[loopx-agent]

Goal And Delivered Outcome

The launch articles overemphasized recent fixes, leaving the early control-plane transaction and execution-authority work, and the benchmark-driven evolution of research decisions, hard to understand. This revision rebuilds the bilingual stories around the full historical progression and concrete failure cases.

• TypeScript: explain provider success before checkpoint, prepared-attempt recovery, atomic claim/lease/receipt handling, evidence-only changes that must not rearm settled work, and semantic producer/consumer evidence. Recent recovery work becomes a continuation of that history. Preserve the existing cohort, contributor attribution and timeline.
• EdgeBench: follow trustworthy feedback → original-goal acceptance → experiment stride → complete checkpoint and fresh decisions → late-run outcomes and complete costs. Bring earlier native results and counterexamples into the main narrative; retain comparison conditions and distinguish initial advantage from subsequent improvement. Existing anchors are retained.
• Content capability: extend the existing artifact-revision guidance with chronological selection, canonical-source reuse, whole-film information density, listening, actual-player compatibility and explicit untested judgments. This is guidance, not a new runtime owner, renderer, permission or approval flow.

Author Declaration

• Written by: model_agent — OpenAI GPT-6 / Codex.
• Specification and revision: the editorial request summarized in this PR is the selection basis; AGENTS.md, docs/development/design.md, and existing creation-outcome guidance at base fb331f45c355152a4c1658af1e097f60a5840154 supply the delivery frame.

Criterion	Disposition	Path	Evidence
Full historical selection and synchronized bilingual articles	implemented	Four public blog documents	Chronological PR references, paired-case readback, source/condition preservation
Whole-page readability and first-screen evidence	implemented	Existing public blog surfaces	Desktop/narrow before and after below; populated and expanded details reviewed
Improve the existing creation capability's video revision guidance	implemented	loopx/capabilities/content_ops/README.md, packages/loopx-media/README.md	Guidance reuses the current owner and separates technical, creative and audience evidence

The existing whole-film workflow also checks the player-selected audio mix, opening narration and inserted shots: separate narration tracks may be selected as alternatives rather than heard together. Per-passage audio presence complements, and does not replace, full listening. This is documentation for the existing creation workflow, not a new media-provider runtime guarantee.

Scope And Continuation

Complete within this public documentation scope. A public release and final film/audio qualification remain separate work; this PR neither publishes a film nor claims that its creative review is complete. No runtime implementation or benchmark semantics changed.

Validation

Original article validation head: 5c30b51dd6a81fe6ff1d11db009c9082299a6e65. Public site TypeScript/build, SEO smoke, bilingual PR references, HTML IDs/fragments and public-boundary checks passed. The TS recovery example now distinguishes a conclusive version rejection followed by a failed read from an uncertain commit response, with immutable implementation/regression links. Blank-note omission is distinguished from explicit list/boolean intent; the repeated planning-migration paragraph is consolidated. This is an explanation of historical behavior, not a new runtime repair. Historical test source was inspected; those backend suites were not rerun. PR base remains main; the original branch merge base was fb331f45c355152a4c1658af1e097f60a5840154. At that validation, fetched main was 0d4bf99a4aa6080f2774f3c4477fb4adc1b5ac46; no intervening changes touch these two EdgeBench articles or their visual-evidence directory.

• Public site TypeScript/build and SEO publication smoke passed.

• Four HTML documents: one H1, unique IDs, valid local fragment links, existing anchors preserved, matching bilingual PR references.

• Actual browser review at 1440×1000 and 390×844; populated article content and expanded details checked. No document-wide horizontal overflow; wide tables retain their scroll containers. New TS examples read as concrete questions followed by recovery/execution mechanisms; benchmark conditions remain adjacent to the outcomes.

• Diff whitespace and added-text public/private scans passed. Screenshots show only public static article content. Generic existing .local documentation examples are unchanged.

• No benchmark rerun, new causal attribution, deployed-site readback, media-provider runtime certification or audience-comprehension claim. The separate canonical-film refinement and launch review bundle remain local candidates and are not part of this public diff.

• Scope-fit: reuse the existing content-ops and media review boundary; no new editor or abstraction. No RFC pair changed. Runtime suites are outside this documentation-only diff. CI/review/merge readiness remain separate.

• Run state: finished. Input class: public_fixture.

Check kind	Result	Evidence / limitation
static	passed	Public site build, SEO smoke, fragment/ID checks and public-boundary scan
manual	passed	Public article first screens, populated sections and expanded details at desktop/narrow sizes
real_backend	not_applicable	Documentation only; no backend contract changed

Review media relocation

Current head: 185e513305285172d3e55dcc7cfbca690e1a2de7; integration baseline: ce4c78088ea464cb31231513882bed6bb6933156 on main. The twelve review PNGs are excluded from the final repository diff. The four articles and two creation guides are byte-for-byte unchanged from 5c30b51dd6a81fe6ff1d11db009c9082299a6e65; no tracked source references the removed review-image directory. Original branch history is retained.

The same twelve images are preserved below as GitHub PR attachments: desktop 1440×1000, narrow 390×844. Capture revisions and original visual-review scope remain recorded beside them. This relocation changes evidence hosting, not article composition or the original review claims. The images were individually inspected for public-only content. Site TypeScript/build, SEO smoke, four-document ID/fragment and bilingual PR-reference checks, whitespace checks and the seven-file public-boundary scan passed on the integration; backend and audience tests were not rerun. The boundary scan retains its unrelated README credential-reference warning.

Frontend / Visual Evidence

Before: source baseline e19d94ca04ead2bb919aaa438e61bf229a6cfad1. The original first-screen captures below show 0df9e7266e2c6789bb3f014a762bd8e24bd8418f; the TS first viewport was rechecked at 441083a3c507854ab6e68b40c7609ac5c1284591 and retains the same composition. The revised recovery passage is captured separately at 441083a3c507854ab6e68b40c7609ac5c1284591. These are static source previews, not evidence of installation or production deployment.

Surface	Before	After
TS desktop
TS narrow
EdgeBench desktop
EdgeBench narrow

Current TS recovery passage: ordinary body text explains two failure sequences, with the decisive distinction in bold and source links adjacent. Inspected at 1440×1000 and 390×844; the English passage and companion review article were also checked at 390px. Both languages wrap fully without document-wide horizontal overflow. This is a composition assessment, not independent audience-comprehension evidence.

Current passage	Desktop	Narrow
Preserve known rejection vs commit uncertainty

• UI impact: changed editorial presentation; source data: public_fixture.
• Attention/craft review: keep one article title and the established layout. Concrete failure cases and their consequences occupy the narrative; methods and later implementation detail remain expandable. Numeric conditions and negative results remain visible near their claims. No new cards, motion or status system.

The EdgeBench narrative also restores the October 7 settled-work-Turn review repair (#5879): why completed-Todo cadence can miss a long task, why unsettled or duplicate records must not advance the counter, and how evidence can support continuing the same open task. The merged historical regression was inspected, not rerun; no benchmark result or frozen cadence was rewritten. The first viewport retains its composition and was rechecked at 5c30b51dd6a81fe6ff1d11db009c9082299a6e65 in desktop/narrow views. The new passage and English narrow view were also inspected, with no document-wide overflow.

Earlier research mechanism	Desktop	Narrow
Review before a long task completes

Type of Change

• Documentation update

LoopX Area

• Public docs or presentation surface (README, protocols, dashboard)
• Capability or extension (guidance only)

Shared-authority RFC fixture impact

N/A: this explains historical work; it claims no new migration, promotion, routing or provider-conformance progress.

Boundary Checklist

• No private state, credentials, raw traces, internal links or local machine paths in the public diff, description or screenshots.
• No duplicate benchmark implementation or new experiment.
• Scoped to the editorial and creation-guidance request.
• No paired RFC changed.
• Visual evidence included above.
• Every commit has a DCO sign-off.