diff --git a/docs/specs/2026-08-19-78-longhorizon-run-evidence-boundary.md b/docs/specs/2026-08-19-78-longhorizon-run-evidence-boundary.md
new file mode 100644
index 0000000..2d84f44
--- /dev/null
+++ b/docs/specs/2026-08-19-78-longhorizon-run-evidence-boundary.md
@@ -0,0 +1,522 @@
+# LongHorizon Run Evidence Boundary
+
+## Traceability
+
+- Spec ID: longhorizon-run-evidence-boundary
+- Story: #78
+- Status: Draft
+
+## Intent
+
+Define a small cross-repository integration proposal for how a future Better
+Harness source profile could read completed or otherwise terminal
+LongHorizon-Harness run artifacts as read-only session evidence. This document
+does not add an adapter, declare LongHorizon-Harness support, or make Better
+Harness part of LongHorizon-Harness execution.
+
+The proposal follows the scope invited by the
+[`phodal` MEMBER comment on Story #78](https://github.com/QoderAI/better-harness/issues/78#issuecomment-5262439512):
+describe the boundary and how a completed run could be read as evidence. That
+comment authorizes a proposal for discussion, not a merge commitment or a
+support commitment.
+
+## Pinned Evidence
+
+This Draft uses only facts visible at these two immutable commits:
+
+- Better Harness main:
+  [`72594cca8046d914dcf5e00f4da1f36bf0900021`](https://github.com/QoderAI/better-harness/commit/72594cca8046d914dcf5e00f4da1f36bf0900021).
+  Its [architecture](https://github.com/QoderAI/better-harness/blob/72594cca8046d914dcf5e00f4da1f36bf0900021/docs/ARCHITECTURE.md)
+  and [proposed ADR-0003](https://github.com/QoderAI/better-harness/blob/72594cca8046d914dcf5e00f4da1f36bf0900021/docs/adrs/harness-run-evidence-bridge.md)
+  are the Better-side evidence.
+- LongHorizon-Harness main:
+  [`be2e7b42523c4f35291f1ed57b683f6c03a29cdc`](https://github.com/AMAP-ML/LongHorizon-Harness/commit/be2e7b42523c4f35291f1ed57b683f6c03a29cdc).
+  The inventory below cites its
+  [`manager.py`](https://github.com/AMAP-ML/LongHorizon-Harness/blob/be2e7b42523c4f35291f1ed57b683f6c03a29cdc/src/lh_harness/manager.py),
+  [`types.py`](https://github.com/AMAP-ML/LongHorizon-Harness/blob/be2e7b42523c4f35291f1ed57b683f6c03a29cdc/src/lh_harness/types.py), and
+  [`control_bus.py`](https://github.com/AMAP-ML/LongHorizon-Harness/blob/be2e7b42523c4f35291f1ed57b683f6c03a29cdc/src/lh_harness/supervisor/control_bus.py).
+
+Better Harness main can also own Harness-as-Code runtime capabilities. The
+phrase "Better Harness is the evidence layer" in this proposal describes only
+the owner boundary for ingesting evidence from the external
+LongHorizon-Harness project. It is not a general restriction on Better Harness.
+
+LongHorizon-Harness PR #53, including commit `c2605bd`, is unmerged and is not
+an authoritative input. Its routing fields may be considered only as an
+optional future extension after merge and versioned conformance; they are not
+required by this proposal.
+
+## Ownership Boundary
+
+| Concern | LongHorizon-Harness owns | Better Harness owns |
+| --- | --- | --- |
+| Execution | Starting and running Manager, GUI/CLI Executor, and Auditor roles | No execution or role control |
+| Mutable state | Task state, task contract, round progress, operator control, and all run mutations | No writes to LongHorizon-Harness state |
+| Lifecycle | Terminal-status production and the meaning of retry, resume, cancellation, failure, and completion inside LongHorizon-Harness | Read-only qualification of an already-produced artifact set; no lifecycle commands |
+| Artifacts | Local and remote artifact production, file layout, schemas, and compatibility | Source-profile validation, normalization, and report projection for accepted inputs |
+| Product claims | LongHorizon-Harness behavior and artifact guarantees | Whether a validated projection is presented as Better Harness session evidence |
+
+The bridge is one-way: LongHorizon-Harness artifacts may flow into a future
+Better Harness reader, but Better Harness never writes normalized data,
+decisions, annotations, or control commands back into LongHorizon-Harness.
+
+## Architecture Precedent
+
+[ADR-0003](../adrs/harness-run-evidence-bridge.md) has `Status: Proposed`. It is
+an architecture precedent for owner-preserving, one-way evidence normalization,
+not an accepted or implemented decision for this integration. The existence of
+the current
+[`harness-run` analyzer](https://github.com/QoderAI/better-harness/blob/72594cca8046d914dcf5e00f4da1f36bf0900021/scripts/session-analysis/platforms/harness-run.mjs)
+does not change the ADR status. That analyzer reads Better Harness native
+HarnessRun artifacts; its file layout and `NormalizedToolActivityV1` mapping
+must not be imposed on LongHorizon-Harness artifacts.
+
+## Current LongHorizon Evidence Inventory
+
+The pinned Manager establishes a local log directory, writes run-level and
+round-level artifacts beneath it, and separately attempts remote mirror writes.
+Local terminal records remain authoritative when a later remote synchronization
+fails. Remote directory creation and writes catch errors and log that the trace
+write was skipped, so the remote copy is best-effort rather than a replacement
+authority.
+
+| Owner | Path at the pinned source boundary | Observed semantics | Versioning gap | Candidate input |
+| --- | --- | --- | --- | --- |
+| LongHorizon-Harness Manager | `<log_dir>/report.json` (also written to `<log_dir>/role_orchestration/report.json`) | Terminal run record. The success path [writes both local copies before attempting its two final remote mirrors](https://github.com/AMAP-ML/LongHorizon-Harness/blob/be2e7b42523c4f35291f1ed57b683f6c03a29cdc/src/lh_harness/manager.py#L943-L953). [`_final_report`](https://github.com/AMAP-ML/LongHorizon-Harness/blob/be2e7b42523c4f35291f1ed57b683f6c03a29cdc/src/lh_harness/manager.py#L1449-L1496) and the [crash boundary](https://github.com/AMAP-ML/LongHorizon-Harness/blob/be2e7b42523c4f35291f1ed57b683f6c03a29cdc/src/lh_harness/manager.py#L1547-L1636) establish `schema_version: 2`, a harness-level `status`, separate `completion_satisfied`, abort/failure fields, task state/contract, and serialized rounds. An existing local terminal report is preserved if a later operation fails. | Version 2 is explicit, but the pinned source does not publish a cross-project compatibility profile for Better Harness. | Required candidate authority, subject to the future source-profile qualification rules. |
+| LongHorizon-Harness Manager | `<log_dir>/role_orchestration/rounds.jsonl` | [Append-only local ledger](https://github.com/AMAP-ML/LongHorizon-Harness/blob/be2e7b42523c4f35291f1ed57b683f6c03a29cdc/src/lh_harness/manager.py#L2164-L2178). Each line is `asdict(ManagedRound)` and contains the [route, plan text, executor output, auditor report text, state, contract, related refs, and role status dictionaries](https://github.com/AMAP-ML/LongHorizon-Harness/blob/be2e7b42523c4f35291f1ed57b683f6c03a29cdc/src/lh_harness/types.py#L92-L104). | `ManagedRound` and the ledger have no independent explicit schema-version field at this commit. | Conditional supporting input after exact shape, bounds, and privacy validation. |
+| LongHorizon-Harness Manager | `<log_dir>/role_orchestration/events.jsonl` | Append-only event stream. Each [emitted event record](https://github.com/AMAP-ML/LongHorizon-Harness/blob/be2e7b42523c4f35291f1ed57b683f6c03a29cdc/src/lh_harness/manager.py#L2251-L2284) includes `schema_version: 1`, an event id, timestamp, event name, and event-specific payload. | The record version is explicit, but no profile-level whole-file or run-segmentation contract is observed. | Excluded from V1. A future event extension requires a separate approved versioned contract. |
+| LongHorizon-Harness Manager | `<log_dir>/role_orchestration/rounds/round_NNN/manager_input.txt`, `auditor_input.txt`, and `auditor_report.txt` | Selected orchestration text files with direct write sites for [Manager input](https://github.com/AMAP-ML/LongHorizon-Harness/blob/be2e7b42523c4f35291f1ed57b683f6c03a29cdc/src/lh_harness/manager.py#L306-L343), [Auditor input](https://github.com/AMAP-ML/LongHorizon-Harness/blob/be2e7b42523c4f35291f1ed57b683f6c03a29cdc/src/lh_harness/manager.py#L726-L741), and [Auditor report](https://github.com/AMAP-ML/LongHorizon-Harness/blob/be2e7b42523c4f35291f1ed57b683f6c03a29cdc/src/lh_harness/manager.py#L847-L904). These files can contain natural-language task or role content. | The files have no shared versioned envelope or independent schema marker. | Excluded from the initial profile because raw prompt/task/report text is outside the privacy boundary. |
+| LongHorizon-Harness Manager | `<log_dir>/role_orchestration/rounds/round_NNN/<role>_raw_trajectory.jsonl`, trajectory-derived screenshot artifacts, and `<role>_metadata.json` | [`_save_role_result`](https://github.com/AMAP-ML/LongHorizon-Harness/blob/be2e7b42523c4f35291f1ed57b683f6c03a29cdc/src/lh_harness/manager.py#L1742-L1829) persists role diagnostics and metadata. Screenshot artifacts are conditional on trajectory content and successful extraction; they are not guaranteed for every role or round. | These debug artifacts have role-specific formats and no common per-round schema version. | Excluded from the initial profile as sensitive debug evidence. |
+| LongHorizon-Harness environment mirror | `<harness_dir>/report.json`, `<harness_dir>/orchestration/report.json`, and `<harness_dir>/orchestration/rounds/round_NNN/round.json` plus selected round files | The success path names the [two final report mirrors](https://github.com/AMAP-ML/LongHorizon-Harness/blob/be2e7b42523c4f35291f1ed57b683f6c03a29cdc/src/lh_harness/manager.py#L948-L953); round recording [mirrors `ManagedRound` as `round.json`](https://github.com/AMAP-ML/LongHorizon-Harness/blob/be2e7b42523c4f35291f1ed57b683f6c03a29cdc/src/lh_harness/manager.py#L2165-L2178). The [remote helpers catch write/setup failures](https://github.com/AMAP-ML/LongHorizon-Harness/blob/be2e7b42523c4f35291f1ed57b683f6c03a29cdc/src/lh_harness/manager.py#L2181-L2221), so a missing mirror does not invalidate an existing local terminal record. | Delivery is best-effort and has no separate mirror-completeness manifest at the pinned commit. | Not an independent authority. A future profile may use it only under an explicit, validated discovery rule. |
+| LongHorizon-Harness Supervisor | `<run_dir>/control/owner.json` and `status.json` | The Supervisor [reserves a run with the source run id and workspace in both records](https://github.com/AMAP-ML/LongHorizon-Harness/blob/be2e7b42523c4f35291f1ed57b683f6c03a29cdc/src/lh_harness/supervisor/service.py#L1465-L1489), while the ControlBus provides [atomic owner/status persistence](https://github.com/AMAP-ML/LongHorizon-Harness/blob/be2e7b42523c4f35291f1ed57b683f6c03a29cdc/src/lh_harness/supervisor/control_bus.py#L672-L708). | These files have no profile-specific version field at the pinned commit. | Required only as the V1 qualification envelope for run identity, workspace, and lifecycle binding. Raw task, command, path, agent, and model values are not session evidence. |
+| LongHorizon-Harness control bus | `<run_dir>/control/commands.jsonl` and `command_receipts.jsonl` | [Separate control-plane records](https://github.com/AMAP-ML/LongHorizon-Harness/blob/be2e7b42523c4f35291f1ed57b683f6c03a29cdc/src/lh_harness/supervisor/control_bus.py#L1-L5). A receipt is the terminal authority for a control command, not for task success or run completion. | Control-bus revision and receipt semantics are owned by LongHorizon-Harness and are not a session-evidence schema. | Excluded from V1. |
+
+The following meanings stay distinct:
+
+- A **terminal run** is a local terminal report with a terminal `status` under
+  LongHorizon-Harness semantics.
+- **Artifacts available** means candidate files exist and pass the future
+  source-profile validation. It does not imply completion.
+- An Auditor `complete` result is one role-level judgment. `ManagedRound`
+  persists `auditor_report` as natural-language text, so that text alone is not
+  structured proof of success.
+- `completion_satisfied` is a separate harness-level field. Its presence does
+  not let Better Harness redefine LongHorizon-Harness lifecycle semantics.
+- **Task success** is a product-level interpretation that this inventory does
+  not infer from file presence, a natural-language report, or one status field.
+
+## Proposed Source Profile
+
+The normative proposal name is `longhorizon-supervisor-run-v1`. It is a Better
+Harness reader profile tied to the pinned LongHorizon-Harness commit and its
+observed artifact shapes. It is not a producer feature or a claim that
+LongHorizon-Harness publishes this profile.
+
+### Selection Boundary
+
+The future reader MUST accept one explicitly selected, Supervisor-managed,
+run-local directory. The selected directory basename is the candidate source
+run id. The reader MUST NOT scan a user home, enumerate a runs root, select a
+latest directory by modification time, guess a run from a fuzzy name, or adopt
+a shared direct-CLI log directory. Direct CLI output is outside V1 because it
+does not carry this Supervisor qualification envelope.
+
+This is a deliberate profile restriction. The pinned Supervisor's
+[`_run_dir`, `_run_logs_dir`, and run-id checks](https://github.com/AMAP-ML/LongHorizon-Harness/blob/be2e7b42523c4f35291f1ed57b683f6c03a29cdc/src/lh_harness/supervisor/service.py#L655-L704)
+support run-local selection, but do not constitute an upstream promise that
+future releases preserve this Better Harness profile.
+
+### Qualification Set
+
+| Class | Fixed run-local input | V1 rule |
+| --- | --- | --- |
+| Required qualification | `control/owner.json` | MUST parse as an object and bind the selected directory basename and requested workspace. It is not projected as session content. |
+| Required qualification | `control/status.json` | MUST parse as an object, repeat the same run/workspace binding, and contain a documented terminal Supervisor lifecycle status. It is not the native report-status authority. |
+| Required evidence authority | `logs/report.json` | MUST pass the exact V1 version, shape, terminal, and consistency rules below. It is the only native terminal-report authority. |
+| Conditional duplicate | `logs/role_orchestration/report.json` | MAY be absent. If present, it MUST parse and be canonical-JSON identical to `logs/report.json`; otherwise the set is `invalid`. The root report always wins by authority, never by guessing between conflicting copies. |
+| Optional supporting source | `logs/role_orchestration/rounds.jsonl` | MAY establish only allowlisted structural round consistency after the complete file validates. Absence makes an otherwise valid set `partial`; malformed or conflicting presence makes it `invalid`. |
+| Excluded | `logs/role_orchestration/events.jsonl`; remote `<harness_dir>` mirrors; `control/commands.jsonl`; `control/command_receipts.jsonl`; per-round text, state, contract, output, report, trajectory, screenshot, and metadata files; orchestration transcripts; `worker.log`; `tmp/task.md` | MUST NOT be read by V1. Owner/status qualification is the only control-plane exception. |
+
+Canonical JSON equality means equality after strict JSON parsing and recursive
+object-key ordering, while preserving array order, primitive types, and exact
+string values. It does not trim, coerce, repair, or discard unknown fields
+before comparing the two report copies.
+
+Within owner/status, V1 consumes only `run_id` and `workspace`, plus
+`status.status` for lifecycle qualification. Raw task, process id, command,
+agent, model, role configuration, idempotency, failure text, and other fields
+MUST be discarded after the bounded parse and MUST NOT affect evidence.
+
+### Identity, Workspace, and Path Safety
+
+Before any evidence projection, V1 MUST verify all of the following:
+
+1. The selected directory basename, `owner.run_id`, and `status.run_id` are
+   identical strings accepted without normalization or coercion. The string
+   MUST be 1-128 characters, not `.` or `..`, and contain no separator, NUL, or
+   control character, matching the pinned Supervisor validator.
+2. The canonical identity of the explicitly requested workspace equals both
+   `owner.workspace` and `status.workspace` after platform-correct canonical
+   path resolution. Only the match result and a privacy-safe opaque target
+   reference may leave qualification; the source paths MUST NOT.
+3. Every fixed descendant resolves beneath the exact selected run directory.
+   Every path component MUST be opened without following links, and every
+   selected file MUST be a regular file. Traversal, NUL, symlink/junction,
+   special-file, containment, or identity failure is fail-closed.
+
+The qualification result MAY use a deterministic opaque alias derived from the
+profile version and qualified source identity for safe internal diagnostics. It
+MUST NOT expose the raw source run id or workspace path, and it MUST NOT be
+interpreted as source retry lineage. This Draft does not map that alias to the
+existing Better Harness `sessionId`; any durable identity mapping belongs to a
+separately approved owner-extension or adapter spec.
+
+### Version and Shape Contract
+
+V1 accepts only the pinned LongHorizon-Harness shape with root report
+`schema_version` exactly integer `2`, `variant` exactly
+`lh_harness_role_managed`, and `mode` exactly `role_orchestration`. The report
+MUST carry typed `status`, `completion_satisfied`, `rounds_run`, `max_rounds`,
+and `rounds` fields. Native report status is closed to `complete`, `incomplete`,
+`blocked`, `cancelled`, or `failed`; booleans are not integers for numeric
+validation. `rounds_run` MUST equal the array length. `max_rounds` remains an
+observed native value only: the pinned human gate can extend the active round
+budget, so `rounds_run > max_rounds` MUST NOT by itself make the set invalid.
+`complete` and `completion_satisfied: true` MUST occur together.
+
+Each embedded round MUST have a strictly increasing one-based `round_index`, a
+`next_step` in the pinned `gui`, `cli`, `done`, `blocked`, `invalid`, or `ask`
+set, and object-valued role-status fields. V1 may validate bounded string fields
+for type and size, but MUST discard their contents rather than project them.
+Unknown unconsumed report fields MAY be ignored only after the required shape
+passes; they MUST NOT be projected or treated as support for a capability.
+
+The pinned report has no producer-version field. Consequently, V1 is specified
+only from the pinned commit, while runtime qualification can observe only the
+explicitly selected profile, schema 2, and exact conformance shape. It cannot
+verify or claim a broader LongHorizon-Harness release range. A future schema
+version is `unsupported`; same-version missing fields, type drift, new status
+values, or failed invariants are `invalid`, never loosely coerced into support.
+
+`rounds.jsonl` has no independent version. If present, V1 MUST parse the entire
+bounded file. Its non-empty line count MUST equal both root-report `rounds_run`
+and `len(report.rounds)`; when that count is zero, only an empty ledger is
+valid. Ledger `round_index` values MUST be unique, in file order, and exactly
+`1..N`. Each line MUST pair one-to-one with the report round at the same index,
+with exact type and value equality for every profile-allowlisted structural
+field, including `round_index`, `next_step`, and the object presence/type of
+the three role-status fields.
+
+Free text and role-status contents are never compared as evidence or
+persisted. If bounded parsing must inspect them to establish JSON shape, they
+remain in memory only until that line is validated and are then discarded. An
+empty ledger for a nonzero report, truncated prefix, extra suffix, duplicate,
+gap, reorder, or structural type/value mismatch makes the whole set `invalid`,
+not `partial` or `validated-terminal`.
+
+V1 does not read `events.jsonl`. Although the pinned records carry a schema
+version, the stream has no profile-level whole-file or unambiguous run-segment
+contract. Event support requires a separate approved versioned extension with
+its own complete event and sequence contract.
+
+### Authority and Lifecycle Consistency
+
+Within the profile, `logs/report.json` is the only authority for the native
+terminal-status and completion-satisfied qualification facts.
+`control/status.json` is a separate Supervisor process-lifecycle qualification
+fact. V1 MUST accept only a terminal status from the
+pinned Supervisor vocabulary: `completed`, `failed`, `cancelled`, `blocked`, or
+`incomplete`. A Supervisor `completed` status requires report `complete` plus
+`completion_satisfied: true`. `blocked` and `incomplete` require the
+corresponding report status. Supervisor `failed` or `cancelled` may override a
+different Manager report outcome because process failure or an operator action
+has higher lifecycle authority; the reader preserves both observations and
+MUST NOT relabel either as task success.
+
+This distinction follows the pinned Supervisor, where a clean process exit is
+not sufficient and a `complete` report without completion evidence becomes a
+protocol failure. The
+[terminal-status projection](https://github.com/AMAP-ML/LongHorizon-Harness/blob/be2e7b42523c4f35291f1ed57b683f6c03a29cdc/src/lh_harness/supervisor/service.py#L413-L458)
+is evidence for the distinction, not a schema imported into Better Harness.
+
+### Evidence States
+
+`evidenceState` is a closed profile-internal qualification value:
+
+| State | Meaning | Eligible for a separately approved future projection? |
+| --- | --- | --- |
+| `unavailable` | A required file is absent or cannot be read through the safe fixed path. | No |
+| `unsupported` | The explicitly selected profile or report schema version is outside V1. | No |
+| `invalid` | Present input is malformed, unsafe, identity-mismatched, internally inconsistent, or conflicts with another selected input. | No |
+| `partial` | The required terminal envelope is valid, but optional `rounds.jsonl` is absent. | Yes, for the internal required facts only, with a safe reason code. |
+| `validated-terminal` | The required envelope and present `rounds.jsonl` pass complete validation. | Yes, for the internal terminal and structural round facts. |
+
+No state means task success, and no state writes to a current Better Harness
+session/evidence owner. `partial` does not silently hide a malformed present
+ledger: malformed or ambiguous presence is `invalid`.
+
+### Failure and Decision Matrix
+
+| Condition | State | Future eligibility; current owner | Required decision |
+| --- | --- | --- | --- |
+| Required owner, status, or root report missing/unreadable | `unavailable` | Not eligible; none | Report a safe missing/unreadable code. |
+| Future report schema or non-V1 profile | `unsupported` | Not eligible; none | Require a separately reviewed profile revision. |
+| Present required file malformed or required field/type/status drifted | `invalid` | Not eligible; none | Fail closed; do not coerce or repair. |
+| Conditional report copy malformed or canonically different | `invalid` | Not eligible; none | Keep root authority but reject the artifact set. |
+| Any traversal, NUL, link, special file, or containment failure | `invalid` | Not eligible; none | Stop before parsing evidence. |
+| Directory, owner, or status run id/workspace mismatch | `invalid` | Not eligible; none | Stop before qualification; do not expose raw values. |
+| Valid required envelope with ledger absent | `partial` | Eligible internal required facts; none | Retain only a safe internal absent-ledger reason. |
+| Ledger present but empty for nonzero rounds, truncated, extra, duplicated, gapped, reordered, malformed, unsafe, or mismatched | `invalid` | Not eligible; none | Never downgrade malformed presence to absence. |
+| Status, report, or ledger qualification facts conflict | `invalid` | Not eligible; none | Retain no eligible result and use a safe conflict code. |
+| Required files and ledger valid for a non-success terminal status | `validated-terminal` | Eligible internal terminal/round facts; none | Preserve the internal native status; task success remains unobserved. |
+| Required files and ledger valid with report `complete` and completion satisfied | `validated-terminal` | Eligible internal source-reported completion facts; none | Do not upgrade source-reported completion to independent task success. |
+
+### Normalization Allowlist
+
+At the pinned Better Harness main, the existing `NormalizedToolActivityV1` and
+episode owners do not define LongHorizon terminal status, completion, round,
+workspace-match, evidence-state, abort-category, or task-success fields.
+ADR-0003 remains Proposed. Therefore V1 keeps every fact below inside
+qualification and projects none of them into a current provider-neutral owner.
+It also MUST NOT reuse the native HarnessRun file schema merely because the
+current `harness-run` analyzer exists.
+
+| Classification | Profile-internal fact | Current Better owner treatment |
+| --- | --- | --- |
+| Observed | Supervisor lifecycle status; root-report native terminal status and completion-satisfied; round count and native max-round value | Unobserved and not projected. |
+| Observed | Valid ledger round index and route labels | Unobserved and not projected. |
+| Derived | Opaque diagnostic alias, workspace-match result, opaque target reference, `evidenceState`, and closed safe reason | Unobserved and not projected. |
+| Derived | Closed abort/failure category: `none`, `user-cancelled`, `provider-failure`, `manager-blocked`, `round-limit`, `worker-failure`, or `other-redacted` | Unobserved and not projected; free-form text is discarded. |
+| Unobserved/excluded | Task, plan, output, Auditor report/result, state, contract, commands/receipts, error text, paths, model/provider credentials, tool payload, token/cost data | Never projected. |
+| Unobserved | Independent task success | Always unobserved; Manager done, Auditor text, file presence, or source-reported completion cannot change it. |
+
+This Draft defines only the qualification/validation boundary and future
+eligibility. Turning any internal fact into durable Better Harness evidence
+requires a separate maintainer-approved owner-extension or adapter spec that
+selects the canonical owner, schema, compatibility rules, and tests. That spec
+MUST preserve unobserved facts and MUST NOT create a parallel generic report
+schema.
+
+### Read and Privacy Policy
+
+V1 is read-only and fixed-allowlist. It MUST perform one bounded read per
+selected file, enforce fixed tested limits for total bytes, record count, line
+size, JSON depth, and string length, and reject unsafe or over-limit input. It
+MUST NOT repair, truncate-and-accept, rewrite, touch timestamps, create indexes,
+or modify any LongHorizon-Harness file. Exact numeric limits are an
+implementation-spec decision that MUST be frozen and tested before code lands;
+unbounded defaults are forbidden by this contract.
+
+Diagnostics MUST contain only a closed code, safe field name, evidence state,
+and opaque alias where needed. They MUST NOT contain a source path, raw run id,
+workspace, task, command, prompt, error/failure body, transcript, credential,
+or rejected input body.
+
+### Retry and Resume Boundary
+
+The pinned Supervisor `resume` creates a new run from saved task/config and
+writes `resumed_from` plus `resume_kind: retry`; its source comment explicitly
+says round state was not resumed
+([source](https://github.com/AMAP-ML/LongHorizon-Harness/blob/be2e7b42523c4f35291f1ed57b683f6c03a29cdc/src/lh_harness/supervisor/service.py#L2067-L2121)).
+V1 qualifies every selected run independently and does not consume or project
+that relation. It does not merge rounds, restore routing state, or call a retry
+a checkpoint resume. A future lineage feature requires separately validated
+source lineage and a reviewed profile revision.
+
+Post-merge PR #53 routing metadata remains an optional future extension. It
+requires separate versioned conformance and cannot widen V1.
+
+## Resolved Design Questions
+
+| Former question | V1 decision | Evidence or rationale |
+| --- | --- | --- |
+| Producer/report versions | Specify V1 from pinned commit `be2e7b42523c4f35291f1ed57b683f6c03a29cdc`, then qualify only explicit profile selection, report schema 2, and exact conformance shape. | No producer-version field exists in the pinned report, so runtime cannot verify or claim a broader release range. |
+| Conflicting local reports | Root `logs/report.json` is authority; a present duplicate must be canonical-JSON identical or the set is `invalid`. | This closes ambiguity without choosing by time or convenience. |
+| Stable event/round fields | Validate the ledger as an exact, complete, one-to-one structural companion to report rounds; exclude events entirely. | Ledger has no version, while the event stream lacks a profile-level whole-file/run-segmentation contract. |
+| Per-round file intake | No per-round file is an input to V1. | The files are unversioned and contain sensitive natural-language/debug content. |
+| Evidence available versus task success | `partial` and `validated-terminal` establish only future projection eligibility; current owners receive nothing and task success stays unobserved. | Artifact validity and source-reported completion are not independent success proof. |
+| Run identity and retry lineage | Qualify with private source run id/workspace, retain at most an internal opaque diagnostic alias, and treat every run independently. | Pinned resume is a new retry run, not restored round state; V1 does not join its optional relation. |
+
+## Conformance Scenarios
+
+These scenarios are the closed V1 conformance set for a future implementation.
+They are design requirements, not fixtures or tests implemented by this Draft.
+"Projection: none" means every profile-internal fact remains unobserved in the
+current Better Harness owners.
+
+| ID | Input | Expected qualification state | Current Better projection | Security or honesty signal | Future implementation evidence |
+| --- | --- | --- | --- | --- | --- |
+| C-1 | Valid matching owner/status and schema-2 root report; ledger absent; conditional duplicate absent or matching | `partial` | None | Safe `ledger-absent` internal reason; root remains authority | Sanitized required-envelope fixture with both duplicate-presence variants |
+| C-2 | `rounds_run: 0`, empty report rounds, present empty ledger, and matching duplicate when present | `validated-terminal` | None | Empty means zero records, not one blank/malformed record | Zero-round golden fixture and exact empty-file assertion |
+| C-3 | `rounds_run: N > 0` with exactly N ledger lines, indexes `1..N`, and one-to-one structural matches | `validated-terminal` | None | Only structural fields survive qualification in memory | Multi-round golden fixture with exact count/order/value assertions |
+| C-4 | Required owner, status, or root report absent or unreadable through the fixed safe path | `unavailable` | None | Closed missing/unreadable code without source values | One fixture per required file and read-failure class |
+| C-5 | Non-V1 explicit profile or report schema other than integer 2 | `unsupported` | None | No best-effort downgrade or version coercion | Future-schema and wrong-profile negative fixtures |
+| C-6 | Present required JSON malformed; required type or terminal enum invalid; run id or workspace binding mismatched | `invalid` | None | Closed invalid/type/identity code; no raw value | Table-driven malformed, type, enum, identity, and workspace fixtures |
+| C-7 | Conditional duplicate absent, matching, malformed, or canonically different | Absent/matching follows C-1..C-3; malformed/different is `invalid` | None | Never select a copy by timestamp or convenience | Canonical-key-order positive fixture plus malformed/value/type mismatch negatives |
+| C-8 | Present ledger is a truncated prefix, extra suffix, empty for N>0, duplicate, gap, reorder, malformed record, or allowlisted structural mismatch | `invalid` | None | Present corruption never degrades to optional absence | One mutation fixture for every listed ledger failure |
+| C-9 | All required invariants pass and `rounds_run > max_rounds` | C-1 or C-3 according to ledger presence | None | Native max is observed internally but does not reject human-extended runs | Positive extended-budget fixture proving this fact alone is not invalid |
+| C-10 | Selected path has a symlink/junction, traversal, special file, canonical escape, NUL, or fixed bound exceeded | `invalid` (missing/unreadable required input remains C-4) | None | Fail closed with only safe code/field/opaque alias | Cross-platform path attacks and over-limit fixtures after numeric limits are frozen |
+| C-11 | Excluded events, per-round text/state/output/report, trajectories, screenshots, remote mirrors, receipts, task file, or worker log exist | State is determined only by required inputs and ledger | None | Excluded files are not opened and cannot change qualification | Read-spy fixture proving zero opens for every excluded path |
+| C-12 | Any pinned native report terminal status; `complete` paired with either valid or invalid completion-satisfied value | Valid enum follows C-1..C-3; inconsistent completion pair is `invalid` | None | Status/completion remain internal; task success stays unobserved | Enum matrix plus complete/boolean cross-field negatives |
+| C-13 | Owner/report contain raw task, command, error/free text, paths, run id, provider/model, or credential-shaped values | State follows structural rules | None | Durable result and diagnostic contain none of the raw values | Canary-secret and private-path non-retention assertions |
+| C-14 | A direct CLI/shared log directory is selected | `unsupported` as outside `longhorizon-supervisor-run-v1` | None | No fuzzy adoption of a directory missing the Supervisor envelope | Direct-CLI/shared-log rejection fixture |
+| C-15 | A Supervisor retry/resume produces a new run with `resumed_from` and `resume_kind: retry` | Each selected run is a separate qualification instance | None | No round merge, lineage projection, or checkpoint claim | Two-run fixture proving independent aliases/state and zero joins |
+| C-16 | PR #53 routing fields are absent, or appear only as otherwise unconsumed extra fields in a conforming set | Base qualification does not require them; extras grant no capability | None | No routing projection or support upgrade | Base fixture without fields and inert-extra-field fixture with duplicate consistency preserved |
+| C-17 | Qualification reaches `partial` or `validated-terminal` while no approved Better owner extension exists | Internal state is retained only for the qualification call | None | No session evidence, durable report field, or support declaration | Negative contract test over every current owner serialization surface |
+
+Any future implementation MUST realize this table with sanitized fixtures and
+must not weaken a negative row by treating it as a warning-only success.
+
+## Acceptance Scenarios
+
+- **AC-1 (traceability):** The final proposal names Story #78, both pinned
+  commits, and source permalinks for every required artifact claim. Unmerged
+  PR #53 fields are absent from the required profile.
+- **AC-2 (ownership):** A review can assign execution, role behavior, state
+  mutation, lifecycle, and artifact production to LongHorizon-Harness, while
+  assigning only read-only qualification, validation, normalization, and report
+  projection to Better Harness. No path writes back to LongHorizon-Harness.
+- **AC-3 (source profile):** The normative contract identifies the local
+  terminal report separately from the round ledger, event stream, per-round
+  files, and best-effort remote mirror, with explicit required, optional, and
+  excluded inputs.
+- **AC-4 (semantic honesty):** Tests and documentation keep terminal status,
+  artifact availability, Auditor status, `completion_satisfied`, and task
+  success as separate facts. Natural-language Auditor output alone cannot
+  produce a structured success claim.
+- **AC-5 (normalization):** A future source profile emits a bounded,
+  profile-internal qualification result only. Every LongHorizon-specific fact
+  remains unobserved and unprojected in current Better owners; a separate
+  approved owner-extension/adapter spec is required before durable evidence,
+  instead of borrowing the native HarnessRun schema or fabricating fields.
+- **AC-6 (trust and privacy):** Validation fails closed on unsupported versions,
+  malformed or oversized records, unsafe paths, and disallowed sensitive
+  fields. The projected evidence contains no raw prompt, task, transcript,
+  credential, trajectory, or absolute home path.
+- **AC-7 (lifecycle and retry):** Related LongHorizon-Harness attempts remain
+  distinct candidate artifact sets. Better Harness does not consume
+  `resumed_from`, label a retry as checkpoint resume, or reconstruct mutable
+  round, routing, or lifecycle state.
+- **AC-8 (future conformance):** A later implementation supplies pinned,
+  sanitized conformance fixtures for each accepted schema version and proves
+  missing, partial, malformed, and future-version inputs fail with bounded
+  diagnostics. Optional post-merge routing metadata is tested separately from
+  the base profile.
+- **AC-9 (product honesty):** README support tables, adapter matrices, host
+  catalogs, CLI help, and package metadata remain unchanged until an adapter is
+  implemented and its conformance and privacy checks pass. This Draft never
+  appears as a support declaration.
+- **AC-10 (validation):** Review evidence maps every normative claim to a pinned
+  source, every acceptance scenario to an executable or documentary check, and
+  verifies links, repository status, privacy scans, and the final diff before
+  acceptance.
+
+## Risks and Controls
+
+| Risk | Fail-closed control | Review signal |
+| --- | --- | --- |
+| Schema 2 drifts without changing its version | Exact required types, enums, and invariants; same-version drift is `invalid` | Mutation fixtures for every required field and invariant |
+| No producer-version field exists | Scope V1 only to pinned evidence plus exact conformance; claim no release range | Pinned commit remains explicit in Traceability and version tests |
+| The two local reports conflict | Root report is authority, but any present duplicate must be canonical-JSON identical | Canonical-order positive and value/type mismatch negative fixtures |
+| Ledger is truncated or append state is ambiguous | Whole-file exact count, `1..N` ordering, and one-to-one structural comparison | Prefix/suffix/empty/duplicate/gap/reorder mutation fixtures |
+| Direct CLI or shared logs are mistaken for a supervised run | Explicit one-directory selection plus required owner/status envelope; no scanning or guessing | Direct-CLI/shared-log rejection fixture |
+| Filesystem paths escape or change type | Canonical containment, no-follow component walk, regular-file requirement, and fixed bounds | Windows/macOS/Linux link, traversal, special-file, and over-limit cases |
+| Owner/report fields expose private content | Bounded internal parse, fixed allowlist, immediate discard, safe diagnostics, and zero current projection | Canary task/path/run-id/provider/model/credential non-retention checks |
+| Native completion is presented as independent success | Separate lifecycle/report facts; task success always unobserved | Every terminal-status fixture asserts zero success projection |
+| Existing Better owners cannot represent these facts honestly | Current projection is none; owner extension requires separate maintainer-approved spec | Negative serialization tests over `NormalizedToolActivityV1` and episode owners |
+| Unmerged PR #53 changes optional source fields | Ignore unconsumed extras, require duplicate consistency, and grant no capability | Base-without-routing and inert-extra-field conformance fixtures |
+| Numeric safety bounds are not yet frozen | This Draft forbids unbounded reads; implementation is blocked until a follow-up spec fixes tested values | Maintainer-approved numeric table and boundary tests before code |
+
+## Non-goals
+
+- Add an adapter, runtime behavior, test fixtures, or production code.
+- Start, stop, control, resume, retry, or otherwise modify a
+  LongHorizon-Harness run.
+- Change README support declarations, the adapter matrix, host catalog, CLI,
+  package metadata, release metadata, or generated documentation.
+- Claim that Better Harness currently supports LongHorizon-Harness.
+- Design Story #63's generic completion model or define task success for every
+  provider.
+- Copy raw prompts, tasks, transcripts, trajectories, credentials, secrets, or
+  absolute home paths into Better Harness evidence.
+- Depend on PR #53 routing fields or treat them as part of the pinned base
+  contract.
+- Describe a retry as checkpoint resume or claim that Better Harness can
+  rehydrate LongHorizon-Harness routing or lifecycle state.
+- Define a new generic report schema or copy the upstream schema into this
+  repository.
+
+## Plan and Tasks
+
+1. **Gate 1 - inventory and ownership:** Pin both repositories, record the
+   observed LongHorizon-Harness artifacts and gaps, define the one-way owner
+   boundary, and freeze the acceptance-scenario skeleton in this Draft.
+2. **Gate 2 - normative contract:** Close the six inventory questions with the
+   narrow V1 decisions above; define source qualification, version handling,
+   normalization, semantic labels, trust/privacy limits, and lifecycle
+   treatment without implementation or support claims.
+3. **Gate 3 - conformance and validation:** Specify sanitized conformance cases,
+   map acceptance scenarios to checks, run documentation validation and final
+   independent review, and decide whether the proposal is ready to authorize a
+   separately scoped implementation.
+
+## Forward Path and Rollback
+
+This Draft is a maintainer discussion artifact only. It does not authorize an
+implementation, adapter, runtime integration, evidence owner extension, or
+support declaration.
+
+If maintainers accept the direction, the next change MUST be a separate
+owner-extension/adapter spec. That spec must freeze numeric bounds, select the
+canonical Better Harness owner and schema, define sanitized fixtures, require
+cross-platform filesystem tests, and set an explicit support-claim gate before
+implementation starts.
+
+If maintainers reject the direction or pinned upstream facts drift, rollback is
+limited to deleting or revising this one spec. There is no runtime, stored
+Better evidence, migration, or cleanup obligation.
+
+This single-file diff does not overlap PR #94 at the file level. It changes no
+LongHorizon-Harness file, so it does not overlap PR #53 at the file level;
+conceptually, PR #53 remains only a possible future optional extension. These
+observations do not guarantee GitHub mergeability or remove the need to review
+concurrent changes.
+
+## Test and Review Evidence
+
+This Draft adds no executable conformance fixtures. The future-test column is
+an implementation gate, not a claim that those tests exist in this change.
+
+| AC | Current document-review evidence | Future executable evidence | This Draft's validation |
+| --- | --- | --- | --- |
+| AC-1 | Traceability and Pinned Evidence name Story #78, both immutable commits, and claim-level source links; PR #53 is optional only. | Link and pinned-fixture checks remain required in any follow-up. | Documentation link graph plus manual pinned-link/source review. |
+| AC-2 | Ownership Boundary and Read and Privacy Policy make the bridge one-way and prohibit control or mutation. | Read-spy and unchanged-files tests around every qualification state. | Single-file diff review confirms no LongHorizon or runtime file changed. |
+| AC-3 | Selection Boundary and Qualification Set define required, conditional, optional-ledger, and excluded inputs; C-1..C-11 cover their decisions. | Sanitized qualification fixtures for every input class and state transition. | Cross-section review of inventory, input table, states, and decision matrix. |
+| AC-4 | Authority and Lifecycle Consistency separates Supervisor lifecycle, native report status, completion, evidence state, and unobserved task success. | Terminal-enum/completion matrix and conflict mutation fixtures. | Pinned Manager/Supervisor source review plus semantic-label sweep. |
+| AC-5 | Normalization Allowlist records the negative proof: pinned `NormalizedToolActivityV1` and episode owners have none of these fields, so current projection is always none. | Negative serialization tests across every current owner; a separately approved owner-extension spec before any positive projection test. | Pinned Better owner/analyzer source inspection and stale-projection-claim scan. |
+| AC-6 | Identity, Workspace, and Path Safety plus Read and Privacy Policy define fail-closed paths, bounded reads, discard, and safe diagnostics. | Cross-platform filesystem attacks, numeric boundaries after freeze, and canary secret/path non-retention tests. | Spec privacy scan for local paths, key shapes, private-key markers, and raw-transcript claims. |
+| AC-7 | Retry and Resume Boundary treats each run as a distinct qualification instance and forbids lineage joins or checkpoint claims. | Two-run retry fixture proving independent qualification and zero state merge. | Pinned `_resume_once` review and lifecycle terminology sweep. |
+| AC-8 | Version and Shape Contract, Failure and Decision Matrix, and C-1..C-17 define the future conformance suite. | Every scenario row becomes a sanitized positive or negative fixture; no warning-only weakening. | Scenario-to-contract review and marker/stale-claim scans. |
+| AC-9 | Product honesty remains gated; Forward Path requires a separate approved spec before implementation or support. | Support-surface snapshot must change only in the later support-gated contribution. | Status/name-only checks show only this new spec; README, matrix, catalog, CLI, package, and ADR remain unchanged. |
+| AC-10 | Plan, scenarios, risks, future tests, and this matrix complete the Story-to-review chain. | Follow-up readiness must run its executable conformance, package, and cross-platform suites. | Generator hash/diff check, focused link test, whitespace checks, privacy scans, and independent global review. |
+
+Actual Draft-validation commands:
+
+```text
+node scripts/doc-link-graph/cli.mjs skills/better-harness
+npx vitest run test/skills-docs/doc-link-graph.test.mjs
+git diff --check -- docs/specs/2026-08-19-78-longhorizon-run-evidence-boundary.md
+```
+
+The direct `npx` command cannot start in this independent worktree because its
+local `node_modules` does not contain Vitest. No dependency is installed or
+upgraded. The same repository's existing Vitest binary and config are therefore
+also run with `--root` fixed to this worktree; that is the focused executable
+result reported for this Draft.
+
+Review Readiness requires: Draft status; Story and pinned-source traceability;
+stable AC and conformance IDs; explicit non-goals; risks tied to controls and
+future tests; no implementation/support claim; one-file status/diff; and no
+commit, push, pull request, or issue comment. Until the independent global
+review passes, this document remains a Draft proposal and does not authorize
+implementation.