Long-horizon progress: a product reading journey

• Stage: proposed product refinement; no shipped UI, evaluator qualification, metric promotion or experiment result is established by this note.
• Ownership: research leads the overall assessment design and calibration; product owns the reading journey and presentation refinement. Repository integration must reconcile both contributions against the current owners.
• Language: 中文版.
• Existing owners: Live Team Workspace, Goal Workspace, research evaluation and auto-research metrics.

The question to answer

“The score looks flat. Did the later work change anything useful, why did the direction change, and what actually supports that account?”

The default reading object is the whole recorded research history, including a study lasting well over ten hours: competing routes, multiple attempts, evidence, turning points, reversals, waiting and gaps in one graph. A research episode is an inspectable detail within that history: a question, a discriminating attempt, an observable result and the next decision. An episode may span Turns; concurrent attempts retain their identities. The proposed research unit and rubric require research qualification before metric use. A displayed grouping does not create an evaluation unit or a new writable graph.

Show the original outcome curve together with the whole-history Explore graph; selection expands an episode's evidence and decision. Preserve the authority of the native scorer and existing board. A graph, model summary, completed Todo or accepted artifact cannot rewrite its score or establish causal benefit.

Shortest ordinary journey

Count deliberate user actions from an already-open Goal to an answer backed by inspectable evidence. This is an acceptance target, not a measurement of the current product. Opening the Goal from elsewhere has its own navigation cost.

Actions	What the person can do	Product requirement
0	Read the full covered history and native result, including competing routes, failed/unfinished attempts, reversals, later work and gaps	Default whole-history Explore in the existing Goal; no setup, provider choice, extra model call or manual graph construction; requested/observed coverage and unloaded pages explicit
1, optional detail	Select a route, attempt or episode and answer what changed, why, which artifact/version was checked and whether anyone used it	Open decisive inline evidence and highlight the same history window/branch; return to the full graph without losing context
1, optional analysis request	Ask the Agent for a deeper account of this history	Reuse the benchmark owner's deep trajectory evaluation and a valid matching assessment, or its existing authorized analysis path; no second evaluator, manual ids or extra approval when intent and scoped authority already exist
2, optional	Compare that episode/window with a baseline or earlier checkpoint, or inspect its original source	Preserve selection and return focus; label whether a matched comparison is available rather than manufacturing one

Source opening is optional only when the inline detail already contains the decisive audience-authorized evidence. If the fact cannot be shown, say what is missing; do not count a summary badge as an evidenced answer. Extra actions need concrete value, such as choosing a comparison basis. Reading and replay never start analysis or a run, submit feedback or add an approval. The optional analysis request uses its existing execution, permission and budget owner; it is not an effect of selecting an episode. Any later write uses its existing action owner and authority boundary.

One focus across time, branches and evidence

The existing Goal workspace owns navigation and selection. The Live Team Workspace projection owns its timeline/map/list rendering. Explore and the capability's evidence records own relationships and source identities.

Existing research Goal
  Default: whole-history Explore + native outcome curve + observed coverage
    Questions / competing routes / attempts / evidence / decisions / gaps
                    select once (optional)
  Focused episode / branch  <-->  inline source, artifact version, check,
                                  next decision, adoption and missing evidence
                    request analysis once (optional, through existing authorized execution)
  Benchmark deep-trajectory assessment --> generated explanation + citations
    Same versioned inputs/result; factual history remains available

Keep one primary reading focus. The score curve, small branch context and detail panel share a selection; do not present three independently navigated dashboards. Avoid a new wall of summary cards, unqualified progress percentages, activity animation or an unreadable all-node view. Fit the full covered interval first; semantic grouping and progressive detail reduce density without dropping old anchors, failures, abandoned/unfinished routes, later work, waiting or gaps. Routine work is collapsible; retirement, invalidation and disagreements remain discoverable. Unloaded pages or retention limits prevent a complete-history claim. Task dependencies, research links and outcome attribution remain distinct; missing attempt times or explicit relationships stay unknown. Compact screens use the same ordered evidence chain vertically, with keyboard access, visible focus and reduced-motion parity. Meaning cannot depend on color or motion alone.

When only Explore records are available, label the view as recorded Explore history. recorded_at is record time, including delayed analysis and historical backfills; it does not establish when a native attempt occurred or complete route/adoption coverage. Parent/topic containment is grouping, including older projections that encoded it as supports with materialized_from=node_parent_id; it is not scientific support. Preserve explicit source relationships without treating their presence as independent verification.

The display must preserve these distinctions:

Reading lens	Decisive observation	What it does not establish
Result	Native outcome value, metric direction, source and scoring conditions	Generalization or improvement under a different evaluator/version
Information	A supported exclusion, uncertainty change or discriminating result linked to a next decision	Benefit merely because a failed attempt or negative row exists
Adoption	A specific receiver used a specific artifact/version in an observed action	Delivery alone, scientific correctness or improved outcome
Realized benefit	An independently qualified downstream outcome under the owning acceptance	Causal uplift from temporal adjacency or joint enablement

Reliability and cost qualify all four lenses; they are not merged into a scalar progress score. Evidence badges separate an observed fact, a model interpretation and a version-specific independent check. An interpretation cites its basis and can abstain. A disputed interpretation remains disputed; selecting it does not promote it to a fact. These are presentation distinctions, not new protocol enums.

Record relationships before the next study

This refinement is based on the whole-history proposal #6648 at f657a65c6abd03f1f6b87937c327f4ccfb2ccde2: a default record-time view, optional relationship view, and inline recorded evidence. That source is a presentation proposal, not installed behavior or completed V11/V12 acceptance. It does not manufacture missing research relationships or automatically compare native outcomes. Its timeline target-selection defects and independent review remain separate product work; a corrected head needs its own readback.

The first useful improvement is to retain why the next experiment follows from the previous result. At a meaningful conclusion or route change, the research Agent should reuse the existing question and attach an evidenced relation during ordinary result writeback. Users do not enter IDs, draw edges or approve individual links. This is study guidance, not a new mandatory attachment on every Turn or a graph-density quota.

Question the relation answers	Existing representation	Evidence needed
What question or prerequisite does this attempt address?	Explicit question/hypothesis/experiment nodes; answers or depends_on, with direction matching the assertion	Attempt identity, tested inputs and an actual prerequisite or answer
Which claim does this result support or contradict?	supports or refutes between existing nodes	Scoped finding, permitted result reference, uncertainty and applicable input revision; a failed run alone is not a refutation
Why was this next route selected?	Explicit decision references; leads_to only for an evidenced consequence	The decision cites its basis; later artifact or behavior separately establishes adoption
Which topic contains this work?	subtopic_of or an existing parent projection	Topic membership only; never count it as scientific support

Use the existing result attachment: question_update and up to four relations may accompany a complete result. Each relation runs from the current question to an existing same-Goal node; its summary states the observed basis. This hook does not create an arbitrary experiment/artifact graph. Materialize meaningful hypothesis, experiment or artifact objects through the existing Explore owner when the case requires those identities; do not recast every finding as another question just to add a line. Missing targets, wrong scope, self-links, duplicates and prohibited cycles must retain their existing rejection behavior. A successful replay does not append another relationship or claim another adoption.

The current evidence boundary is incomplete. Findings carry evidence references and tested input revisions, but the edge writer has no direct evidence_refs or endpoint-version parameters. An edge summary is an assertion, not a durable version binding. Keep the edge event identity and source snapshot with the permitted finding/artifact references in the research analysis sidecar; if that association cannot be reproduced, report it as unknown. Do not send invented fields to the current strict writer. A future contract extension belongs to the existing Explore typed owner and must qualify writer, reader, checkpoint replay, CLI and packaged presentation together. It is not delivered by this document.

The default graph shows explicit source relations and parent grouping distinctly. Any later inferred association stays in the on-demand analysis layer with its sources, method revision and uncertainty; it does not silently become a source edge. Same-checkpoint lineage, matched comparison, implementation inheritance and actual adoption are different facts. Neither supports nor leads_to is a catch-all for them. Record existing exact identities in the sidecar until a real caller and accepted contract justify extending the vocabulary.

Before launching a new study that claims relationship-informed decision quality, research owns one end-to-end collection canary; product owns the corresponding reading check:

1. Follow one permitted real case from question/attempt through result, an explicit relation and the next decision. Reproduce its source, artifact/input revision, record time and actual attempt time when available. Capture both the relation and the relevant evidence before replan; restore the checkpoint and verify the receiver can read the same basis without duplicate writes or future evidence.
2. Read it through the actual result owner and the packaged #6648-derived history view: the selected item must open its own evidence. Preserve missing links, isolated nodes, rejected writes, delayed backfills and changed-input limitations. If that view remains defective, report product acceptance open; do not treat screenshots or a private custom rendering as the shipped journey.
3. Inspect the receiver's next action separately. A relation being present or cited proves neither adoption nor a more useful experiment. Keep unsupported inferences and no-next-opportunity cases visible. Graph classification versus runtime penalty, Portfolio proxy versus native transfer, and checkpoint capture versus actual wake input supply three distinct diagnostic questions; select authorized evidence, not predetermined favorable conclusions.
4. Freeze the same collection instructions, fields and overhead accounting across compared arms. Agent-authored relationship guidance can change decisions and is therefore part of the treatment manifest; passive recording alone is not the same intervention. Keep TurnEnvelope separately off unless explicitly studied. Do not retrofit the new guidance into old frozen runs.

The launch requirement is the scoped collection and restore check for the claims being tested, not completion of all V12 deep analysis or a general provenance schema. Native-outcome studies with weaker observability must declare that limit. Coverage diagnostics count unresolved associations, missing versions and erroneous links against the declared opportunities; edge count and connectivity are not progress metrics. Public documents use generalized cases; raw trajectories and private study identities remain in the authorized research store.

Two layers, one trajectory-assessment basis

The Goal workspace contract owns the two-layer composition. Layer 1 projects existing authorized Explore records and explicit relationships with zero extra model calls. Layer 2 is a product consumer of the benchmark research owner's deep trajectory evaluation: one method and referenced versioned assessment, then a generated explanatory graph and conclusions. The whole run and prospectively fixed comparison windows own coverage, native outcomes and marginal cost; research defines cross-Turn episodes and explicit source/submission/artifact bindings in the same snapshot. Product rendering does not define a new unit, rubric or progress score.

The existing behavior findings/report support exploratory observations and interpretations, not scientific scoring. The toolkit's compact native Goal lifecycle summary is not a full history graph or quality assessment; the experiment board still owns native results and comparability. The current toolkit supplies versioned trajectory_assessment_input_v1 / trajectory_assessment_v1 artifacts and their store; the Goal reader exposes six uncalibrated judgments and bound citations. This bounded integration is not a unified scientific evaluation service or a calibrated model rubric. It cannot establish native utility or causal benefit.

When extending integration, inventory the existing producer and reconcile source snapshot/revisions, requested and observed time coverage, episode/event/artifact bindings, permitted audience, input digest, method/code revision, model/configuration only if used, qualification, actual cost, counterevidence, uncertainty, abstention and result identity. This is an interface requirement, not scientific evaluator qualification. Observations, Agent interpretations, independent checks and actual adoption retain separate provenance. Unqualified model confidence is not a scientific probability.

Keep explicit episode eligibility distinct from judgment uncertainty. Unknown eligible episodes remain in the supplied denominator; explicitly ineligible episodes do not become no or unknown. Show any declared exclusions beside the metric and identify the affected episodes in the same optional detail. Missing episode metadata or the difference between sample size and denominator cannot establish an exclusion. The reader formats the producer's eligibility and counts; it never recalculates scientific eligibility or the metric.

Retain a generated result's historical input basis. New inputs or method make its current applicability unestablished; a new assessment cannot silently overwrite it. Keep the factual graph readable through denied access/budget, in-progress analysis, cancellation, failure and retry. Repeated requests/navigation must not duplicate execution. Only the existing authorized Agent/execution owner may perform an analysis; opening this page grants no model budget, source access or intervention in a frozen benchmark. Analysis cannot borrow the solver's budget, cooldown or feedback opportunities. The source/evaluator restrictions below still apply.

Time and score alignment

Use the research owner's predeclared budget/checkpoint to define the later window. Show continuation elapsed time, inherited prefix and total source age separately. Do not choose “the latter half” after observing the best score, align incomparable restarts by wall clock alone, or charge concurrent work twice.

Keep recorded score points at their actual measurement/checkpoint and distinguish the best-so-far candidate from the current submitted candidate. Show gaps, waiting and censored endings; do not interpolate missing measurements as an improving or flat result. Selecting an episode highlights its interval and related recorded points, without drawing a causal edge to them. An unchanged best score can coexist with a new tested implementation or reusable exclusion.

A comparison exposes its task, model/harness, source, feedback, evaluator, budget and observation basis. Changed controls or unequal coverage remain visible. Without comparable measurements, say “comparison unavailable” rather than zero lift. Preserve the common denominator and the research owner's resource-weighted and episode-count views, including failures, incomplete work and abstentions. Do not silently discard unobservable baseline work. Human/model/tool cost, elapsed/active/wait time and assessment overhead retain their units and coverage; unknown cost is not zero. Existing usage owners provide interval costs; subtracting unqualified rolling totals is not an alternative ledger.

Synthetic example: exclusion, implementation and outcome diverge

Everything in this example is illustrative. It is not a real trajectory, benchmark result, calibrated rubric or installed screen.

flowchart LR
  Q[Question: can either approach improve the candidate?] -->|explicit experiment reference| A[Approach A tested]
  Q -->|explicit experiment reference| B[Approach B artifact v2]
  A -->|check identifies a violated assumption| N[Negative evidence]
  N -->|later decision cites that evidence| P[Exclude A from this next search]
  B -->|check bound to v2 and its inputs| V[Native check passed; best outcome unchanged]
  V -. adoption observation absent .-> U[Downstream use and benefit unknown]

In the declared later window, A changes a decision through supported exclusion. B reaches a version-specific native check without improving the incumbent. That is neither “no work happened” nor “the run improved.” A one-selection detail shows the assumption/check, exact artifact version, cited next decision and absent adoption observation. The original score remains unchanged. Dotted absence is not an adoption edge; a time-neighboring later score is not an attribution claim.

The real pilot should include an authorized exclusion, a genuine implementation that fails native transfer, a positive proxy under changed controls, and useful baseline work with weaker mechanism instrumentation. Research chooses the cases, rubric, independent coding and held-out calibration. This note supplies no weights, thresholds, automatic interventions or new experiment authorization.

Source, access and recovery

Research and experiments use the existing Goal workspace. Within one status source, the registry's declared parent_goal_id organizes a recursive sidebar: children start folded and the selected Goal's ancestors reopen on reload or browser Back. Goal ordering moves siblings only; it never changes parentage. Orphans and cyclic ancestry remain visible with an unresolved-parent indicator.

Selecting a parent exposes Only this Goal and its children. It opens the same frontend with a shareable rootGoalId display scope; All Goals returns to the current source's full workspace. A second frontend window can open that URL without copying, moving or rebinding live Goals. An unavailable root shows a retryable gap, without substituting unrelated Goals. The grouping conveys no Agent assignment, dependency, access, budget or lifecycle inheritance; each Goal retains its own execution state and acceptance. Child counts do not settle the parent's outcome. This same-source view does not itself connect experiment instances held in other registries or runtimes; their source-qualified read bridge remains a separate integration boundary.

An inline evidence detail needs the producing owner, source identity/revision, artifact version, observation/check time, freshness/coverage, explicit relationship basis and separate check/adoption observations. Inventory existing producers before defining a new field; reuse the shared typed presentation owner when a real caller requires an extension. Python may adapt permitted observations, not invent a second assessment or progress authority.

Case	Required reading behavior
Missing or incomplete observation	Show the gap and affected claim; unknown is neither zero nor negative
Restricted source	Explain limited access without leaking the source body; never replace it with a confident generated account
Artifact/input revision changed	Retain the historical check basis and mark current validity unestablished; do not keep a current-green badge
Model assessment disagrees with a check	Show both provenance and disagreement; original check/scorer authority remains intact
Stream gap, refresh, back navigation	Resnapshot within scope, preserve or explicitly invalidate selection/window, and never replay execution effects
No granular events or unsupported host	Show the available snapshot and freshness; do not simulate live progress

Only audience-authorized projections enter the product. The existing metrics boundary against parsing raw private traces/transcripts or protected evaluator bodies remains intact. Research may analyze permitted private material under its own contract and return safe versioned evidence/assessment references with coverage. Hidden evaluator material is eligible only after the solver has terminated and scoring has completed; it never returns to solver context or runtime feedback. Opening the view grants no additional source access, model budget or control power.

Smallest useful implementation and its acceptance

1. Design contribution now: reconcile this reading journey with the research design in existing roadmap/RFC/metrics owners. The synthetic example and link checks qualify documentation only; calibration and packaged behavior remain open.
2. Read-only product slice: one existing Goal entry with the default whole-history Explore graph, original score/window, complete/unknown coverage and optional one-selection evidence/check/adoption detail. Preserve multi-hour competing routes, attempts, negative/unfinished work and gaps over the existing projection. No new writable graph or automatic evaluator. Missing relationships stay visible; an implementation prerequisite must name its owner.
3. Packaged qualification: use the real backend and an authorized case to carry the 0/optional-1/optional-2-action reading journey through refresh/back, narrow viewport, keyboard use, old/missing anchors, absent timestamps/causal bindings, pagination, duplicate/out-of-order events, disconnect, Goal isolation, restricted/missing data and changed artifact inputs. Verify no execution or hidden evaluator feedback from reading/replay. Separately integrate the optional Agent request with the same pinned benchmark assessment: one authorized request reuses a matching result or executes once, with versioned citations, uncertainty/abstention, stale-input/method handling, denied budget/access, cancellation and failed/retried analysis while Layer 1 remains usable. Specification alone does not qualify that integration.
4. Comprehension and scientific qualification separately: compare against the existing list/map using predeclared questions about outcome, turning point, source, validation and uncertainty; record answer errors, time and action count, including false-progress and false-no-progress cases. Research independently qualifies the rubric, matched baseline, cost/window coverage and calibration.

Successful comprehension does not validate an evaluator; evaluator agreement does not prove a usable UI. Neither closes original small-team/live-stream acceptance, installs the proposal or qualifies downstream benefit. This design must not delay the existing L1/R2 correction journey until every research question is settled.

Open decisions for research/product integration are the episode unit and labeling rubric, window/censoring policy, comparable baseline coverage, source permissions, cost coverage and which existing producer can supply a versioned decision link. Do not turn these questions into mandatory user setup or pretend they are approved.