docs: plan goal-centred replan study and intermediate measurements

[loopx-agent]

Goal And Delivered Outcome

The original Goal is already present in replan context; presence alone does not prove that an agent reassesses the most consequential remaining gap. This bilingual Draft separates additional whole-goal reflection (G) from a source-bound reading index (C), retaining the original Goal and complete ordinary input in all four cells.

The proposal now specifies an executable research sequence and measurement design: necessary correctness and new late-checkpoint qualification, the primary retained/fresh comparison, stride, G/C, then the lower-priority outcome-refinement candidate. It connects intervention delivery to actual adoption, research evidence, retained candidates, native outcomes and complete costs. It reuses the existing six trajectory judgments, with explicit episode eligibility, denominators, unknowns and analyzer qualification. This is documentation only: no shipped runtime, setting, scoring or experiment behavior changes.

Author Declaration

• Written by: OpenAI Codex, model agent, using the LoopX Agent contribution identity.
• Specification: the maintainer-requested whole-goal/context study and comparable intermediate measurements; existing measurement/episode acceptance in docs/architecture/rfcs/long-horizon-harness-benchmark-research-program-v0.md.
• Ownership: existing benchmark-toolkit with the native provider, TypeScript replan/context and trajectory-assessment owners. G/C remain local study labels, with no new protocol or scorer.

Criterion	Delivered / remaining
G0C0/G1C0/G0C1/G1C1	Preserved; fresh reconstruction, complete ordinary input and accessible evidence remain common. No extra retrieval, summary model or hidden stride activation.
Staged experimental program	Specified with late-point/source admission, paired capacity, common windows and cumulative budget accounting. The document launches no jobs.
Outcome-refinement candidate	Lower priority; inspected #6784 revision is named. No overall enabled switch is invented; optional checks are not the whole intervention. Equivalent questions merge; an incremental comparison needs an isolatable compatible common base.
Six process judgments	Existing goal_alignment, benefit_test, discrimination, information_gain, evidence_adoption, redundancy; metric-specific eligibility/window, yes/no/unknown/eligible, N/A and missing-data bounds. All remain uncalibrated_model.
Facts beside judgments	Delivery/use, feedback lag, proxy-to-native transfer, actual retention, initial/window-best/latest-valid/final, operational failures and full costs are separated, with implemented versus pending reductions stated.
Analyzer improvement	Curated positive/negative/unchanged/unadopted/missing/failure diagnostic designs, independent episode/citation review, blind assessment, disagreement and held-out calibration before scientific comparison. No measurement feedback enters frozen solvers.
Actual warm receiver and late scientific comparison	Unfinished author work. Synthetic transport qualification and older-source model characterization do not substitute.

Boundary And Remaining Author Work

The disposable prototype invokes the destination CLI once, preserves its current response and identity, and composes only the declared advisory from that response. Its framing/off/error/replay checks are bounded engineering evidence. Matching a few source files, a generated model-visible file, an accepted replan or a stored checkpoint does not prove current client delivery, complete recovery, later adoption or native benefit.

Remaining author work: qualify the actual pinned warm receiver/model-client path; freeze compatible late points, exact treatment inputs, episode boundaries/rubric and matched windows before outcomes; then complete the authorized continuation comparisons and independent assessment. Current-source model transport previously failed before tool use, and an older synthetic denominator remains unqualified. Preserve those failures and costs. Do not borrow an existing fresh arm as an unplanned control, expand the factorial without capacity, reset cooldown/opportunities or add another 36h after restoration.

Keep Draft for this concrete unfinished receiver and empirical work. The documentation revision does not mark the scientific study complete or change product defaults. Review, merge, installation and scientific acceptance remain separate.

Future-facing pass: reuse the typed reader, pre-decision selector, six-judgment vocabulary/key/reducer and existing native/usage owners. Automatic episode discovery, semantic citation entailment, causal adoption links, cross-session full accounting and independent evaluator calibration remain explicit gaps. No speculative production abstraction or parallel authority was added. A later caller-facing option still requires its frontend/CLI/Lark readback and disable journey.

Validation

Documentation head: fd90ba87bc3cb439f4badb7debf443f492b39724, synchronized with target main 55d6337485d4924f1cbb8b41e51a8c02687c332a. The complete PR diff contains only the two study documents and the existing lower-page benchmark-guide link.

Check	Result and scope
Relative links and anchors, fences, four cells, six existing metric names	Passed on both final documents.
Bilingual meaning and public boundary	Reviewed; no private trajectories, raw measurements, local paths, credentials or internal links published. First viewport and remaining-author-work text are unchanged in this increment.
Semantic advisory	Passed with zero changed source carriers; no shared vocabulary introduced. Advisory scope limits remain.
Bilingual RFC staged and exact-head checks	Passed; no existing RFC pair changed.
Whitespace, branch/PR base, DCO	Passed; target-main merge base verified and every branch-only commit signed off.
Prior isolated CLI transport	Bounded File/SQLite framing, unchanged values, off/error parity and same-guard lost-display recovery remain historical engineering evidence, not rerun or promoted to science here.
Runtime, model, native and scientific tests for this documentation increment	Not run. No production behavior changed, model analysis or scientific cell launched. Current warm receiver, matched late-point adoption/native/full-cost comparisons and calibration remain unqualified.

Frontend / Visual Evidence

UI impact: none. Plain Markdown research planning and an existing lower-page guide link. This increment leaves the opening viewport unchanged. It adds no packaged frontend, installed configuration or browser interaction.

Type Of Change

• Documentation update

LoopX Area And Technical Direction

Benchmark research protocol and public docs; existing roadmap S5/S11 and long-horizon measurement/episode acceptance. No new roadmap, authority store, provider migration, runtime routing, setting or shared protocol vocabulary.