lanyue-llk · Sep 22, 2026 · Sep 17, 2026 · Sep 17, 2026 · Sep 17, 2026 · Sep 17, 2026
diff --git a/.agents/skills/benchmark-editorial/references/html-delivery.md b/.agents/skills/benchmark-editorial/references/html-delivery.md
@@ -2,6 +2,8 @@
 
 The dashboard reading page and standalone HTML are two presentations of the same authored article. Use the current report renderer and preserve its integrity, safety and localization contracts.
 
+Generate Chinese by default. English prose, illustrations, rendering and language-specific checks run only on an explicit manual request. An absent English edition gets a not-generated notice and a return link, not stale translated pages. Preserve translation sources without distributing unused English assets. Apply cross-language parity checks only when both editions are generated; always test the notice in a Chinese-only release.
+
 ## Reading layout
 
 Choose a clear hierarchy for title, introduction, sections, body and captions. Align headings and their supporting figures deliberately, keep body lines readable, and check the result at actual display sizes. Do not create an oversized heading from an editing instruction or shrink a necessary qualification into unreadable text.

diff --git a/.agents/skills/benchmark-report/SKILL.md b/.agents/skills/benchmark-report/SKILL.md
@@ -11,6 +11,8 @@ Read the repository's `AGENTS.md` first. Project rules override this general ski
 
 ## Working contract
 
+Repository reports default to Chinese-only output and checks. English generation (including translated illustrations) requires an explicit manual request. Bind generated languages in the report manifest; English navigation opens a not-generated notice when absent. Do not publish stale English pages. Run English-specific validation only for an explicitly generated English edition, retaining full parity checks when both editions exist.
+
 1. Establish the decision being supported, the measured unit, cohort, versions, execution location, freshness policy, and intended deliverable. Inspect the actual dataset, runner, outputs, scorer, prior review evidence, and current report. Do not infer them from screenshots or prose.
 2. Freeze a machine-readable evaluation contract before expensive calls. It must identify the exact cases and methods, run count, timeout, concurrency policy, input and output contract, scoring provenance, reference state, and publication target when one exists. For source-dependent tasks, fail closed: verify and freeze references immediately before calls, verify the same declared scope afterward, and certify the cohort only when both checks pass. Partial or uncertified cohorts remain diagnostic evidence and cannot enter a release.
 3. Let each method use its strongest documented route to produce the required output. Equalize the task, inputs, output contract, call budget, and scoring. Do not require identical internals. Record material asymmetries such as local versus remote execution, cold versus warm start, bundled conversion, network transit, account tier, and billable time.

diff --git a/.agents/skills/benchmark-report/references/report-and-release.md b/.agents/skills/benchmark-report/references/report-and-release.md
@@ -1,5 +1,11 @@
 # Report and release
 
+## Language generation
+
+All tracks generate and validate Chinese by default. English requires an explicit manual trigger; do not automatically translate prose, generate English illustrations, render English routes or run English-specific checks. A manifest records which languages were generated. When English is absent, retain only a clear not-generated notice with a return link, and test that control. Remove stale English output and unused translated assets from the new bundle while preserving source translations. Parity and English localization checks apply when the English edition is explicitly generated; Chinese-only releases retain all Chinese evidence, semantics, links and visual gates.
+
+For the current trajectory report, `make trajectory-replay-report EVIDENCE=/absolute/path/results.json` builds Chinese only. Explicit `ENGLISH=1` (or the build script's `--include-english`) enables English. Report, visual and packaging checks use the generated manifest, so they do not regenerate an absent edition.
+
 ## Reader-facing structure
 
 Organize around reader questions, not storage tables. A useful report usually makes these answers easy to find:

diff --git a/AGENTS.md b/AGENTS.md
@@ -1,7 +1,8 @@
 # Report maintenance
 
-- Keep report-specific illustrations and review payloads outside Git, alongside frozen evidence. Commit only their compact path/hash bindings. WCXB builds read the profile-bound files from `<evidence>/report-assets/` and preserve the same relative paths in the public package; restore missing files from the profile's pinned Artifact Site version, never from an unversioned latest report. Do not silently fall back to a repository copy.
-- Remove retired report-specific generators and their exclusive tests when no supported entry point calls them. Preserve published reports and historical evidence. Keep behavioral tests for the supported build, package, and publication boundaries; do not replace them with assertions about private helper structure.
+- Keep report-specific prose, illustrations and review payloads outside Git, alongside frozen evidence. Commit compact path/hash bindings instead. The replay build reads `research/trajectory-replay-content.json` from the `--content` directory (default: the data home's `report-assets/trajectory-replay/`); restore missing content from the pinned Artifact Site version, not an unversioned latest report. Do not silently fall back to a repository copy. Compact task definitions, acceptance contracts and provenance annotations remain versioned inputs.
+- Remove retired report-specific generators and their exclusive tests when no supported entry point calls them. Preserve published reports and historical evidence. Keep behavioral tests for supported build, package and publication boundaries; do not replace them with assertions about private helper structure.
+- WCXB builds read profile-bound files from `<evidence>/report-assets/` and preserve the same relative paths in the public package; restore missing files from the profile's pinned Artifact Site version, never from an unversioned latest report. Do not silently fall back to a repository copy.
 
 ## Common report presentation
 
@@ -28,7 +29,7 @@ These requirements apply to **all** report tracks; browser-specific rules below
 - A reference, scoring or acceptance correction must fix the shared cause, not patch stored verdicts or one provider. Add paired executable controls to `evaluation/scoring-regressions.json` where both error directions are plausible, run them through the public `score_attempt` and task predicate, and bind the root cause plus regression IDs in every new reference/scoring revision receipt. A revision without this provenance fails before evidence is changed.
 - A heterogeneous same-window latency cohort keeps every predeclared method-task call. Report the common-denominator all-call distribution separately from each method's conditional successful-call distribution; an all-method common-pass intersection is diagnostic only and must never drive post-outcome task selection or a ranking.
 - Evaluation credentials come only from the repository `.env`. Artifact Site publishing is the sole user-scoped exception: read explicit `ARTIFACT_SITE_TOKEN`, then `~/.config/artifact-site/tokens/<target-host>`; the previous `ARTIFACT_HUB_TOKEN` and `~/.config/artifact-hub/` locations are read-only migration fallbacks. Never load `.env` for publishing. Save new authorizations per target host without overwriting another deployment's token. Keep `.env.example` current and publish only account requirements, official acquisition links and variable names.
-- Chinese and English are two localized views of one report model. Their HTML route inventories, section order, tables, values, filters and evidence links must match one-to-one. Localize interface and explanatory prose for each language; keep task names, IDs, URLs, references, provider output and other measured evidence unchanged.
+- All report tracks default to Chinese-only generation and checks. English prose, illustrations, rendering and validation require an explicit manual request. When English is absent, its navigation control must open a not-generated notice with a return link; do not ship stale English pages or unused English-only assets. Preserve translation sources for manual use. Record generated languages in the report manifest. Run release/visual checks for those languages and the not-generated notice. When both editions are generated, their HTML route inventories, section order, tables, values, filters and evidence links must match one-to-one. Keep measured evidence unchanged.
 - Web, dynamic, popular-site and document text use shared `score_attempt` orchestration and one task-result predicate. Task-specific requirements live in case configuration, not provider- or track-specific scoring branches. Binary acceptance and text diagnostics remain separate; missing reference coverage is unavailable, not zero or perfect.
 - Dynamic-content reference capture follows `harness/web_fetch/README.md`: retain paired initial HTTP bytes and independently rendered visible text, bound completion and the full declared scope. Distinguish embedded state from follow-up fetching; never require JavaScript as the only valid route.
 - Popular-site stable text must be independently observed and task-relevant. Do not freeze volatile ranks, prices, member counts or current list membership. Title diagnostics join returned entities to independent names by detail URL and never substitute for body quality, recall, freshness or task acceptance.
@@ -57,5 +58,5 @@ These requirements apply to **all** report tracks; browser-specific rules below
 - Split environment preparation, browser-session setup, Agent run, BrowserTool wall time, evidence finalization and evaluator time at collection. Publish the self-hosted deployment topology and image identity; claim per-execution reset/readiness only when a preparation receipt binds reset, readiness and application health.
 - Formal report prose is reader-facing: no prompts, reasoning, internal work narration, conversational rebuttals or references to parallel sessions. Review pages contain review rounds and corrections, not per-case audit tables; release notes contain reader-visible effects.
 - English routes may preserve measured task names, IDs and evidence verbatim, but all report-authored interface and explanatory prose must be localized. Release checks must reject known untranslated interface markers and untranslated accessibility attributes.
-- Before release, inspect every core route class in Chinese and English at desktop and narrow widths in both themes. Prevent vertical table labels, clipped controls and unreadable evidence; use ImageGen raster assets or Mermaid rather than SVG for Browser Use diagrams.
+- Before release, inspect every core route class in each explicitly generated language at desktop and narrow widths in both themes, plus the English not-generated notice when applicable. Prevent vertical table labels, clipped controls and unreadable evidence; use ImageGen raster assets or Mermaid rather than SVG for Browser Use diagrams.
 - Keep routine report checks fast and deterministic; they cover structure, evidence, links, terminology, and static readability guards without launching a browser. Run the separate full release check before packaging or publishing; it adds the complete cross-language, cross-theme, desktop and narrow-width visual gate.
diff --git a/Makefile b/Makefile
@@ -39,7 +39,7 @@ ci: naming-check test fixture-smoke
 
 report: runtime-check
 	@test -n "$(EVIDENCE)" || (echo 'EVIDENCE=/absolute/path is required' >&2; exit 2)
-	$(PY) scripts/build_report.py --evidence "$(EVIDENCE)" --output reports/webfetch
+	$(PY) scripts/build_report.py --evidence "$(EVIDENCE)" --output reports/webfetch $(if $(filter 1,$(ENGLISH)),--include-english,)
 
 report-check: runtime-check
 	$(PY) scripts/check_report.py reports/webfetch
@@ -54,7 +54,7 @@ wcxb-release-check: runtime-check
 
 browser-use-report: runtime-check
 	@test -n "$(EVIDENCE)" || (echo 'EVIDENCE="/absolute/path [/absolute/path...]" is required' >&2; exit 2)
-	$(PY) scripts/build_browser_use_report.py $(foreach path,$(EVIDENCE),--evidence "$(path)") --output reports/browser-use
+	$(PY) scripts/build_browser_use_report.py $(foreach path,$(EVIDENCE),--evidence "$(path)") --output reports/browser-use $(if $(filter 1,$(ENGLISH)),--include-english,)
 
 browser-use-review: runtime-check
 	@test -n "$(EVIDENCE)" || (echo 'EVIDENCE=/absolute/path is required' >&2; exit 2)
@@ -73,7 +73,7 @@ browser-use-release-check: browser-use-report-check
 
 trajectory-replay-report: runtime-check
 	@test -n "$(EVIDENCE)" || (echo 'EVIDENCE=/absolute/path is required' >&2; exit 2)
-	$(PY) scripts/build_webarena_replay_report.py --results "$(EVIDENCE)"
+	$(PY) scripts/build_webarena_replay_report.py --results "$(EVIDENCE)" $(if $(filter 1,$(ENGLISH)),--include-english,)
 
 trajectory-replay-report-check: runtime-check
 	$(PY) scripts/check_webarena_replay_report.py reports/trajectory-replay
@@ -106,7 +106,7 @@ resource-image:
 
 public-benchmark-report: runtime-check
 	@test -n "$(EVIDENCE)" || (echo 'EVIDENCE=/absolute/path/to/reviewed-public-benchmark-evidence is required' >&2; exit 2)
-	$(PY) scripts/build_report.py --evidence "$(EVIDENCE)" --output reports/webfetch-public-benchmarks
+	$(PY) scripts/build_report.py --evidence "$(EVIDENCE)" --output reports/webfetch-public-benchmarks $(if $(filter 1,$(ENGLISH)),--include-english,)
 	$(PY) scripts/check_report.py reports/webfetch-public-benchmarks
 	$(PY) scripts/check_report_links.py reports/webfetch-public-benchmarks
 

diff --git a/evaluation/browser-use/current-report.json b/evaluation/browser-use/current-report.json
@@ -35,6 +35,10 @@
     "source_path": "research/browser-use-failure-cause-review.json",
     "source_sha256": "a11134919f5e7b9a65c601374211e20e7e940f82e6f363ed025c49b52336e721"
   },
+  "trajectory_replay_bridge": {
+    "source_path": "research/browser-use-trajectory-replay-bridge.json",
+    "source_sha256": "731c04735e6b58d271d4ed5d43514300060c3ecfca0f54d7bc826c1300dd2ac5"
+  },
   "publication": {
     "base_url": "https://artifact-hub.app.lexmount.com",
     "site_slug": "7l77EeXPRfml",

diff --git a/evaluation/trajectory-replay/webarena-verified/README.md b/evaluation/trajectory-replay/webarena-verified/README.md
@@ -1,10 +1,26 @@
 # WebArena-Verified verified trajectories
 
-`trajectories.json` is the compact, versioned trajectory portion of the benchmark dataset.
-Each trajectory is one verified feasible BrowserTool path, not the task's only correct path.
+`current-run.json.dataset.path` and its SHA-256 identify the authoritative trajectory
+input for a formal build; currently this is `trajectories-corrected.json`.
+`trajectories.json` is the retained predecessor used by qualification and correction
+provenance. It is not a second current dataset. Each trajectory is one verified feasible
+BrowserTool path, not the task's only correct path.
+
+WebArena task definitions and shared evaluator fixes belong in the
+`lexmount/webarena-verified` fork. BrowserTool actions, selector compilation, local
+qualification and method comparisons belong here: the upstream task dataset does not
+define those browser-tool trajectories. Historical correction catalogs in `revisions/`
+are immutable inputs bound by retained cohorts, not copies of report output; removing
+them must not make historical contracts impossible to resolve.
+
+Authored report content and illustrations live outside Git. The report command takes
+`--content` (default `$BROWSER_EVAL_DATA_HOME/report-assets/trajectory-replay`, with the
+data home defaulting to `~/.browser-eval`). `research/trajectory-replay-content.json`
+binds every consumed file. The prior Git revision retains migration recovery copies;
+published report versions and physical execution evidence remain unchanged.
 
 `current-run.json` is the run authority for the next formal replay. It freezes the same 90-task
-selection used by the current Browser Use evaluation, the two registered browser methods, resource
+selection used by the current Browser Use evaluation, the registered browser methods, resource
 limits, and the isolation schedule. The private ready-to-run seed manifest must contain exactly one
 canonical candidate for each of those 90 tasks. The current inventory is assembled from 60 clean
 autonomous successes and 30 reviewed repairs; successful executions from different methods are
@@ -34,23 +50,33 @@ candidate evidence, not additional tasks.
   `BROWSER_EVAL_DATA_HOME`; this directory stores deterministic actions and provenance hashes.
 - The current self-hosted WebArena deployment has one fixed-port site stack, so a formal suite owns
   the physical environment exclusively. It runs every task for one browser before starting the next
-  browser, restores the official fixture before every task, and tears down both the browser and site
-  containers between browsers. Parallel methods are valid only after separate site stacks also have
-  separate containers, data stores, ports, gateways, and lock namespaces; multiple browser
+  browser, restores the site stack once before each browser method, and tears down the browser and site
+  containers between methods. It does not reset the site between tasks. Parallel methods are valid
+  only after separate site stacks also have separate containers, data stores, ports, gateways, and
+  lock namespaces; multiple browser
   containers against one site stack are not independent environments.
-- Formal execution requires clean browser-eval and CowAgent checkouts and binds their commits, the
-  resolved browser and site image identities, and the frozen CPU, memory, and sampling limits into
-  the resumable receipt. A resume with different runtime identity is rejected. Browser storage,
-  cache, cookies, permissions, service workers, and IndexedDB are cleared before every task; failed
-  teardown aborts the suite. Any pre-action context recovery writes a new physical-evidence
-  directory instead of overwriting the first one.
+- Formal execution binds each method to frozen actions, a browser-binary hash, the site map, and a
+  method-scoped site reset receipt. A fresh BrowserTool context is created for each task; browser
+  storage does not carry between tasks, while site data is not reset between tasks. Every physical
+  execution and its evidence is retained separately.
 - More than one trajectory may be retained for a task when paths are materially different and
   independently satisfy the same promotion gate.
 
-Run the frozen suite with `scripts/run_webarena_trajectory_replay.py`. Diagnostic subsets are
-explicitly non-promotable. `--qualification-engine chromium` runs the complete frozen task set on
-the reference browser and is the only scope that may promote the standard trajectory dataset.
-After all 90 qualify, promote that receipt with `scripts/promote_webarena_trajectories.py`, then run
-the measured method matrix with `--qualified-dataset` and no scope option. The runner rejects formal
-execution when that dataset is missing, incomplete, or differs from the frozen profile, seeds, or
-canonical actions. Never edit a verdict or verification claim by hand.
+Freeze each method's exact 90-task inputs with `scripts/freeze_webarena_three_method_inputs.py`,
+run each complete method cohort with `scripts/run_remote_three_method_cohort.py`, and bind the
+certified cohorts with `scripts/aggregate_webarena_three_method_cohorts.py`. Each method
+receives one site reset before its first task. The runner rejects a changed binary, task set,
+action sequence, or reset policy. Diagnostic subsets never enter the formal matrix. Never edit a
+verdict or verification claim by hand.
+
+The report presents every registered method with its recorded protocol. Diagnostic-only cohorts cannot
+enter its score-bearing matrix or add a separate result appendix.
+
+Adding a browser version appends a method and its complete cohort; existing results remain
+unchanged. Historical failure diagnostics may retain their original cohort binding through
+`inherited_result_cohorts` in the review record. Each inherited cohort must be hash-bound, use
+the same dataset, and have every logical result preserved exactly in the expanded matrix.
+When appended cohorts use a corrected evaluator, declare the exact method-specific correction ID
+and corrected task hash under `acceptance_revisions`. The task page retains both acceptance
+versions, and measurement notes identify the affected task IDs. Identical frozen trajectory
+hashes do not prove identical evaluator revisions. Historical verdicts are not rewritten.