Hardened the grouped tool with client-coordinate screenshots, queried Per-Monitor-V2 DPI evidence, a fresh screenshot check immediately before every input/save, exact PID/HWND focus checks, and partial-input cleanup after transport failure. Ran open → observe → click → observe → abort through the registered FastMCP server; separately ran open → observe → F7 → observe → F5 → observe → finalize through the public Python adapter.	The mouse smoke used a 1400×1000 client image at 144 DPI and preserved the source. Normal NetEdit animation changed 0.295% of global pixels, while the protected 49×49 click region stayed identical. The F7/F5 edit changed only the candidate and loaded in SUMO, but the reused surface audit found 18 introduced/focus overlaps and Connection Mode reported new_target_scope_review_findings; promotion remained blocked.	Adopt the 3D-editor MCP loop, not its arbitrary-code surface: observed viewport plus persisted object state, one atomic edit, undo/abort, explicit save, then independent machine feedback. A successfully delivered GUI action is never equivalent to a valid road network.
2026-07-20	NetEdit MCP safety closure	Froze GUI settings and selections into hash-bound session snapshots; required F7 to be the first edit with selection junctions exactly equal to the declared source scope; rejected injection while any physical key or mouse button is held; constrained the MCP schema with literal operation/action/object enums; and made wrong-session errors non-destructive. Semantic shortcuts/save require exact live viewport equality. Finalize binds source/candidate hashes around SUMO-load, surface, Connection Mode, junction-identity, outside-scope preservation, and audit-integrity checks.	Even when every machine gate passes, a GUI candidate remains review_required; a failed geometry/connection/scope gate propagates as top-level fail. Validation mistakes remain observable/retryable, while uncertain delivery/focus/evidence failures abort the target session.	This is the reusable control-system boundary: immutable inputs, one atomic controller action, explicit state observation, independent machine feedback, and no automatic promotion. The local screenshot artifact is viewable by Codex, but generic MCP ImageContent remains a later interoperability enhancement.
2026-07-20	Hardened FastMCP F7 feedback run	Re-ran open → observe → F7 → observe → F5 → observe → finalize through the registered FastMCP server with the frozen four-junction selection and declared replacement ID. The session used a fresh evidence directory and preserved both the source and frozen preload hashes.	F7 selection lock, junction identity, SUMO load, and audit-integrity gates passed. Surface comparison and Connection Mode failed, so the top-level result was fail; the screenshot still shows the oversized central polygon. Candidate SHA-256 is fe151ba844defdb380e31219c2ded4786a4886175b97ab21202c6c37bf570454, audit-summary SHA-256 is 9cf71b273ade0523fe847a27cccd4481de36280c725344eb8ccc25cbf2c2dd0a.	The larger MCP can execute and observe a real NetEdit hypothesis, but the feedback controller—not input delivery—decides acceptance. This four-node merge remains rejected and cannot be reused as cleaned topology.
2026-07-20	Full physical-key safety probe	Re-ran the final FastMCP path after expanding the input preflight from modifiers to every Windows virtual key. The first F7 attempt was blocked by an exact-viewport change; subsequent attempts were blocked because Windows continuously reported VK_SUBTRACT (0x6D) as physically down.	No F7 was delivered, the candidate SHA remained equal to the frozen source, and only the NetEdit process created for this test was stopped. Torii did not synthesize a key-up or bypass the user's input state.	Validation/preflight failures are retryable, but physical input ownership remains fail-closed. A GUI controller must never mix an unledgered human key with an MCP action merely to complete a smoke test.
2026-07-20	Outside-scope preservation re-audit	Re-audited the previously saved F7 candidate with the new exact outside-scope XML gate. It compares non-target junctions, non-incident external edges and lanes, out-of-scope connections, unaffected TLS programs, and remaining top-level network state.	The scope gate passed with 244 outside junctions, 320 outside edges, 1,438 outside connections, and 62 unaffected TLS programs unchanged. The overall candidate still failed because the bounded surface and Connection Mode gates failed.	The rejected merge was locally bounded but geometrically wrong. Scope preservation and local geometry correctness are independent feedback dimensions and both are required.
2026-07-20	Post-review MCP boundary hardening	A final adversarial review found three boundary gaps: candidate-side incidence could expand the declared edit scope, a delivered action followed by screenshot failure could remain unledgered, and the public MCP accepted a caller-supplied NetEdit executable name. Froze external edge/TLS scope from the source network, required exact declared-junction ownership for candidate internal edges, made every post-delivery evidence failure abort the session, removed the executable parameter from the MCP schema, and made unknown audit statuses fail closed.	The focused suite increased to 71 passing tests. New regressions cover candidate reattachment, numeric-ID substring collisions, unrelated/candidate-injected internal-edge mutation, unknown audit states, and post-input/post-save evidence loss.	A GUI feedback controller must freeze authorization before editing, separate pre-input retryability from post-input uncertainty, and never expose an executable-selection surface to MCP callers.
2026-07-21	Node 5 complete conflict-core estimate	Combined the same-location Ingolstadt OSM/cleaned comparison with Hamburg MAP a1-a4 cuts and the exact OSM graph.	The physical core is exactly 16 OSM/SUMO nodes and 16 internal OSM edge fragments; the eight external OSM approaches remain the immutable road skeleton.	A join proposal is a bounded conflict-cell hypothesis, not a list of short edges. The estimator must output the complete absorbed core and explicit retained cuts before the controller may rebuild it.
2026-07-21	Teacher replay outside-scope correction	Traced a 10.86 m boundary gap to full-network netconvert clipping an ordinary off-scope edge while Torii restored only internal artifacts. Reused the existing off-scope restore path, extended it to ordinary lane geometry and edge centerlines, and limited geometry blending to the local conflict-cell endpoint.	Iteration 034 has zero outside-scope structural regressions, zero introduced surface overlaps, SUMO load pass, and exact 25/25 official TLS movement parity.	OSM owns remote approach endpoints and long-road geometry; the official teacher owns core-side endpoints, lane cardinality, movements, internal/via topology, and TLS/linkIndex. Normalization side effects outside that split authority are restored or rejected.
2026-07-21	Ingolstadt-style lane-transition boundary repair	Identified downstream node 9799586527 as a real, non-TLS Steinhauerdamm/B75 3→2 lane transition rather than part of the node-5 conflict core. Added an evidence-authorized Torii controller that keeps the node, edges, lanes, and connections unchanged and replaces only its polygon with the convex hull of adjacent lane endpoints.	Iteration 037 resolves the inherited 4.815165 m² overlap with zero introduced findings. Topology SHA is unchanged; both retained movements (1→0, 2→1) arrive with zero teleport/collision; background NetEdit captures are hash-bound.	Never enable global --junctions.minimal-shape as a cleanup shortcut. The estimator classifies a linear lane transition, the controller applies one local shape action, and surface/Connection/SUMO/smoke/NetEdit feedback decides whether to retain it.
2026-07-21	Existing Ingolstadt runner optimization	Audited run_ingolstadt_corridor_teacher.py before adding new code. Kept its original bounded single-junction mode, then added an opt-in reference-matched mode that delegates to Torii's existing full OSM cleanup workflow.	The new mode preserves raw OSM, human reference, aggregation candidates, and teacher replay candidates as separate hash-bound layers; it rejects an explicit candidate net because full comparison must rebuild from OSM. The current same-bbox audit locates 132 of 149 human join cases, but still labels them with one coarse action family. Teacher materialization remains separately opt-in, and join confirmation now uses exact status tokens so unconfirmed cannot be accepted by substring.	Use this runner as an estimator of human edit patterns, not as a Hamburg generator. Its next output contract must describe absorbed core nodes/edges, retained boundaries and storage segments, applicable conditions, and counterexamples; Hamburg official evidence decides whether an action transfers.
2026-07-21	Ingolstadt teacher-action contracts	Reused the existing full reference-join report instead of adding another matcher. Projected every human case into an action contract containing absorbed source nodes/internal edges, retained OSM boundaries, applicability evidence, excess/missing scope and rejection reasons. The v6 runner additionally binds the exact bbox and SHA-256 of the source OSM, TUM teacher, comparison net and join report, and verifies source/teacher hashes before and after execution.	The frozen same-bbox rerun succeeds as an estimator: 149 contracts comprise 29 exact bounded conflict-core priors, 61 source-identity reviews, 42 incomplete-source abstentions and 17 unmatched abstentions. The earlier count of 77 was too broad because 48 candidates extended beyond the human teacher core; those cases are now withheld. execution_status=pass, input_parity=pass, evidence_status=review_required, and promotion_gate_status=blocked.	Feed only the 29 exact-core cases to the Hamburg estimator as non-authoritative structural priors. Reviews and abstentions are negative controls. A contract may propose one bounded edit, but Hamburg official geometry/movement evidence and the full feedback gates must independently authorize it.
2026-07-21	Product-scope reset	Re-audited the accumulated Hamburg artifacts after validation corridors had started to dominate the work plan.	Locked the only product target back to Am Sandtorkai 2349→2394→2403; Ingolstadt and the two alternative corridors remain method fixtures only.	Every iteration and promotion report must carry the product scope ID; a validation fixture cannot become the product target through workflow reuse.
2026-07-21	Named-corridor teacher-free estimator	Reused the existing bbox signal-cell discovery on the immutable Am Sandtorkai OSM snapshot, without a teacher, materialization, or expected topology.	Found 24 cells from 96 signal anchors. Near the targets, 2394 has one consistent suggestion, while 2349 has three overlapping interpretations and 2403 has five. Legacy fixed-radius IntersectionIR views inflate approaches, so they are observations rather than cell-selection authority.	Conflict-cell discovery, fixed-radius IR, official identity, and controller domain are separate estimator channels; disagreement must produce abstention, not a merge.
2026-07-21	2403 controller abstention	Wrote a Torii corridor edit ledger containing one hash-traceable review marker and a no-change rollback for the 2403 OSM cell.	The ledger validates and explicitly requires review; no source/candidate network was modified or generated. Five overlapping cell hypotheses, storage-capable connectors, and missing current MAP/OCIT/TLD movement assets block a join or TLS installation.	A no-op is a valid controller action. The feedback loop advances its knowledge state without forcing a topology mutation when the estimator is non-identifiable.
2026-07-21	Named-corridor OSM lineage and approach restoration	Rebuilt the compact Am Sandtorkai OSM candidate with --output.original-names, then compared every 2349/2394 official MAP lane against the retained OSM approaches.	The estimator found that the earlier crop had removed complete OSM approach chains 61649647#0/#1/#2 and 127716467#0/#1. Restoring only those five source segments changed the official lane binding from three low-confidence lanes to 21/21 active, high-confidence bindings.	Do not redraw the road surface or replace it with MAP geometry. Preserve complete OSM ways and use official MAP evidence to detect missing cuts and authorize the smallest restoration.
2026-07-21	2349/2394 official static topology candidate	Applied the existing Hamburg compound topology adapter to the restored OSM skeleton. For 2349, joined only the two 0.20 m opposing micro-cores; for 2394, preserved the compound multi-owner layout.	The first candidate had 16/16 official movements and exact external-lane geometry, but Connection Mode found six OSM routing connections outside the complete official vehicle inventory. The adapter had demoted their signal ownership without removing their routing permission.	Treat every successful materialization as a controller hypothesis. The feedback audit can still reject it when the official movement contract and the remaining OSM lane graph disagree.
2026-07-21	2349/2394 feedback-corrected topology	Reused the existing 2394 routing-removal inventory, added the one equivalent 2349 removal, and wrote SUMO's official <delete> connection patch form. Added exact official evidence hooks for two lane fan-outs and one destination merge.	Iteration 009 removes exactly six unsupported connections, adds exactly five MAP-supported connections, preserves 21/21 lane bindings and 16/16 official movements, keeps all 133 external lane geometries unchanged, and loads in SUMO. Iteration 010 Connection Mode passes with zero new target or outside-scope findings. Hash-bound background NetEdit Inspect/TLS/Connection captures for 2349 and 2394 preserve the candidate SHA and show no long cross-block links or lane-surface distortion.	Freeze SHA d87eeac6a0f6c5c082f59917fc40e8ead71d5e97cfcda565301bd7c1b8871913 as the accepted 2349/2394 topology sub-scope. This does not promote the whole corridor or claim historical signal timing.
2026-07-21	2403 physical-layout evidence	Compared the OSM core with Google satellite imagery and the same-bbox Ingolstadt teacher cases C046, C058, and counterexample C039. The imagery is review-only; C046/C058 are structural precedents, not Hamburg truth.	All sources support preserving the divided-road approaches, storage/channelization links, and multiple physical owners. They reject the proposal to collapse the complete 2403 compound into one large junction. C039 is rejected because its source/teacher identity is unstable and its join would swallow independent stoplines.	A shared controller never implies one physical SUMO junction. Teacher transfer needs an explicit retained-boundary/stopline contract; otherwise the controller must choose no-op or a geometry-only repair.
2026-07-21	2403 scoped junction-shape repair	Rejected the global netconvert --junctions.minimal-shape probe because it changed all 133 external lane geometries. Added a reuse-first Torii controller that accepts an explicitly enumerated junction set, reuses the existing external edge/lane and Connection Mode signatures, verifies hash-bound TLS semantics, and copies only junction shape/customShape from the same-topology reference.	Iteration 016 changes six authorized 2403 polygons and nothing else: external lane deviation is 0 m, normal/controlled movement deltas are 0, SUMO load and Connection Mode regression pass, and local surface findings fall from 11 lane + 5 junction overlaps to 0 + 0. Iteration 017 background NetEdit review used an English keyboard, no global input, and preserved candidate SHA 9e4ba2580f7f1ff2e0cb5be5d453f88d6075f50d100364139ac8eb275069b11a.	Treat SUMO minimal-shape output as a geometry reference only, never as the candidate network. Promotion remains review_required: official 2403 MAP/OCIT movements and controller ownership are still missing, so this closes polygon geometry but not legal movement or signal topology.
2026-07-21	2403 provisional movement inventory	Reused Torii's external-lane graph and bounded BFS on the immutable iteration-016 candidate instead of writing another path finder. Enumerated four ingress edges, four egress edges, seven internal channelization edges, 33 physical lane connections and 17 bounded ingress-to-egress lane paths. Cross-checked the source OSM way tags and restriction relations.	The Am Sandtorkai ingress has three lanes and turn:lanes=left|through|through;right, which agrees with its current left/through/right lane branching. Brooktorkai is internally inconsistent: OSM declares four lanes but supplies five turn-lane tokens. No core turn-restriction relation resolves that conflict.	Keep geometry frozen. Convert each ingress lane's OSM token/path comparison into estimator evidence; auto-accept only exact lane-count/order matches, retain the Brooktorkai movements as provisional, and never treat the absence of a restriction relation as official permission.
2026-07-21	2403 official Brooktorkai lane/movement evidence	Compared the frozen OSM/SUMO approach with the current LSBG March-2024 Brandstwiete/Bei St. Annen plan, the implemented July-2022 Am Sandtorkai/Brooktorkai plan and current satellite imagery. The government plans are authoritative; imagery is only a visual cross-check.	Both plans show one separate cycle lane plus exactly four Kfz lanes at the Brooktorkai stop line. The five-token OSM turn:lanes value is stale/malformed and must not create a fifth motor lane. In SUMO right-to-left lane order the supported permissions are lane 0 through;right, lane 1 through, lanes 2-3 left. The frozen candidate already contains every one of these movements except 25444569#2 lane 0 -> 395272650#0 lane 0 (straight).	Preserve all four lane geometries and the compound physical owners. Apply one hash-bound PlainXML connection addition, then require exact outside-scope preservation, zero surface regressions, Connection Mode pass, SUMO load/routeability and background NetEdit review. This can promote legal road topology only; 2403 signal ownership/timing remains blocked without MAP/OCIT/TLD evidence.
2026-07-21	2403 feedback-loop movement closure	Rejected iteration 018 because whole-network netconvert moved 36 lanes, and rejected iteration 019 because restoring the external OSM geometry left 19 internal path endpoint gaps. Added a reusable scoped Torii primitive that restores existing internal movement geometry from the accepted parent and reanchors only the declared new movement to immutable external-lane endpoints.	Iteration 024 preserves all 133 external lane shapes and lengths exactly, adds exactly one ordinary movement and no controlled movement, introduces no surface finding, passes SUMO load and Connection Mode with the hash-bound engineering-plan merge evidence, and sends 10/10 movement-smoke vehicles through with zero teleport/collision. Fresh background NetEdit Inspect/TLS/Connection captures used the English keyboard layout and preserved candidate SHA df3c6de1936c9ff1bd16d363c1d2e3cc19af44302cff2b074575f7a0153e7548; no new cross-block connection or lane-surface distortion is visible relative to the accepted parent.	Freeze iteration 024 as the reviewed three-intersection road-topology candidate. This closes movement topology, not signal control: 2403 MAP/OCIT/TLD controller ownership and timing remain a separate blocked stage.
2026-07-21	Detector binding on frozen topology	Reused the existing MAP-to-SUMO and count-stream binders. Collapsed only one-to-one serial cuts of the same signed OSM way, preserving true parallel-lane ambiguity; for 2394, supplied the frozen official MAP file instead of using nearest-lane geometry.	Binding improves from 5/19 to 17/19 active. All 9 streams at 2394 now use official MAP ingress-lane identity. Only 2403 Z.1/Z.2 remain blocked because its MAP is unpublished and their nearest parallel lanes differ by less than the 1 m ambiguity gate. E1/E2 output remains withheld.	A detector at an OSM edge cut is one physical lane hypothesis, but a different OSM way or parallel lane is never collapsed. Official MAP identity overrides geometry only when the detector point is still within the strict distance gate.
2026-07-21	Shared-lane detector semantics gate	Re-ran the detector binder on iteration 024 with the 2394 MAP and the 2403 four-lane constellation evidence. The official MAP and LSBG plan retain the existing 2394 motor-lane counts, so duplicate Z fields do not authorize another road lane.	All 19 fields have lane hypotheses, but only 17 (node, lane) identities exist: 2394 Z.1/Z.2 share MAP lane 2 and Z.5/Z.6 share MAP lane 7 while their counts differ. The v7 manifest now keeps the mappings active but blocks E1/E2 materialization and automatic promotion instead of silently summing them.	Lane identity and aggregation semantics are separate evidence layers. Multiple official fields on one lane may be serial, redundant, or otherwise non-additive; Torii requires explicit official aggregation evidence before constructing one virtual sensor.
2026-07-21	2403 two-conflict-core closure and W1 handoff	Rejected the single-core/hull probes and the first two-core variant because they distorted surfaces or added two SUMO-guessed movements absent from the accepted parent. Iteration 027 keeps separate north/south conflict cores and deletes only those two guessed movements.	Preservation parity is 23/23 boundary movements with no loss/addition; surface, SUMO load and Connection Mode gates pass; 23/23 smoke vehicles arrive with zero teleport/collision; both owners have immutable background NetEdit reviews. The SHA-bound W1 handoff is review_ready with execution gate pass.	Join authorization is per physical conflict core. W1 promotion consumes one candidate SHA plus preservation, geometry, connection, runtime and NetEdit evidence; it still cannot prove official signal control.
2026-07-21	Final-OSM signal binding	Reused the existing iteration-009 compound TLS derivation instead of re-running path inference or comparing final OSM edges with hh-map-* teacher IDs. The W2 binder now accepts the hash-bound compound plan and checks every selected physical connection against iteration 027.	2349 binds 8/8 streams to 9/9 controlled OSM connections; 2394 binds 8/8 to 13/13. The two-hour official weekday history has 16/16 initialized streams and 3,671 exported events. W2 remains non-promoting because 2403 has no published MAP/OCIT/TLD control asset.	Preserve official movement-to-OSM evidence as a reusable contract. One movement may own multiple physical SUMO connections; exact set parity replaces edge-name equality.
2026-07-21	W3 official count-station truth correction	Audited Hamburg's Anzahl_Kfz_Zaehlstelle_15-Min SensorThings layer before materializing detectors. Kept the 5-minute Zählfeld streams only as lane/location diagnostics and switched the intended demand/validation truth to the processed directional station streams.	The selected Saturday window contains 10/10 bins for all nine station streams. Six direction-1/2 streams represent real directional cross-sections; three direction-0 streams are totals used only for QA, never additional route constraints. Four directional cross-sections can be uniquely materialized on W1; the two 2403 arm-2 sections remain review_required because their geometry is unique but their movement semantics cannot be promoted without the missing 2403 MAP/OCIT evidence. Node 2349 has no published station stream.	A detector field is not a traffic-total contract. Bind one official station to one directed SUMO edge cross-section, place E1 on all passenger lanes, compare the lane sum with the single station observation, and keep direction-0 totals and raw fields outside routeSampler constraints.
2026-07-21	W3 composition-first sensor identity correction	Replaced station-midpoint snapping with a hash-closed identity join: processed station stream → official zusammensetzung → physical 5-minute Zählfeld stream/Thing → official MAP, engineering-plan constellation, or unique official point evidence → frozen OSM-derived SUMO edge/lane. Added an explicit Hamburg↔SUMO sensor-type crosswalk.	Six directional station groups resolve to 15 unique physical fields and 16 memberships; 2403-Z.1 is intentionally reused by two processed groups. All 60 selected 15-minute bins exactly equal the sum of their 5-minute component fields. Fifteen E1 count equivalents and separate E2 queue/occupancy diagnostics load in SUMO. Only streams 28579, 29389, and 30599 directly cover every passenger lane of one co-located edge; 27143, 32721, and 36436 remain aggregation-only validation groups. Five field identities remain promotion-review items because four 2403 constellation source hashes are not closed by the upstream manifest and one 2394 detector point is 4.23 m from its official MAP lane. Node 2349 is still missing.	A processed Zählstelle midpoint is not a physical detector. Create one E1 per unique physical Zählfeld, aggregate E1 outputs by the official composition, use E2 only for queue/occupancy diagnostics, keep TLD request detectors separate from vehicle counts, and admit an edge constraint only when one co-located member exists on every passenger lane. The old midpoint-based W3 probe is superseded.
2026-07-21	W3 composition-derived route support	Reused Torii's existing network-boundary route enumerator, interval edgeData writer, constraint-structure audit, and SUMO routeSampler; supplied only the three direct full-lane station groups from the new identity manifest.	The corrected input contains 30 rows (3 edges × 10 bins) and 3,273 vehicles. Fifty-eight candidate routes give 100% constraint match, zero deficit/overflow, and zero structural conflict. The prior 40-row/4,312-vehicle midpoint probe is invalidated.	Zero deficit proves only that a plausible realization exists for the direct measured sections. It does not identify a unique OD matrix; the three multi-edge processed groups remain E1 aggregation comparisons only.
2026-07-21	W3 route inverse-problem formulation	Expressed route generation as a measured-edge-by-candidate-route incidence system before choosing any demand. Added rank, nullity, route-signature equivalence and unobserved-route diagnostics, and defined a multi-solution routeSampler ensemble followed by SUMO E1 scoring.	For the frozen Hamburg W3 inputs, the static matrix is 3 × 58, has rank 3 and nullity 55, and contains only eight distinct detector signatures; 18 candidate routes are invisible to the three direct hard constraints. Counts therefore identify aggregates of route-equivalence classes, not individual trips or a unique OD matrix.	Keep all plausible branch/through routes, generate one or more non-negative integer feasible realizations, and select only after simulation. routeSampler input deficit is a construction gate; co-located physical E1 output aggregated by official station composition is the acceptance evidence.
2026-07-21	W3 feasible-route ensemble	Reused the unchanged 58-route whitelist and SUMO routeSampler with seeds 17, 42, 73, 101, 202; no optimization or fabricated boundary flow was added.	All five versioned runs have zero deficit and five distinct interval-route fingerprints. They contain 2,302–2,339 vehicles and activate the same 40 physical route sequences, but pairwise interval-route L1 allocation distance is 838–971. The earlier --optimize full baseline contains 2,337 vehicles on only three routes and converges to one seed-independent sparse solution.	Retain multiple exact construction solutions. Sparse/full optimization is not automatically more realistic; choose among solutions only with SUMO E1 composition residuals, runtime quality and an explicit route prior.
2026-07-21	W3 route-ensemble outer-loop diagnostic	Ran all five zero-deficit routeSampler solutions on the exact frozen W1 network with per-run absolute E1 outputs. Used native --tls.all-off only to isolate topology, route allocation and detector timing; no placeholder TLS was treated as historical control.	Every run inserts and completes all 2,302–2,339 vehicles with zero teleport/collision. Formal two-hour six-station MAE ranges from 15.50 to 17.96 vehicles/15 min; seed 73 is the diagnostic leader. The residual is concentrated in the three aggregation-only station groups that were absent from routeSampler's edge constraints.	A zero-deficit edge solution is not a six-station solution. Keep selection=null: no-signal ranking is diagnostic and cannot replace same-window official TLS replay.
2026-07-21	W3 empirical sensor-response matrix	Probed every one of the 58 candidate routes with one isolated SUMO vehicle and all 15 co-located physical E1s, then aggregated the fields using the frozen six-station membership contract. Solved the resulting non-negative integer system with minimum L1 change from routeSampler seed 73.	The empirical station matrix is 6 × 58, rank 6, nullity 52, with 17 distinct route signatures. Unlike the coarse edge matrix, all ten 15-minute systems have exact integer solutions. The prior-preserving solution contains 2,677 vehicles and matches all 60 station bins exactly (5,928 = 5,928, MAE/RMSE/max error 0, GEH<5 100%). Four of five within-bin schedule permutations also finish with zero teleport/collision; schedule seed 74 is the clean diagnostic representative.	Sensor response must retain physical lane identity; a shared field cannot be expanded to a whole-edge count. The empirical matrix is an estimator, integer prior-distance minimization is the controller, and SUMO E1 is feedback. Promotion remains blocked because probes and replay disabled TLS and complete same-window official control for 2403 is unavailable.
2026-07-21	W3 two-node official-signal proxy	Replayed the exact 2,677-vehicle schedule with the complete 2349/2394 weekday official event stream shifted onto the count window; kept 2403's existing OSM controller explicitly as a proxy.	All vehicles arrive with zero teleport/collision. The six-station audit has 59/60 exact bins, total 5,927/5,928, MAE 0.0167, max error 1. The run is a cross-date signal sensitivity experiment, not a historical twin; manifest SHA is e6e2c8d455471b9664675f36a1e2bdf884ee86875cb7319022d263cdd999dcad.	Signal timing can change physical-lane choice even when the static route-incidence solution is exact. Keep the route allocation fixed and update only the smallest evidenced control variable; do not promote while 2403 and same-window timing are missing.
2026-07-21	W3 legal-speed dimension	Used TraCI detector vehicle data to identify the single missing contribution: s74_00929 touched 2394 Z8, but its lateral detector crossing produced an impossible apparent 57.4 m/s traversal and was excluded from nVehContrib. A first one-vehicle speed edit was invalidated because implicit SUMO speed draws shifted for later vehicles. Re-ran with every vehicle assigned an explicit factor at or below 1.00, then varied only the target over 0.85/0.90/0.95/1.00.	Every controlled case completes 2,677/2,677 with zero teleport/collision and all 60/60 station bins exact (5,928/5,928, MAE 0). The result proves reproducibility and a legal speed axis, but not a measured speed distribution; manifest SHA is 2eba7bd8126267cdbee8e4f8637fd4e17637e0fb3d05ec81543d204e705e89d1.	Flatten the speed tensor into a sparse matrix whose columns are route, departure pattern, and a discrete factor in (0,1]. Require explicit factors for the full fleet so random-number consumption cannot masquerade as a speed effect. Speed remains a regularized latent variable until official spot-speed/travel-time evidence exists.
2026-07-21	Superseded v20 NetEdit comparison	Captured Inspect, TLS and Connection views for the separate official-geometry v20 candidate (559a5752...). The user-provided screenshot and accepted W1/W2 contracts instead reference iteration 027 (aa1676df...), so v20 is not evidence about the active signal binding.	The mismatch was detected before modifying either network. The correct iteration-027 network was reopened through Torii's existing launch_netedit helper and its exact window title was confirmed.	Always resolve the reviewed network from the current hash-bound W1/W2 manifest before launching NetEdit; never infer the active candidate from the newest artifact directory.

Route inference as a sensor-matrix inverse problem

For one time bin, Torii defines a route-incidence matrix

[ A_{sr}=\text{number of times candidate route }r\text{ crosses measured section }s, \qquad A x_t \approx y_t, ]

where x[r,t] is a non-negative integer route multiplicity and y[s,t] is the official count. Most entries are binary, but the definition permits a route to cross a section more than once. Equal columns are observationally indistinguishable. A non-zero nullity means that the measurements admit many route-flow solutions, so Torii must report that ambiguity rather than assigning the label "observed OD" to one arbitrary result.

The reusable workflow is:

1. Build a whitelist of legal through, one-intersection, two-intersection and branch-exit routes from the frozen SUMO topology. Keep source/sink and route provenance; do not create traffic on an unmeasured boundary merely to make the picture look complete.
2. Audit A: rank, nullity, identical route columns, unobserved columns and necessary count-consistency conditions. This is the estimator's static identifiability report.
3. Generate an ensemble of non-negative integer feasible route multiplicities. Different random seeds or explicit priors may choose different members of the feasible set. A maximum-entropy/prior-distance objective is a selection rule, not new evidence.
4. Run every retained solution in SUMO. Replace static A with the simulator's time-expanded response: departure-bin vehicles may reach downstream sensors in later bins because of signals, queues and travel time.
5. Compare every physical E1 with its official Zählfeld, then sum E1 members using the official Zählstelle composition. Missing members make the whole station/bin missing. Choose or update routes from those residuals; never validate by summing unrelated serial detectors.
6. Promote a route solution only when the declared count tolerance, interval coverage, signal-history coverage, vehicle completion, teleport and collision gates all pass. Multiple statistically equivalent solutions may be retained as an uncertainty ensemble.

For compound or lane-specific stations, the production matrix is an empirical response matrix H, not the coarse edge incidence A. Torii sends one isolated vehicle over each candidate route, records every physical E1 response, and then applies the frozen official field-to-station membership. This preserves cases where one processed station contains fields on different SUMO edges or shares a physical field with another processed group. The controller solves H x_t = y_t with non-negative integer x_t, minimizing L1 distance to a plausible routeSampler prior. SUMO then replays the full mixed demand; a changed lane choice, bin shift, queue or collision returns a residual to the next controller iteration.

Speed adds a finite latent dimension rather than a continuous free variable. For speed classes k with factors such as 0.85, 0.90, 0.95, 1.00, Torii uses

[ H_{(s,t),(r,b,k)}=\text{SUMO E1 contributions at station/bin }(s,t) \text{ from route }r\text{ departing under pattern }b\text{ at speed class }k. ]

The tensor is flattened so rows remain (station, observation bin) and columns are (route, departure pattern, speed class). Decision variables are non-negative integer vehicle counts. Every factor is at most 1.00, so the desired free-flow speed never exceeds the lane limit; car-following, queues, turning speed and red-light compliance may reduce the realized speed further. Because interacting vehicles make the full response nonlinear, isolated probes are only an estimator. The controller must run the complete mixed solution in SUMO, measure the residual, and refresh local columns or a finite-difference Jacobian when the additive approximation fails.

This follows SUMO's own warning that traffic counts do not define a unique solution and its recommendation to use a plausible route whitelist with routeSampler (Routes from Observation Points, routeSampler). The simulation outer loop follows dynamic OD-estimation work that treats SUMO as the inner assignment model and minimizes sensor residuals outside it (Andersen, Chiarandini and Debrabant, 2022). Flow conservation, positivity and prior trajectory structure are kept as separate constraints rather than being confused with measurements (Michau et al., 2016).

Current topology state

• The only product scope is Am Sandtorkai 2349→2394→2403; Ingolstadt and the alternative Hamburg corridors are method fixtures, not replacement targets.
• The active baseline is the immutable OSM bbox source. OSM owns continuous road shapes and remote approaches; official Hamburg data may authorize bounded conflict-core movements, lane counts, controller ownership, and signal semantics.
• 2349 and 2394 now have an OSM-derived official static-topology candidate: 21/21 official lanes and 16/16 vehicle movements are bound, all 133 external lane geometries are unchanged, SUMO loads the network, and no local surface finding was introduced.
• The 2349/2394 sub-scope is now machine- and NetEdit-reviewed and frozen by hash. It is two-of-three topology accepted, not a corridor digital twin; its all-red programs are structural placeholders and do not claim historical signal replay.
• 2403 is no longer an unresolved physical-shape problem. Imagery, the OSM road skeleton, and the Ingolstadt preserve-split teacher rule support a compound four-arm layout rather than one merged junction. A scoped controller changed only six authorized junction polygons; all 133 external lane geometries, every connection, and every TLS payload remain unchanged. The 11 lane-surface and 5 junction-surface findings are now zero, SUMO loads the candidate, and the generic Connection Mode regression has no blocker.
• The July-2022 and March-2024 official engineering plans close the remaining Brooktorkai lane-permission gap. Iteration 027 then replaces the unresolved micro-node cluster with two bounded physical conflict cores while preserving all 23 source-authorized boundary movements. The W1 topology is frozen at SHA aa1676df2182026a87d261e633d2cf8bd100c9b2a8c3ee6ac9d9d3dd22d49a33; 2403 controller ownership and historical timing remain unknown because no current machine-bindable MAP/OCIT/TLD package is published.
• W3 now keeps two different official entity types separate. Fifteen unique physical Zählfeld fields become co-located SUMO E1 count equivalents (with optional E2 diagnostics); six processed directional Zählstelle streams are aggregation groups over those fields, not additional detector points. The 60-bin composition audit has zero differences. Three groups are direct single-edge constraints and three are validation-only. Node 2349 has no published station stream, so corridor-wide promotion remains blocked.

2403 topology status split

Layer	Status	Evidence boundary
Physical compound layout	pass	OSM continuity, current imagery and Ingolstadt preserve-split teacher agree on multiple physical owners.
Junction/lane surface geometry	pass	Authorized six-polygon repair resolves 16 -> 0 findings with 0 m external-lane deviation.
Generic SUMO connection graph	pass	No connection delta, no Connection Mode blocker, and SUMO load passes.
Official legal-movement parity	pass	Two hash-bound LSBG engineering plans prove the four Kfz lanes and the one missing lane-0 straight movement; iteration 024 adds exactly that movement.
Signal controller ownership	blocked	Official signal groups and controlled movements are unavailable; existing TLS data cannot be promoted.

2026-07-21 evidence publication

• Published a compact GitHub evidence bundle for W1/W2 containing the frozen SUMO network, its raw OSM bbox source, an attributed OSM rendering, the official Hamburg geobasemap, two hash-bound NetEdit review captures, W1/W2 manifests, signal events, and acquisition provenance.
• Kept generated caches and experimental candidates outside version control; the bundle does not change the W1 topology or relax the W2/W3 gates.

2026-07-21 stop-line and repeated-control audit

• Reopened the hash-bound iteration-027 network rather than the superseded v20 candidate and traced all 16 official 2349/2394 vehicle movements through the final SUMO lane graph.
• Seven movements are currently controlled twice by the same official HH_* / linkIndex. Four duplicates were introduced because a MAP-proven connectivity repair was also treated as signal-control evidence; three came from consecutive OSM micro-TLS links being absorbed into one official group.
• Projection of the official MAP ingress stop lines onto iteration-027 separates the real and duplicate arcs. At 2394 the first controls are 0.58--1.93 m from the official stop lines, while the downstream duplicates are 8.13--9.94 m away. At 2349/C9 the corresponding distances are 4.00 m and 22.94 m.
• The repair policy is therefore: MAP connection repairs provide route topology only unless they carry explicit stop-line authority; one official movement keeps the first stop-line-aligned control arc and demotes later controls.
• 2349/C5 and C6 remain a topology exception: their only controlled branch is 14.03 m downstream of the official stop line. The intervening external edge must be represented as internal channelized connection geometry so both movement-specific links begin at the real stop line.
• Three additional 2349 OSM micro-TLS controllers precede the official owners by only 4.11--4.74 m on the same straight OSM ways. They are duplicate approach controls and must be retired in the corrected candidate. Equivalent 2403 pairs remain review-only because no official MAP/OCIT package identifies the survivor.

2026-07-21 official stop-line ownership correction

• Projected every official MAP KML ingress endpoint B into the exact iteration-027 SUMO coordinate frame instead of inferring the stop-line owner from the current HH_* controller. This corrected the provisional retirement interpretation above: the three 2349 OSM micro nodes are nearer the official stop lines and must be taken over by HH_2349, not deleted as duplicates.
• The same check found that 2394 MAP lanes 6/7 place their stop line near the start of 381540198#1; the physical control therefore belongs on the unique predecessor connections from 381540198#0, while the later conflict-core connections remain routing-only.
• Torii will treat a MAP ingress binding as a stop-line cross-section rather than automatically treating the mapped lane's downstream end as the signal owner. If the official anchor is nearer the lane start, the estimator may shift to exactly one proven predecessor; ambiguous predecessors fail closed.
• One local geometry exception remains. Official lane 10 at 2349 serves C5 and C6 with different signal expressions after a shared channelized path. The corrected PlainXML candidate must absorb only 61649647#1/#2 into the local logical cell and create direct movement-specific internal connections from 61649647#0, preserving the concatenated OSM geometry. No movement is collapsed to a central straight-line shortcut.

Next steps

1. Treat iteration 027 SHA aa1676df…49a33 as the immutable W1 topology; later signal or detector work must prove exact road-topology preservation.
2. Preserve the completed 2349/2394 static and two-hour historical signal bindings; keep 2403 signal ownership/timing explicitly blocked until a machine-bindable official asset is available.
3. Use the frozen composition-first identity index: one E1 per unique physical field, E2 only as a queue/occupancy diagnostic, and official station totals as group comparisons. Generate routeSampler edge constraints only for the three direct, full-lane, single-edge groups; keep multi-edge groups out.
4. Preserve the exact empirical-matrix allocation and clean schedule-74 route evidence. Use explicit legal speed classes for every vehicle, not SUMO's implicit random speed distribution, and extend the response columns by departure pattern and speed class. Re-run the complete mixed demand after every estimator update.
5. Productize only the reusable estimator/controller/audit steps; Ingolstadt priors remain non-authoritative and no validation corridor replaces Hamburg.

2026-07-21 scope reset: systematic Ingolstadt human-cleaning curriculum

• Suspended every previous Hamburg W1 promotion claim. The latest compact-core join experiment is retained only as a negative sample: it preserved external lane geometry after repair, but introduced unsupported connection fanouts and changed controller ownership without official 2403 MAP/OCIT evidence.
• Audited the existing same-bbox Ingolstadt runner before adding code. Torii already computes topology clusters, road hierarchy, scope, Connection Mode, movement/TLS deltas and review artifacts; the missing layer was a curriculum that connects those observations to typed human-cleaning actions.
• Materialized outputs/ingolstadt_teacher_curriculum_v1: 149 existing raw-versus-human cases become hash-bound state/action records. The current v1 evidence contains 29 exact positive bounded-core joins, 61 review counterexamples, and 59 blocked/unidentified examples. Leakage grouping keeps cases sharing a nearby spatial cell, absorbed source node, internal OSM way family, or human controller in one fold; the resulting split is 123 train and 26 held-out cases. Eighteen diverse cases form the first NetEdit A/B queue.
• The curriculum reuses the existing topology estimator and composable intersection/road layers. .net.xml remains the exact NetEdit/SUMO payload; the higher-level semantic state should reuse CanonicalNetworkSnapshot and CorridorResearchBundle, not introduce another competing network language.
• Found a safety defect in the legacy aggregation path: a reference_matched Ingolstadt case can currently be treated as a confirmed target join even though the teacher action contract explicitly blocks transfer. The controller must convert teacher results to prior evidence and a reversible hypothesis, then require independent target-city geometry, stop-line, storage/channelization, movement, boundary and official-data gates.

Next curriculum increment:

1. Freeze representative NetEdit A/B views for T3, X4, offset/dog-leg, channelized, shared-controller and abstain cases using identical view rules.
2. Add typed v2 teacher actions for movement gaps, pedestrian internals, controller split/merge, road hierarchy/type and geometry changes; v1 join labels remain backward compatible.
3. Evaluate JOIN / KEEP_SEPARATE / ABSTAIN and exact absorbed/internal/boundary sets on held-out Ingolstadt families before enabling any Hamburg materializer.
4. Transfer only a rule class that passes held-out tests, then let Hamburg OSM, official road data, MAP/OCIT/TLD, imagery and NetEdit review accept or reject each proposed edit independently.

2026-07-21 Ingolstadt spatial registration and OSM micro-loop lesson

• Corrected the Ingolstadt A/B screenshot protocol. The raw same-bbox SUMO net and the human-cleaned teacher use the same UTM projection but different netOffset values; separately centering their local coordinates did not prove that the screenshots showed the same real intersection. A new read-only registration artifact converts both local frames to common projected UTM and binds each screenshot center to the exact source-network hash.
• Rebuilt the curriculum with that registration: 141 of 149 cases register, eight remain blocked, and 29 cases retain exact blocker-free core-join evidence. Old screenshot pairs whose projected centers differed were invalidated instead of being accepted visually.
• Confirmed the user-observed small-loop pattern at Ingolstadt source nodes 1448519097 and 7254701196. The imported net contains reciprocal external edges 24498193#1 and -24498193#1, each only 0.20 m long, plus a generated dir="t" turnaround at each endpoint. The human-cleaned network contracts the two source nodes into cluster_1448519097_7254701196; therefore this case is a positive example of a false OSM/SUMO micro-fragment rather than a real roundabout.
• The existing generic corridor simplifier does not discover this case because it requires a non-empty common road name and only accepts simple lane-preserving geometry nodes. This pair is unnamed and each endpoint also participates in a junction, so it requires a separate, fail-closed reciprocal-micro-fragment estimator rather than a relaxed name check.
• Hamburg contains the same structural family: the 2394 workflow already records the 0.2 m edge 60578487#0 as consumed by a bounded join. Ingolstadt supplies a reusable prior, but Hamburg contraction still requires its own stop-line, channelization, movement, official-data and boundary-preservation evidence.
• A read-only run on the Hamburg source also classifies the 2403 pair 249707791 <-> 757036954 (-60578489 / 60578489, two lanes in each direction, 0.20 m) as a geometry-fragment candidate. The earlier diagnostic join absorbed it, but that candidate failed later surface-overlap gates, so the classification is not a promotion claim.
• The strongest small-ring signature still visible in the current Hamburg diagnostic candidate is different and must remain protected: at 2349, 194672083 and 59990286 form a 0.82/0.81 m reciprocal pair and netconvert generated one direct dir="t" movement at each end. The pair is controlled by HH_2349, has asymmetric/unproven way semantics, and is used by the official movement/stop-line binding. The estimator therefore returns protected_or_review, not a join. The safe first repair is to compare the two turnarounds with MAP/OCIT and suppress only an officially unsupported turnaround; the short connector remains until independent evidence proves a single physical core.
• This Hamburg ring was produced by a partial node join plus automatic turnaround generation, not by copying a complete OSM roundabout. This is why micro-fragment contraction and turnaround suppression are separate actions.
• Added the protocol actions ABSORB_RECIPROCAL_MICRO_FRAGMENT and SUPPRESS_ARTIFICIAL_TURNAROUNDS. They are review-only until a candidate round trip proves exact non-turnaround movement parity, controller/stop-line preservation, zero outside-scope edits and no new geometry overlap.
• Implemented the independent read-only estimator in external_micro_junction_audit.py. Focused verification across the new audit, owner-layout classifier, spatial registration, curriculum, visual evidence, typed actions and aggregation safety completed with 53 passing tests; Ruff and git diff --check passed. No source network was modified.

2026-07-22 iteration 036: preserve iteration 027 and suppress its artificial small ring

• Returned to the user-selected iteration-027 network instead of rebuilding its road geometry. The official 2349 MAP/OCIT inventory contains eight vehicle movements and none uses the direct 194672083 -> 59990286 or reverse turnaround generated by netconvert.
• Generated an independent candidate that deletes only those two dir="t" connections. All external edge XML, non-turnaround connections, TLS IDs and TLS programs remain unchanged; the two short channelized edges remain in the network because their official movement and stop-line roles are still valid.
• SUMO load passed. The 23-vehicle boundary movement smoke completed with 23/23 arrivals and zero teleports. Torii's bounded surface comparison reports zero introduced findings and zero findings at the edited 2349 owners. Candidate SHA-256: 8bab5e9996cfc48c5d9248a687b44b4adfdc245e12f38f5c06583ee70eff7e71.

2026-07-22 iteration 037: internalize only the 2403 sub-metre road fragments

• Starting from iteration 036, joined only the three near-coincident owners that bound the reciprocal 0.29 m and 0.58 m OSM fragments. This absorbs exactly four external micro edges; the 7.85–12.53 m approach/channelization edges remain external.
• The first native netconvert join inferred one new lane movement (83075275#2_0 -> 111747607#1_1). A second semantic-prune pass removed only that unsupported addition. The final boundary inventory is exactly 18 -> 18 with no lost or added movement.
• SUMO load, Torii collapse/presence gates and bounded surface comparison pass. An all-off topology smoke sends 18 vehicles through all 18 movements with 18/18 arrivals, zero teleport and zero collision. Candidate SHA-256: f7e9fb4c25dc16873a7c3110870b7674f5fc643277700c4882933876945ce3d0. The candidate remains diagnostic because 2403 still lacks a published MAP/OCIT package for official TLS promotion.

2026-07-22 iteration 038: Ingolstadt-style single-core probe for 2403

• Reused the existing Ingolstadt teacher curriculum. Human-cleaned case04 collapses 14 source OSM nodes into one physical cluster_* junction while retaining the external approach cuts.
• Applied the same bounded hypothesis to 2403 without replacing iteration 037: merged the existing central core with four stop-line owners and their one intermediate node; five short core edges became internal junction geometry.
• SUMO initially dropped one parent movement. Restored only that exact lane-to-lane connection (25444569#0_0 -> 395272650#1_0). The final probe preserves all 18 boundary movements, introduces no surface overlap, loads in SUMO, and all 18 topology-smoke vehicles arrive with zero teleport/collision.
• The candidate SHA-256 is 2da03214eea32571669d5a0cbac80fd88a4bd9079a76be36290a8c180df5c559. It remains diagnostic because 2403 has no published MAP/OCIT controller package. Background NetEdit A/B capture is pending: the capture guard observed the active foreground keyboard layout as Chinese (0x0804) and correctly refused to send mode shortcuts.

2026-07-22 NetEdit screenshot default

• Made Torii's hash-bound background NetEdit reviewer the mandatory default for screenshots; it does not use global keyboard or mouse input.
• The legacy foreground/full-screen capture command now fails closed and points to netedit_background_review.py.
• Interactive NetEdit viewing remains a separate CLI-only action and is never used as the screenshot path.
• Removed the background reviewer's dependency on the user's foreground input method: shortcut safety now checks the private NetEdit window's keyboard layout, so a Chinese ChatGPT input context no longer blocks silent capture.
• If that private NetEdit window inherits Chinese input, Torii requests English for that window only with WM_INPUTLANGCHANGEREQUEST; it never changes or types into the user's foreground application.

2026-07-22 junction Inspect correction

• Continued iteration 038 after restoring the one dropped parent movement. The candidate still passes 18/18 movement parity, SUMO loading and the 18-vehicle topology smoke; this proves connection semantics, not junction-shape quality.
• Verified that the Ingolstadt teacher network was used read-only. Its SHA-256 remains cc629a9aaa7158cebafaa044971f82677506616476a487967a53543353efac5d and Git reports no modification to the teacher file.
• Rejected a first screenshot experiment that drew the junction footprint as an additional polygon. A polygon is not a SUMO junction and cannot be used as Inspect evidence. Those images are diagnostic only.
• SUMO's junction is the network node object; its shape determines where incoming edges end and outgoing edges begin. The current background runner can hash-bind and select junction:<id>, but it cannot yet prove that NetEdit's left attribute pane entered Net: junction without taking foreground input. The report now states this limitation instead of claiming successful object inspection.
• Added a stable canvas-window lookup so the reviewer does not attach to a transient NetEdit top-level window. Background screenshots remain non-mutating and promotion remains blocked until a real junction Inspect view is verified.
• Reused the existing NeteditTargetSession rather than extending the background reviewer with synthetic FOX messages. Its hash-bound real click produced a Hamburg screenshot whose left pane is Net: junction; Ctrl+J then switched from the node bubble to the compiled junction shape. The session was aborted without saving and the source candidate hash remained unchanged. This is now the required fallback only when attribute-pane proof is requested; ordinary screenshots remain background-only.

2026-07-22 repository report and reusable workflow

• Added a curated README report containing the official 2024 LGV aerial, the July-2022 LSBG construction plan, the OSM-derived input and a background NetEdit full-network Connection Mode capture of iteration 038. The local Net: junction capture remains a companion identity proof rather than the main topology image. Both reference candidate SHA-256 2da03214eea32571669d5a0cbac80fd88a4bd9079a76be36290a8c180df5c559.
• Split public showcase material from repository evidence. Curated images, provenance, schemas, source/tests and a compact hash summary are versioned; repeated API caches, NetEdit sessions and generated simulation directories remain rebuild-only artifacts.
• Added a portable config facade to the existing Hamburg W0-W5 execution ledger. Stage paths are relative to the config, upstream manifest/feedback changes still invalidate downstream stages, and the facade does not duplicate the focused road, signal, detector or replay implementations.
• The report keeps the product boundary explicit: exact route-matrix and zero-teleport diagnostics are shown, while missing 2403 MAP/OCIT and low-confidence detector-lane identities continue to block promotion.

2026-07-23 workflow DAG v2

• Split legacy W3 into real existing producers: W3a count acquisition depends on W0, while W3b detector binding depends on W1 and W3a. Route incidence, demand, and replay remain together in W4 until independent producers exist.
• Re-hash W1/W2/W3b/W4 network files and reject any downstream SHA-256 that differs from W1. Stage schemas are now explicit instead of accepting any JSON object with a passing status.
• W4 now consumes the hash-bound W3a count values and W3b detector mapping, verifies its selected W2 binding, rejects unresolved shared-lane aggregation, and rejects mutated or cross-binding TLS event history. Its replay manifest and the hash-bound signal-observation manifest are versioned as v2 because these are breaking input/provenance contracts.
• W5 is generated as the terminal capability-summary record. Resume now invalidates only already-materialized strict descendants, keeps that invalidation until descendant bytes change, never invalidates the newly supplied changed stage, and checks promotion gates independently of whether execution has another action.