Categories overlap. Runtime includes authored fixtures and negative controls. Family labels group related evidence; they are not counts of independent applications.
Frozen pilot after adjudication: named 18/24; prose 18/24. Initial unsupported-claim count: 2 per condition (historical only); adjudicated flags remain qualitative, with no revised numeric count. Independent human validation: pending. Decision time: unknown. [Adjudication](../../research/evaluations/language-pilot-01/submitted-review-01/adjudication.md).
One frozen development pilot with submitted model review; six matched briefs. Equal aggregate scores do not establish equivalence. Stable anchors are consumed development cases, not fresh held-out tasks.
Ten selected retrospective outcomes (not a historical census), followed by prospectively registered bounded cycles. Retrospective effort remains unknown.
Declared groups require a review of shared evidence tier, question granularity, evaluation and effort accounting. A matching label is not proof of comparability. [Machine-readable window audit](../../research/evolution/ComparisonWindows.json).
Unknown or zero effort is not dropped to create a favorable window. Qualitative review can retain those cycles, but effort comparisons need an explicitly reviewed window with measured positive research time. No trend or plateau is inferred.
One bounded question counts at most one consequential outcome. Corroboration and inconclusive work remain visible separately. Fewer than three declared cycles: qualitative review trigger pending; no trend assessment.
One bounded question counts at most one consequential outcome. Corroboration and inconclusive work remain visible separately. Fewer than three declared cycles: qualitative review trigger pending; no trend assessment.
One bounded question counts at most one consequential outcome. Corroboration and inconclusive work remain visible separately. Fewer than three declared cycles: qualitative review trigger pending; no trend assessment.
One bounded question counts at most one consequential outcome. Corroboration and inconclusive work remain visible separately. Fewer than three declared cycles: qualitative review trigger pending; no trend assessment.
One bounded question counts at most one consequential outcome. Corroboration and inconclusive work remain visible separately. Fewer than three declared cycles: qualitative review trigger pending; no trend assessment.
One bounded question counts at most one consequential outcome. Corroboration and inconclusive work remain visible separately. Fewer than three declared cycles: qualitative review trigger pending; no trend assessment.
One bounded question counts at most one consequential outcome. Corroboration and inconclusive work remain visible separately. Fewer than three declared cycles: qualitative review trigger pending; no trend assessment.
One bounded question counts at most one consequential outcome. Corroboration and inconclusive work remain visible separately. Fewer than three declared cycles: qualitative review trigger pending; no trend assessment.
Profile hashes and local references passed validation when this report was generated. That checks traceability, not source truth or external URL freshness.
Profiles with unknown last substantive review date: 0. Assessment dates record this metadata review, not a new substantive validation. Recorded substantive dates identify scoped author-model reviews, not renewed empirical or external-source freshness.
Open limits: Authorized retention and export are separate decisions. Complete contract recurrence and fault-tested deployed behavior remain unverified.
Open limits: Process-kill recovery observed in SQLite fixture; power-loss, device-loss backup and full frontend contract remain unverified. Local acceptance is not off-device recovery.
| runtime | authored SQLite / Git recovery fixture | SQLite atomic supersession/restart plus Git identity replay after lost local acknowledgement; one authored integration, no frontend recurrence. | [StorageReceiverRecovery](../../research/reviews/StorageReceiverRecovery.md) |
Open limits: Atomic supersession observed in SQLite fixture; concurrent in-flight supersession, UI states and general receiver retention remain unverified.
Open limits: No native-unload, crash, multiple-tab or access-revocation guarantee. Full base/draft/current merge contract is not verified in deployed frontends.
| runtime | authored release fixture | Related conditional activation-map mechanism in one process, not the full frontend pattern. | [ActivationConflictsAndReload](../../research/reviews/ActivationConflictsAndReload.md) |
| documentary | React Router | Cancellation does not imply server rollback. | [TransactionalEvidenceExpansion](../../research/reviews/TransactionalEvidenceExpansion.md) |
| static-code | Element / Matrix | Pinned client and SDK transaction/echo paths. | [FrontendRecoveryImplementations](../../research/reviews/FrontendRecoveryImplementations.md) |
| runtime | Element / Matrix | Real SDK with synthetic transport: lost acknowledgement, identity reuse and authoritative echo. | [MatrixRecoveryExperiment](../../research/reviews/MatrixRecoveryExperiment.md) |
Make a navigable state reconstructible from its URL.
| Evidence | Family | Scope | Review |
| --- | --- | --- | --- |
| documentary | htmx | History behavior and history-cache misses. | [NavigationRenderingEvidence](../../research/reviews/NavigationRenderingEvidence.md) |
| documentary | Web platform | pushState changes history, not server dereferenceability. | [NavigationRenderingEvidence](../../research/reviews/NavigationRenderingEvidence.md) |
| static-code | Zulip | ZT01-ZT04: traced target-message-lookup and anchor-navigation continuations to their unconditional completion points; no captured-list guard or test coverage located. No reproduction or application execution. | [ZulipTargetLookupCallPath](../../research/reviews/ZulipTargetLookupCallPath.md) |
Open limits: Independent application recurrence and human/AT validation remain unverified. Pending history disclosure, multiple history misses, redirects, nonzero swap delays, bfcache, lifecycle cleanup, interrupted/late enhancement and offline availability remain open. Zulip's target-lookup and anchor-navigation continuations have a traced absence of a captured-current-list guard and no located test coverage for a stale-navigation race (unreproduced, not a defect claim); deferred editing focus and search callback wiring remain separately unverified.
| runtime | htmx in authored public-catalog fixture | NF01–NF06: Tab/Enter navigation, raw versus authored focus/status policy, held-response ownership and background update; DOM evidence only, not AT speech. | [NavigationFocusExperiment](../../research/reviews/NavigationFocusExperiment.md) |
| static-code | Zulip | ZT01-ZT04: traced target-message-lookup and anchor-navigation continuations to their unconditional completion points; no captured-list guard or test coverage located. No reproduction or application execution. | [ZulipTargetLookupCallPath](../../research/reviews/ZulipTargetLookupCallPath.md) |
Open limits: Independent application recurrence and human/AT validation remain unverified. Pending history disclosure, multiple history misses, redirects, nonzero swap delays, bfcache, lifecycle cleanup, interrupted/late enhancement and offline availability remain open. Zulip's target-lookup and anchor-navigation continuations have a traced absence of a captured-current-list guard and no located test coverage for a stale-navigation race (unreproduced, not a defect claim); deferred editing focus and search callback wiring remain separately unverified.
Compute a result from explicit authoritative inputs.
| Evidence | Family | Scope | Review |
| --- | --- | --- | --- |
| documentary | Redux | Selector guidance, not a requirement for a global store. | [PatternBreadthExpansion](../../research/reviews/PatternBreadthExpansion.md) |
Open limits: Snapshot coherence and changed-locale cache test still planned.
| runtime | htmx in authored public-catalog fixture | NF01–NF06: Tab/Enter navigation, raw versus authored focus/status policy, held-response ownership and background update; DOM evidence only, not AT speech. | [NavigationFocusExperiment](../../research/reviews/NavigationFocusExperiment.md) |
Open limits: One implementation family; repeated trials are not independent uses. Animated exits and competing owners remain untested; no production repair claim. Route ownership is a contrast, not independent modal recurrence; AT speech and broader history/lifecycle cases remain unverified. Zulip auxiliary target/anchor callbacks, deferred editing focus and search callback wiring remain unverified; static guards do not prove whole-application ownership.
| static-code | react-microfrontends examples | Root config/navbar are one example family with shared startup gating. | [FeatureDeploymentBoundaries](../../research/reviews/FeatureDeploymentBoundaries.md) |
| runtime | authored release fixture | Publication/activation/rollback across fresh, loaded and lazy sessions. | [FeatureReleaseExperiment](../../research/reviews/FeatureReleaseExperiment.md) |
| runtime | authored release fixture | Conditional activation updates and safe reload policies, including negative controls. | [ActivationConflictsAndReload](../../research/reviews/ActivationConflictsAndReload.md) |
Open limits: No upstream deployment or real team/release-independence validation. Distributed atomicity, retention authorization and native-unload/crash behavior remain open.
Declare a UI capability's semantic surface separately from its runtime-isolation mechanism.
| Evidence | Family | Scope | Review |
| --- | --- | --- | --- |
| documentary | Web Components | CC01: Custom Elements observedAttributes/lifecycle callback mechanism, narrower than a full semantic contract. | [CapabilityContractSources](../../research/reviews/CapabilityContractSources.md) |
| documentary | webpack | CC02: Module Federation expose/consume module and share-scope dependency-negotiation mechanism, one layer below any UI-facing contract. | [CapabilityContractSources](../../research/reviews/CapabilityContractSources.md) |
Open limits: No known use of an explicit capability contract, as distinct from its substrate mechanisms, has been located. Enforcement, versioning and compatibility checking are unverified; a contract with no tooling is only documentation. Only one family per substrate mechanism inspected; no independent recurrence.
Remaining uncertainty: No CDN or upstream deployment validation.
Supersedes: none recorded.
### LEARN-0010 — How do stale activation writers and unfinished drafts constrain updates?
2026-09-26 · stream E · retrospective · revision
Before: Release experiment left activation conflicts and reload policy unexecuted.
Conclusion: Conditional replacement rejects stale writers; explicit reload policies preserve or refuse transition, with schema recheck after navigation.
### LEARN-0011 — Can a public route preserve selection and complete-document availability across native/enhanced entry, history and HTTP caching?
2026-09-27 · stream B · prospective · corroboration
Before: Native href/GET contracts plus representation-aware routing and cache selection should satisfy this narrow obligation; documentary support lacked runtime checks.
Conclusion: Corroborated in pinned htmx fixture: native and history paths work; missing Vary reproduces fragment-as-document reuse in both browsers. Existing obligations stand; no new pattern or promotion.
Remaining uncertainty: One authored application, no independent recurrence or usefulness gain. Accessibility, delayed scripts, interrupted transport, real intermediaries and offline availability remain open.
### LEARN-0012 — Does enhanced route replacement need explicit continuation and feedback policy that respects changed focus ownership?
2026-09-27 · stream F · prospective · corroboration
Before: Native controls alone do not establish enhanced focus/feedback semantics; delayed responses should respect focus moved to another surviving task. Modal return is not automatically route continuation.
Conclusion: Corroborated in 22 cases: raw same-ID link focus survives but this fragment contract leaves stale title and silent failure; authored heading/status policy works and preserves newer search focus. No AT speech or universal heading policy established.
Remaining uncertainty: Author-model DOM evidence, no human/AT validation or independent recurrence. Delayed history focus, competing requests, delayed scripts and interrupted enhancement remain open.
### LEARN-0013 — Can obsolete history/navigation completions affect newer intent, and does separating completion ownership from focus ownership prevent it?
2026-09-27 · stream F · prospective · revision
Before: The single-response continuation policy preserved newer search focus but left competing history/success/error completions unresolved; focus-at-swap alone might not be sufficient.
Conclusion: Revise operational guidance: validate navigation ownership before content/URL/title/feedback effects, including stale errors and newer failure, independently of focus permission. Prior policy misses obligations in 12/12 adversarial rows; authored guard meets 12/12. Pause fixture work for independent source contrast.
Remaining uncertainty: One authored htmx family and author-model review, not production recurrence or AT validation. History adapter is pinned implementation-dependent; pending-state, multiple-miss and lifecycle limits remain open. Three distinct protocols are not a comparable diminishing-returns window.
### LEARN-0014 — Where does independently developed Zulip scope navigation/search result and focus applicability, and where does our model fail to establish a guarantee?
2026-09-27 · stream B · prospective · revision
Before: The authored navigation generation guard protected current-view and focus effects, but independent application recurrence and alternative ownership/cache policies were unverified.
Conclusion: Revise guidance toward per-effect applicability: pinned Zulip preserves noncurrent cache data while guarding main selected-view continuations by list identity. Make client-fragment addressability explicit without inferring no-script enhancement. Shared feedback, auxiliary callbacks and deferred focus limit the inference.
Remaining uncertainty: Static targeted ranges in twelve files, no Zulip runtime/upstream test execution or independent human/AT review. Auxiliary lookup/anchor continuations, deferred focus and callback wiring need a bounded follow-up before any race claim. One Zulip family supplies partial mechanisms, not full-contract recurrence or a no-script guarantee.
### LEARN-0015 — Do the six authority/projection seeds denote reusable patterns or a mixture of scoped acceptance, projection placement and delivery choices?
Before: Authority is scoped and acknowledgement levels differ, but six seed names lack developed contracts and their projection/delivery relationships are not explicitly compared.
Conclusion: Revise the research decision aid to separate scoped acceptance, projection computation and delivery timing with compatibility constraints. Six seed dispositions and two synthetic traces; retain all seeds, zero new drafts or candidate maturity changes.
Remaining uncertainty: Six living primary documents across four families, not independent full-contract implementations; exact revisions unknown. Synthetic VC-01/02 reasoning is not execution or demonstrated usefulness; authority/merge/storage and phase-specific recovery contracts still need application evidence. Existing candidate profiles unchanged; five prospective groups are not a comparable learning trend.
### LEARN-0016 — Does the six-question authority/projection aid distinguish local editable content, remote shared acceptance and subsequent projection in a pinned application?
Before: Scoped acceptance, projection location and delivery schedule are separate questions with compatibility constraints; a local display need not establish shared acceptance.
Conclusion: Corroborated in pinned Zulip content editing: conditional local Markdown echo precedes remote validation; locked optional content checks and server rendering precede save/event scheduling; HTTP and event application are distinct paths. Same displayed object can carry tentative content. No candidate or maturity change.
Remaining uncertainty: Ten selected source/test files from one existing Zulip family, no application/database/upstream test execution or full-contract recurrence. Response/event ordering, lost response after acceptance and current-edit ownership need preregistered reproduction; no defect or safe-replay claim. Six prospective groups are not a comparable trend; original human and milestone gates remain open.
### LEARN-0017 — Do Zulip's target-message lookup and anchor-navigation fetch continuation guard against a stale navigation that occurred while the request was in flight, the way the main narrow-fetch continuation (ZN02) does?
2026-09-27 · stream B · prospective · corroboration
Before: ZN07 located both callbacks and noted no visible captured-current-list guard, but did not trace their callers or check test coverage.
Conclusion: Traced both continuations to their concrete completion points: process_result calls opts.cont unconditionally, and neither the target-lookup channel.get nor load_messages_around_anchor supplies opts.msg_list, so neither participates in the opt-in list-identity guard that message_fetch.ts implements for other callers. No located test exercises a stale-navigation scenario for either path.
Remaining uncertainty: Static source trace only; no application, database or upstream test executed, and no race actually reproduced. show() performs its own state reset on entry, so a stale re-entry's externally observable effect is unverified without execution. Second cycle in comparison group B-independent-navigation-source-v1; still short of three comparable cycles, and the workplan's rule against manufacturing cycles to reach that count still applies.
### LEARN-0018 — Which process-recovery obligations can SQLite plus actual Git receive-pack satisfy, and where does object identity fail to supply an operation receipt?
2026-09-28 · stream D · prospective · corroboration
Before: Atomic local intent/outbox updates and stable receiver identity support recovery within explicit transaction, ancestry and retention boundaries.
Conclusion: Nine checks corroborate atomic SQLite process recovery, Git replay and server ancestry refusal; split-transaction and fresh-identity negative controls retain the limits. One authored integration, not a full frontend use.
Remaining uncertainty: Power-loss, device-loss backup, remote durability and arbitrary operation-ID retention remain unverified. No full frontend contract, independent human validation or M1/M2/M3 acceptance. New runtime group is not comparable with static studies; no learning trend.
### LEARN-0019 — Does named organization change constraint/recovery recommendations versus identical guidance under numbered headings on fresh authored variants?
Before: Consumed pilot-01 model adjudication scored 18/24 in each condition; naming benefit and equivalence remain unestablished.
Conclusion: The preregistered 12-attempt run produced no retained answers. All six pairs are excluded; no quality comparison or naming conclusion is possible. Endpoint health checks failed inside and outside the sandbox.
Remaining uncertainty: No usable generation or empirical rating; failed attempts are not zero scores. Provider usage and billing are unknown; no retries or additional calls authorized. Active research effort unknown; request elapsed time measures failed execution only. Independent human validation and real consumer outcomes remain absent; this new group cannot close a comparable three-cycle window.
