diff --git a/MANIFEST.md b/MANIFEST.md
index a9926034..3665bcb0 100644
--- a/MANIFEST.md
+++ b/MANIFEST.md
@@ -73,8 +73,8 @@ hand-written.
 
 <!-- Generated by `bun run emit:manifest`; do not hand-edit. -->
 
-There are 59 records, ids `0001`-`0059`, alongside
-`0000-template.md`: 50 accepted, 7 proposed, 2 superseded.
+There are 60 records, ids `0001`-`0060`, alongside
+`0000-template.md`: 50 accepted, 8 proposed, 2 superseded.
 
 | Id | Status | Title |
 | --- | --- | --- |
@@ -137,5 +137,6 @@ There are 59 records, ids `0001`-`0059`, alongside
 | `0057` | accepted | Preview the decision portal locally through a preview selector in the published image, published to loopback only |
 | `0058` | accepted | Project the decision corpus into a GitHub Project as a one-way derived view |
 | `0059` | accepted | Aggregate decision corpora across an organization's repositories as a pinned, read-only federation |
+| `0060` | proposed | Drive the probabilistic evaluator passes through the harness with typed questions over a versioned wire contract |
 
 <!-- END GENERATED: adr-inventory -->
diff --git a/docs/adr/0060-drive-the-probabilistic-evaluator-passes-through-the-harness-with-typed-question.md b/docs/adr/0060-drive-the-probabilistic-evaluator-passes-through-the-harness-with-typed-question.md
new file mode 100644
index 00000000..44d428f0
--- /dev/null
+++ b/docs/adr/0060-drive-the-probabilistic-evaluator-passes-through-the-harness-with-typed-question.md
@@ -0,0 +1,469 @@
+---
+schemaVersion: 0.2.0
+id: "0060"
+title: "Drive the probabilistic evaluator passes through the harness with typed questions over a versioned wire contract"
+status: proposed
+date: 2026-10-11
+deciders:
+  - "@mbeacom"
+tags:
+  - evaluator
+  - architecture
+  - ai
+  - calibration
+  - governance
+scope: org
+reversibility: two-way-door
+blastRadius: org
+relatesTo:
+  - "0003"
+  - "0004"
+  - "0007"
+  - "0009"
+  - "0010"
+  - "0014"
+  - "0016"
+  - "0027"
+  - "0031"
+  - "0037"
+affects:
+  - type: path
+    pattern: "packages/evaluator/**"
+  - type: path
+    pattern: "packages/cli/src/evaluate*.ts"
+  - type: path
+    pattern: "specs/011-probabilistic-evaluator-passes/**"
+  - type: path
+    pattern: "specs/012-evaluator-calibration/**"
+    note: "012 owns the snapshot shapes this contract parses into; it is constrained, not changed."
+provenance:
+  authoredBy: agent-drafted
+review:
+  tier: arb
+  tierReason: >-
+    Fixes the only path by which model output can reach the evaluator, which
+    ADR-0027 governs as the gate deciding where human judgment is required.
+---
+
+# ADR-0060: Drive the probabilistic evaluator passes through the harness with typed questions over a versioned wire contract
+
+> **Status: proposed.** Agent-drafted. This record is gate T003 of
+> [`specs/011-probabilistic-evaluator-passes/`](../../specs/011-probabilistic-evaluator-passes/tasks.md).
+> It **ratifies** the harness-driven architecture that spec already selected,
+> and it **resolves** that spec's [Q6](../../specs/011-probabilistic-evaluator-passes/spec.md#q6),
+> the wire contract. It does **not** amend the constitution, and it ships no
+> code. It clears one of 011's four Phase 0 gates. T001 and T002 (012's frozen
+> holdout and its observed-failing gate) and T004 (Q4 and Q5) stay open, so no
+> probabilistic pass is authorized by this record. Under ADR-0027 §2 nothing
+> here describes the evaluator as having more than its deterministic pass; every
+> statement about Passes 1–3 is conditional.
+
+## Context
+
+[ADR-0027](./0027-ratify-the-deterministic-evaluator-and-bind-calibration-reporting-to-the-first-probabilistic-pass.md)
+carries Passes 1–3 forward as intent and forbids escalation by model
+discretion. Spec 011 turns that into an architecture: **the model produces
+evidence and deterministic code does the deciding**, and the model call
+happens in the agent harness that invoked adrkit, never inside adrkit (011
+FR-003, FR-004). 011 makes that choice and leaves two things to a record: the
+choice itself must be ratified, and the bundle/response wire contract (Q6)
+must be settled rather than deferred, because Passes 2 and 3 cannot be built
+without it.
+
+What exists, measured in this tree on 2026-10-11:
+
+- **adrkit has never been a model client.** No package calls a model. Three
+  landed gates enforce that for the evaluator and must stay green unmodified:
+  `packages/evaluator/test/purity.test.ts` traps `globalThis.fetch`,
+  `Date.now`, and `Math.random`; `packages/evaluator/test/contracts.test.ts`
+  bans `node:fs`, `node:net`, `node:http`, `node:child_process`, and every
+  `@adrkit/adapter*` import under `packages/evaluator/src/`; and feature 005's
+  SC-006 bars model, prompt, embedding, and retrieval libraries.
+- **The constitution forbids the alternative without an amendment.**
+  Principle II: "Network-dependent tests and runtime behavior are forbidden".
+  Its rationale names "offline use, IP provenance, and the mechanical boundary
+  between this Apache-2.0 project and any employer-internal IP".
+  [ADR-0003](./0003-ship-as-spec-kit-extension.md) already ships adrkit inside
+  a harness that holds model access, and
+  [ADR-0010](./0010-bun-toolchain.md) makes published artifacts Node programs
+  that need no network.
+- **There is a precedent for a strict structured input.**
+  `packages/cli/src/evaluate-snapshot.ts` validates the
+  `adrkit.pass0.snapshot/v1` bundle: a versioned envelope token, a
+  duplicate-key-aware JSON reader, `rejectUnknownKeys` on every object, and
+  `SnapshotContractError` mapped to exit `2`. Malformed data that is present is
+  an error, and the CLI never invents defaults for it. Absent optional data
+  degrades to inert ([ADR-0009](./0009-affects-resolution-and-catalog-binding.md)).
+- **The spec already fixes most of what the wire must carry.** FR-010 requires
+  each score above 0 to cite a span of the proposal, and the citation to be
+  *validated*, not merely present. FR-011 makes two cap antecedents model
+  evidence. FR-014 requires Pass 3 to be a separate call with four outputs,
+  each present or explicitly absent, and never invented. FR-016 sets
+  `low-confidence` to citation coverage over a constant 8 and forbids a model's
+  self-reported confidence "alone or as a term". FR-017 requires each Pass 3
+  output to carry `bearsOn` over D1–D8. FR-005 makes the snapshot the
+  determinism boundary. Feature 012 owns the `RubricScoreSnapshot` and
+  `AdversarialSnapshot` field shapes (012 T008) and every metric.
+- **No relevance primitive exists** (011 Q3). `packages/mcp/src/search/normalize.ts`
+  says "No stemming, fuzzy, weighting, or ranking". Feature 012 has withdrawn
+  its relevance floor, so `novel-no-precedent` is expected to stay
+  `evidence-absent`.
+- **A class of model now exists that answers typed questions without
+  generating text.** On 2026-10-10 the maintainer asked for an evaluation of
+  TypeSafe's Jev and Microsoft's Microsoft-Decision-1. Both take a `state` plus
+  named questions of three types: yes/no (`noul`), pick-one (`choice`, at most
+  255 options), and ordered levels (`score`, 2 to 10 levels). Both return
+  probabilities, not prose. Microsoft serves the same request shape at
+  `…/providers/microsoft/v1/systemone`, the path TypeSafe uses at
+  `api.typesafe.ai/v1/systemone`. Neither vendor's documentation mentions
+  published weights or self-hosting; both are served only through hosted
+  endpoints. Both charge $0.042 per million input tokens. Their own
+  documentation sets the limits that matter here: neither generates text or a
+  rationale; both say a fractional score is not a calibrated rating
+  (TypeSafe's Jev 1.13 limitations page: "score levels poorly calibrated";
+  Microsoft: "Prefer scores for relative ordering and thresholds rather than as
+  absolute, calibrated ratings"); TypeSafe says "injected or misleading content
+  in the state can move its answers" and that the model leans toward the first
+  option; Microsoft says wording and option order change scores. Neither vendor
+  claim is evidence for this project. ADR-0027 requires adrkit's own figures.
+
+The evaluation showed that the wire contract's real choice is between two
+shapes. One is a free-form prompt that a generative model answers with JSON and
+quoted citations. The other is a set of typed questions whose answers adrkit can
+validate exactly. Only generative models can answer the first. Both kinds of
+responder can answer the typed parts of the second. The second also makes
+several of 011's requirements mechanical instead of hopeful.
+
+## Decision
+
+**We will run every probabilistic pass through the harness, over a versioned
+wire contract of typed questions. adrkit authors every question and validates
+every answer. The record text travels only as data.**
+
+### 1. The architecture is ratified, and the constitution is not amended
+
+adrkit emits a **bundle** and consumes a **response**. The harness that
+invoked adrkit makes the model call. **No adrkit package calls a model or holds
+model credentials, and that includes `packages/adapters/*`** (011 FR-004). A
+relay that forwards a bundle to a hosted endpoint is a harness and lives outside
+this repository. Because adrkit opens no model connection, Principles II and IV
+hold as written, and this record ratifies an architecture rather than amending
+one.
+
+The three landed purity and contract gates named in the Context stay green and
+unmodified. No probabilistic pass may be a precondition for any adrkit command
+completing (011 FR-004, US5).
+
+### 2. Two envelopes, versioned on their own
+
+- The request is `adrkit.evaluator.bundle/v1`. The response is
+  `adrkit.evaluator.response/v1`. These tokens are independent of
+  `schemaVersion` and of the CLI's version. Any change that alters how a valid
+  document is read takes a new token and a record.
+- **One bundle covers one pass of one proposal.** Pass 1, Pass 2, and Pass 3
+  each get their own bundle. A Pass 3 bundle never contains Pass 2 output
+  (FR-014).
+- Every bundle carries `bundleId`, the lowercase hex SHA-256 of its canonical
+  bytes with `bundleId` itself omitted. Canonical means keys sorted by
+  `compareCodeUnits`, no insignificant whitespace, and UTF-8. A response must
+  echo the `bundleId`. A mismatch is a contract error, so a response cannot be
+  replayed against a different proposal or pass.
+- Documents travel as JSON files or over stdio. The CLI flags are 011's to name
+  in implementation, and they are additive under
+  [ADR-0031](./0031-publish-a-narrow-consumer-sdk-as-the-contract-and-document-the-cli-json-as-its-s.md).
+
+### 3. What a bundle contains
+
+- **`state`**: the proposal and, for Pass 1 and Pass 2, the context 011 FR-006
+  assembles. It is data. Record text appears **only** here.
+- **`spans`**: the proposal body cut deterministically before any call (clause
+  4).
+- **`questions`**: a map from question id to a typed question. Ids, types,
+  `instructions`, and option or level descriptions come from **a committed
+  question catalog** in `@adrkit/evaluator`, keyed by id and versioned with the
+  envelope. They are constants. No question text is interpolated from record
+  content. The only per-proposal variation allowed is *which* catalog questions
+  appear, and the span and candidate ids they range over.
+
+That is the injection boundary. Hostile text in a record can try to move an
+answer, as both vendors warn, but it cannot rewrite a question, add one, or
+widen an answer's domain, because the validator checks every answer against the
+catalog and not against anything in `state`.
+
+Five question types:
+
+| Type | Answer `value` | Optional evidence | Read by the kernel |
+|---|---|---|---|
+| `yes-no` | `true` / `false` | `probability` in [0, 1] | `value` |
+| `one-of` | one declared option key | `probabilities` over every option | `value` |
+| `level` | integer index into the declared levels | `probabilities` over every level | `value` |
+| `cite` | list of span ids from this bundle | none | `value` |
+| `text` | string, at most 2,000 code points | none | presence only |
+
+When optional probabilities are supplied, every one must be a finite number in
+[0, 1], and they must agree with `value`. For `yes-no`, `value` must equal
+`probability >= 0.5`. For `one-of` and `level`, there must be exactly one
+probability per declared option or level, keyed by it, they must sum to 1
+within 1e-6, and `value` must be the argmax, with ties broken by declared
+order. A vector such as `[-1, 2]` sums to 1 and is still refused. Any
+violation is a contract error. Any threshold other than 0.5 is a calibration
+decision and belongs to 012, by record. Probabilities are recorded as separate
+evidence, beside the scores and not inside them. 012's FR-019 drift metric is
+defined over raw and post-drop surviving scores and does not read them; any
+metric that does needs 012 to define it first. **They are never a routing
+input.** FR-016 already bars a model's
+confidence from `low-confidence`, and this contract gives no other trigger a
+path to it. A fractional score is never accepted: `level` is an integer,
+because both vendors disclaim the interpolated value.
+
+`bearsOn` (FR-017) is not a set type. It is one `yes-no` question per dimension
+per Pass 3 output. That keeps the contract to primitives every responder can
+answer, and it matches how decision models evaluate questions in isolation.
+
+### 4. Citations are span ids, so validation is membership
+
+The proposal body is the text after the closing `---` of its frontmatter,
+normalized to `\n` line endings and taken as UTF-8. It is cut into **blocks**
+separated by one or more blank lines. A fenced code block (```` ``` ```` or
+`~~~`, matching ADR-0023's fence rule) is one block however many blank lines it
+contains. Blocks are numbered `s1` … `sN` in document order. Each span carries
+`id`, `start`, and `end` as byte offsets into the normalized body (half-open),
+and its `text`. The cut is a pure function with no clock or filesystem, and it
+lives in `@adrkit/evaluator`.
+
+A `cite` answer is a list of ids. The validator refuses an id that is not in
+this bundle's `spans`, a duplicate, or a non-list. A refused citation is a
+contract error. An **empty** list is valid and means *uncited*. That makes
+FR-010's "validated, not merely present" a set-membership check. A model
+cannot cite a span that quotes nothing, because it cannot name one that adrkit
+did not issue. Whether a cited span actually supports the score remains the
+model's evidence and 012's to measure. This record does not claim it.
+
+Citations live in the report and are discarded with it. That is the default
+answer to 011 [Q4](../../specs/011-probabilistic-evaluator-passes/spec.md#q4),
+stated here so nobody assumes an audit trail that does not exist. Keeping
+citations inside a record is a schema change, which makes it a one-way door
+under Principle V. It needs its own record, and the maintainer still has to
+rule on Q4.
+
+### 5. Absent, refused, and malformed are three different outcomes
+
+| Outcome | Example | Result |
+|---|---|---|
+| **Malformed** | Wrong envelope token, unknown key, duplicate key, wrong type, value outside the declared domain, a probability outside [0, 1] or missing for a declared option, probabilities that disagree with `value`, an omitted Pass 3 output, unknown span id, `bundleId` mismatch | Typed contract error at the boundary. Exit `2`, as for `SnapshotContractError`. Nothing from the response is used. **Never `PassAbsence`.** |
+| **Refused** | `{ "refused": true }` with no answers, or no response at all | `PassAbsence` for that pass, with the reason. Evidence was never produced. |
+| **Unanswered** | A question id missing from `answers`, other than a Pass 3 output | That item is `evidence-absent`. Every other answer is used. |
+
+The one exception is the four Pass 3 outputs. FR-014 requires each to be
+present or **explicitly** absent, so an omitted Pass 3 output answer is
+malformed, not unanswered: the explicit absence is an empty-string `value`
+(clause 8). Omission cannot stand in for it, or a harness that dropped a slot
+would read as one that judged it empty.
+
+A malformed response is a broken integration, not missing evidence. Counting it
+as absence would drain 012's marginal denominator silently, which is the
+failure 011 calls "absence reported as evidence". Every object rejects unknown
+keys, as `rejectUnknownKeys` does today. The reader rejects duplicate JSON keys,
+as the snapshot reader does.
+
+### 6. The response declares its responder, and adrkit records it as declared
+
+Each response carries `responder: { kind, model }`. `kind` is one of
+`decision-model`, `generative-model`, or `human`. `model` is required unless
+`kind` is `human`, and it must be the **resolved, versioned** model id the
+provider returned. Both System One APIs return one in their `model` field, for
+example `jev-1.13.0` rather than `jev-latest`. adrkit cannot verify either
+field. It records them verbatim, labelled as declared, because 012's
+per-dimension drift is measured per model version (012 FR-019) and is
+meaningless across an alias that moved underneath it.
+
+### 7. Obligations on the harness that adrkit cannot check
+
+The harness contract states them, because adrkit can only state them:
+
+- **Fresh context per bundle.** No state carries over between a proposal's
+  Pass 2 and Pass 3 bundles, or between proposals. adrkit can guarantee only
+  that it did not put Pass 2 output into a Pass 3 bundle (FR-014). Decision
+  models evaluate each question in isolation by construction. That supports the
+  claim, but the claim still rests on the harness.
+- **Questions are asked as written.** The harness must not rewrite `instructions`
+  or options. It may translate them mechanically, for example `one-of` to a
+  vendor's `choice`, a `cite` to one yes/no or pick-one per span, or `level` to
+  `score` with the argmax taken.
+- **Option order is the harness's to vary.** Both vendors report order effects.
+  A harness may shuffle options when it asks, but it must answer with the
+  declared keys.
+- **Data egress is the operator's decision.** A bundle carries the full
+  proposal and the text of the records retrieved for it. Where that goes is
+  decided by whoever runs the harness. adrkit holds no credential and makes no
+  call.
+
+### 8. What a decision-only responder can and cannot supply
+
+This is a stated limit of that class of responder, not a defect:
+
+| Need | Typed shape | Decision model |
+|---|---|---|
+| Pass 1 relevance of a category (a) candidate (not emitted; clause 9) | `yes-no` per candidate | Yes |
+| Pass 2 level per dimension | `level` (5 levels, the 0–4 anchors) | Yes |
+| Pass 2 citations | `cite` | Yes, by per-span questions |
+| FR-011 cap antecedents (D2 shared assumption, D3 negative consequence) | `yes-no` | Yes |
+| Pass 2 "name the missing thing" for a score below 3 | `text` | **No** |
+| Pass 3 outputs and their presence | `text`, with presence meaning non-empty `value` | **No** |
+| Pass 3 `bearsOn` | `yes-no` per dimension | Only alongside a present output |
+
+A Pass 3 slot is present **only** when its `text` answer is non-empty, and
+explicitly absent when it is the empty string. Every Pass 3 output must be
+answered one way or the other (clause 5), so a slot can never be filled with
+invented content or dropped silently (FR-014). A relay for a responder that
+cannot generate answers a Pass 3 bundle with `refused`, which produces a
+Pass 3 `PassAbsence`, and
+`pass-disagreement` records `evidence-absent`. What an unanswered
+"missing thing" does to a Pass 2 score is 011 and 012's to define. This record
+carries the question and its absence, nothing more.
+
+### 9. Relevance has a shape and a candidate owner, and is not emitted yet
+
+This gives 011 [Q3](../../specs/011-probabilistic-evaluator-passes/spec.md#q3)
+a candidate owner: a relevance judgment for retrieval category (a) can be
+harness-supplied evidence, one `yes-no` per candidate, and adrkit would still
+implement no ranker. The question catalog defines that question's shape.
+
+**Under this record, no Pass 1 bundle emits it.** Admitting a candidate on a
+`true` answer would make the 0.5 rule in clause 3 a ranking function with a
+floor, which is the floor 012 withdrew. Recording answers that admit nothing
+would still be observed ranking output under FR-018, and would move
+`novel-no-precedent`'s absence class without anyone deciding it should. So
+category (a) stays empty, nothing ranks, and `novel-no-precedent` stays
+`evidence-absent` with its class derived from observed state, until 012
+decides by its own record whether harness relevance counts as a ranking
+function and with what floor. Categories (b) and (c) stay rule-based and
+unconditional either way (FR-006).
+
+### 10. Where the code goes
+
+- The span cutter, the question catalog, and the response validator will be
+  pure modules in `@adrkit/evaluator` (under `src/harness/`). They will take an
+  already-parsed JSON value and return either a snapshot in 012's shapes or a
+  typed contract error, and will import nothing the contract test bans.
+- The duplicate-key-aware reader and the file and stdio I/O will stay at the
+  CLI boundary, next to `evaluate-snapshot.ts`.
+- Response parsing must be deterministic: the same response bytes give the same
+  snapshot bytes. Responses themselves are not deterministic, which is why the
+  snapshot is the determinism boundary (FR-005).
+
+## Options considered
+
+### Option A: Typed questions over a versioned envelope, harness-driven (chosen)
+
+| Dimension | Assessment |
+|---|---|
+| Constitution | Unchanged. No adrkit package makes a model call. |
+| Injection surface | Questions are catalog constants. Record text is only data. |
+| FR-010 citation validation | Exact, by span-id membership |
+| Responders | Generative models, decision models, and humans |
+| Cost | A catalog to maintain, and a span cutter whose output the contract depends on |
+
+### Option B: A free-form prompt bundle, answered with JSON and quoted citations
+
+**Pros:** easiest for a generative harness, since it is one prompt per pass;
+nothing to translate; the model can explain itself.
+**Cons:** validating a quoted citation means fuzzy-matching model text against
+the proposal, so FR-010's "validated" becomes a similarity threshold that would
+itself need calibrating. Prompt text mixes instructions with record content,
+so a record can address the model directly. Decision models cannot answer it at
+all. It is also the shape that turns into a single-judge "looks good", which
+ADR-0005 was written against (superseded by ADR-0027, which keeps the point).
+
+### Option C: An injected model-client port inside adrkit
+
+**Pros:** `adr evaluate` could run Passes 2 and 3 itself, with no integration
+work for adopters.
+**Cons:** it requires amending Principle II ("Network-dependent tests and
+runtime behavior are forbidden"), which by the constitution's own Governance
+section needs a record first. It breaks the offline-use and IP-boundary
+rationale, puts credentials in a published package, and fails the three landed
+purity gates. **Rejected unless the constitution is amended first**, as 011
+states. The cost is recorded so a future reader can see exactly what this
+option would take.
+
+### Option D: Adopt the vendors' `/v1/systemone` request shape verbatim
+
+**Pros:** two vendors would need no translation; the format is proven in
+production.
+**Cons:** it binds a published contract to formats this project does not
+version. It has no `cite` or `text` type, so FR-010's citations and every Pass
+3 output would have to be bolted on outside it. It returns fractional scores
+that both vendors call uncalibrated, and it carries a 255-option ceiling. Option
+A keeps the three primitives aligned one-to-one with that shape, so a relay
+harness is a thin translation, without inheriting it.
+
+### Option E: Do nothing, and leave Q6 open
+
+**Pros:** no commitment before 012 freezes its holdout.
+**Cons:** 011 T003 says it plainly: deferring Q6 unblocks nothing, and T020
+would invent the contract ad hoc in the middle of implementation, which is
+exactly when a contract gets fitted to whatever the first model happened to
+return.
+
+## Trade-offs
+
+- **More round trips for a decision model.** A `cite` becomes one question per
+  span. A 40-block proposal scored on eight dimensions is hundreds of yes/no
+  answers. Decision models answer them in parallel and cheaply, but the harness
+  has to compose them.
+- **Generative responders lose some freedom.** A model that would cite a phrase
+  must cite a block. Blocks are coarser than phrases, and a block that only
+  partly supports a score still validates. Precision is traded for
+  checkability.
+- **The span cutter becomes contract.** Changing how blocks are cut changes
+  which ids a citation names, so it takes a new envelope token.
+- **The catalog is a second place rubric wording lives.** It must quote
+  `docs/EVALUATOR_RUBRIC.md`, not paraphrase it. Under ADR-0005 action item 4,
+  carried forward by ADR-0027 §4, a catalog change that alters a rubric
+  judgment is a rubric change and needs a record.
+- **Harness obligations are unverifiable.** Fresh context, unmodified questions,
+  and honest responder declarations are claims adrkit records and cannot check.
+
+## Consequences
+
+- **Easier:** implementing 011 Phase 1 against a fixed contract; testing every
+  validator branch with handwritten responses, with no model involved; running
+  the same bundle through a generative and a decision-model harness and letting
+  012's figures compare them.
+- **Harder:** the first harness integration is real work. The contract is
+  stricter than "return JSON", and a sloppy integration fails loudly with exit
+  `2` instead of degrading quietly. That is the intended direction of error.
+- **Unchanged:** Pass 0, its report bytes, its reason codes, the eight landed
+  triggers, the schema, `adr lint`, `adr check`, MCP, and the CI Action.
+- **How we would know this was wrong:**
+  - After 012's holdout is frozen and a first pass runs under it, more than 5%
+    of responses from any one supported harness are contract errors. That would
+    mean the contract is too strict to integrate against, not that the harness
+    is careless.
+  - Citation coverage, and so `low-confidence`, tracks block granularity rather
+    than proposal quality. 012's per-dimension figures would show that as
+    `low-confidence` firing mostly on proposals with few, long blocks.
+  - A supported harness cannot honor the fresh-context obligation, and nothing
+    lets a reader tell.
+- **Revisit if:** 012 rules on whether harness relevance counts as a ranking
+  function (clause 9), or a responder class appears that can neither generate nor answer
+  typed questions.
+
+## Action items
+
+1. [ ] The maintainer rules on 011 Q4 (citations in the record or not) and Q5
+       (exact per-tier weights, by its own record). Both remain open, and 011
+       T004 stays open with them.
+2. [ ] 012 T008 defines `RubricScoreSnapshot` and `AdversarialSnapshot` so they
+       can express validated span-id citations, per-slot presence, per-dimension
+       `bearsOn`, the declared responder, and optional probabilities as
+       evidence kept separate from the scores FR-019 reads.
+       This record constrains those shapes and does not define them.
+3. [ ] After 011's Phase 0 clears, and no earlier: the span cutter, the
+       question catalog, and the response validator in `@adrkit/evaluator`,
+       with every contract-error branch observed failing against a handwritten
+       malformed response (ADR-0016). No model call is involved.
+4. [ ] After 011's Phase 0 clears: the CLI's bundle emission and response
+       consumption, additive under ADR-0031, mapping contract errors to exit `2`.
+5. [x] Amend 011's Q6 to point here, and Q3 to name the candidate owner.
diff --git a/specs/011-probabilistic-evaluator-passes/spec.md b/specs/011-probabilistic-evaluator-passes/spec.md
index 015e5fe5..e435bd33 100644
--- a/specs/011-probabilistic-evaluator-passes/spec.md
+++ b/specs/011-probabilistic-evaluator-passes/spec.md
@@ -1158,6 +1158,14 @@ to prevent. Q7 was resolved during scoping and is recorded under
   with its own determinism obligations (`compareCodeUnits` ordering, byte-reproducibility) and its
   own record. Both specs therefore carry this as an open question with **no owner**, which is the
   accurate state. **Owner: unassigned — needs a decision on whether it is scoped at all.**
+  **Candidate owner proposed by
+  [ADR-0060](../../docs/adr/0060-drive-the-probabilistic-evaluator-passes-through-the-harness-with-typed-question.md)
+  clause 9:** relevance for retrieval category (a) as harness-supplied evidence, one `yes-no`
+  question per candidate, with no ranker in adrkit and categories (b) and (c) unchanged (FR-006).
+  The record defines the question's shape but has **no Pass 1 bundle emit it**: admitting a
+  candidate on a `true` answer would reinstate 012's withdrawn floor at 0.5. Whether harness
+  relevance counts as a ranking function, and with what floor, is 012's to decide by its own
+  record. Until then category (a) stays empty and `novel-no-precedent` stays `evidence-absent`.
 
 <a id="q4"></a>
 
@@ -1183,12 +1191,17 @@ to prevent. Q7 was resolved during scoping and is recorded under
 <a id="q6"></a>
 
 - **Q6 — `[NEEDS CLARIFICATION]` What is the prompt-bundle / structured-response contract with
-  the harness?** The harness-driven architecture is settled (see
-  [§ Architecture](#architecture-the-model-produces-evidence-deterministic-code-does-the-deciding)),
-  but its wire contract is not: the bundle format, the response schema, how a malformed or
-  refused response degrades (presumably to `PassAbsence` per A3), and whether the contract is
-  versioned independently like the schema. This is the main design surface the architecture ADR
-  should settle or explicitly defer. **Owner: 011, pending the architecture ADR.**
+  the harness?** **Proposed answer, pending ratification:**
+  [ADR-0060](../../docs/adr/0060-drive-the-probabilistic-evaluator-passes-through-the-harness-with-typed-question.md)
+  (**proposed**) ratifies the harness-driven architecture and settles the wire: versioned
+  `adrkit.evaluator.bundle/v1` and `adrkit.evaluator.response/v1` envelopes, one bundle per pass,
+  a `bundleId` the response must echo, typed questions (`yes-no`, `one-of`, `level`, `cite`,
+  `text`) whose wording comes from a committed catalog and never from record text, citations as
+  ids into spans cut deterministically before the call, unknown-key rejection, and a malformed
+  response as a typed contract error (exit `2`) rather than `PassAbsence`. It places the
+  fresh-context obligation (FR-014) on the harness. Until that record is **accepted**, T003 stays
+  open and this question is not resolved. If the record is rejected or revised, this entry changes
+  with it. **Owner: maintainer, by accepting or rejecting ADR-0060.**
 
 ## Decisions recorded during scoping