8.1 KiB
Measured E/F label trials: not promoted
Source d0896ced75fea73bcda6bd14f259a34503840704 is an isolated, externally
selected experiment. Neither profile is accepted as a production rule. F's first
reader gave wrong Request/Report endpoints and retained an alternative allocation
ownership interpretation for the conversation. E's reader identified the intended
participants, undirected conversation, Request and Report, but with moderate
confidence and explicit route ambiguity. These separate single-agent static
readings do not establish a causal comparison or human usability.
Frozen implementation and gates
The profile changes only the selected conversation's derived label presentation. Canonical text, full Project and all routed points remain untouched. Actual Iced Segoe UI size-13 measurement yields E backing 119.174805 x 40 and F 89.18213 x 56; line advance is 16. Canvas/SVG/picking share layout records, and pre-experiment core geometry is retained separately for gesture baselines. See the unchanged initial review and repair review.
The first compile failed with E0063 because a test EdgeRoute omitted attachment fields. After that correction, the added different-edge preview regression failed: refusing wrapping retained stale wrapped hit bounds while painting one line. The repair stores raw core routing before presentation mutation. Both failures remain in the raw gate logs.
At frozen d089 the parent completed 126 selected tests, 0 failed, 0 ignored, 1,426 filtered out, including all six experiment tests; strict desktop all-target Clippy and the GUI development build passed. No full-suite result is claimed. The fresh formatting check exited 0 with empty output; it was run after an earlier summary had named a log that did not exist. commands.json records parent-observed command flags, environment and outcomes; build stdout alone does not establish invocation scope.
Executable SHA256 is 2b6d76d744d7422fc826f7f7ca0009b296a0efb751fe0b3bbc1516b8a624c201;
runtime is f45604a12e8bcd839fd86f667c1be1cc7c4f9da2b3c69ff381ac708f0e1ae7d4.
Windows GNU LLVM dev/unoptimized only. Both original E/F runs exited normally
with code 0, empty streams and unchanged pinned files, as captured by their
launch watchers. Binaries/runtime copies are excluded from this evidence packet.
Native observations and independent byte checks
Read the locked E first reading and F first reading before interpreting captures. The parent reports that clicks in E's two lines selected the same BoundaryExchange. F's three lines were selectable; a declared 30-pixel-right/20-pixel-down drag, Undo, Redo, JSON export and final Undo to clean state were exercised. These are parent-observed native actions, not deductions from source tests or screenshots. The parent observation record distinguishes actual tool observations from captures and a raw event transcript (which was not retained). F's last-line click is tool-observed only. The authored move crossed a foreign stroke; respecting authored offset is not a reading pass. E did not exercise native editing or Undo/Redo.
Both actual native SVGs preserve all twelve edge UUID/path/orientation/drawing records. Removing only the target edge's immediate label presentation leaves the same 6,977 raw bytes as the actual baseline native export: all cards, ports, other labels, root attributes and viewBox remain exact. The target's complete canonical title and ordered Segoe UI size-13 line/baseline positions agree with predeclared measured metrics. The SVG checker does not verify rendered glyphs, font availability on another machine or browser raster parity.
All original full Project inputs retain SHA256
9db1015369f450d5810c96a01fddb8a009d57f6aafbda5182703f15ded23d9c8.
The complete F exported Project differs only by the target route's label:
fraction .55, offset x120.812744/y38.814423; no waypoints, attachments, semantic
fields or formatting profile were added. Exported JSON is not evidence of an
ordinary source Save or save-clean lifecycle. The Model JSON operation reported
export, and F remained dirty until Undo.
Two later fresh-process launches used the saved 0ad401... Project. Under F the
parent observed the same three-line shape at the authored position and exported
reopened-F.svg. The ordinary Baseline launch opened the same saved file in
the preceding F run's folder via Recent Open, not the unused identical copy in
its own folder. It displayed one line, confirming the external formatting
profile was not persisted. No ordinary-profile SVG was exported. Both launches
exited 0 with empty streams and unchanged pins; parent rechecked the actual
loaded file. The observation record
preserves the failed forward-slash chooser attempt, unavailable indexed OK action,
empty filename attempt and successful recovery instead of hiding them.
The independent final evidence review predates these reopen runs. Its then-missing lifecycle checks must not be read as review of newer evidence. The later saved-F review independently confirms the full model delta, all twelve unchanged paths and every SVG byte outside the moved label. Its f32 reconstruction from the original path, saved fraction/offset and prior font metrics predicts the reopened coordinates within .0002 units; five corruption controls were rejected. That review did not observe the GUI or inspect ordinary Baseline reopening. The latter remains a parent-observed native result, not an independent SVG check.
Packet and limits
The inventory covers every raw evidence file
except itself. Scoped -text attributes retain measured bytes through Git.
verify.py checks hashes, process receipts, SVG conservation
and the complete model delta without launching anything:
python harness/experiments/wrapped-label-20261001/evidence/verify.py
# After staging, also check staged bytes; optionally verify frozen source objects:
python harness/experiments/wrapped-label-20261001/evidence/verify.py --index --source
Historical preparation/launch helpers retain original local paths and are not
portable replay commands. The pristine SVG/model checkers use positional paths
and exclusive new output files. The independent reopened-SVG checker retains its
original fixed local paths; evidence/verify.py invokes its read-only check with
packet inputs, providing an offline replay without rewriting historical code.
The first SVG selfcheck setup failure is retained: it chose a
card fill absent from the actual export, then was corrected; native files were
unchanged. Ten in-memory corruptions were rejected by the SVG checker. The model
helper contains an ineffective constant comparison; its meaningful drag bounds
use the displayed rounded 52% zoom, not independently measured exact scale.
Its full parsed-Project equality check establishes the sole changed field set.
It relies on assertions and only writes a success receipt; python -O disables
its checks and failures leave no structured receipt. The inventory records its
hash separately. Do not treat it as a general lifecycle harness.
The earlier SVG report preserves its historical F-only checkpoint; the later E receipt extends that evidence. Historical notes saying gates were pending are retained as original records, not overwritten into retrospective claims.
The external profile is not persisted. Authored fraction/offset are honored while the profile is enabled; the ordinary product's wrapping/persistence contract is unresolved. No general typesetter, winner, performance acceptance, nested/folded coverage, accessibility result, Linux result, glyph/raster parity or complete OA/SA acceptance is established. Native held-pointer cancellation, native different-edge guard refusal and E editing remain unverified. Publication and any further lifecycle work remain parent-owned.