--- name: mc-adv-review-v17 description: "Run one independent act in a CANONQ-NATIVE v1.7 Mission, in mode review (the default) or verify, never as two skills (Definition Section 10; Profile 2.4 and 4.6). Review: exports each target with its id and binding or repository tip, spawns a reviewer in a separate context, transcribes its return lint-clean, publishes REVIEW-{SCOPE}-{NNN} and lands the kind-review ledger record with its structured workOrder. Verify: spawns the measuring context and publishes VERIFICATION-{SCOPE}-{NNN} with its provenance header and a ledger record produced by independent. Refuses to report success without the published asset id. Triggers on \"adversarial review v17\", \"mc-adv-review-v17\", \"run a review\", \"verify independently v17\". Arguments: the mode, then the target asset ids or titles." --- # mc-adv-review-v17 Read this file to the end before acting. A review is evidence, and evidence that is unpublished, uncited or undispatched is evidence nobody weighed; this skill exists because a hand-rolled protocol forgot a step twice, and one forgotten step was a false report of publication. A review is one shape of that evidence and a measurement is another, so both run here (IP-MC17-011 L8). `mc17` is the launcher `client/mc17` two levels above this skill's folder, from a POSIX shell, Git Bash included, and `client/mc17.cmd` from PowerShell or cmd. Every record this skill appends carries `recordedBy`, the name of the person or session that writes it down, on a ledger of either version. ``` mc-adv-review-v17 v0.2.0 (MC 1.7 - CanonQ-native) Parameters: ? | mode (review | verify, default review) | targets (asset ids or exact titles, one or more) | ordered-by (the authority's words ordering the act, verbatim) | brief (verify mode: what is to be measured and returned) Mechanical half: mc17 (version, hash, binding --class artifact, create, lintdoc, evidence) and the ledger append by patch. Contract: in, the targets; out, the published asset id, the ledger record number, the structured result. The structured result is the mode's own shape: in review mode it is the work order, in verify mode it is the measurement's items, each with its command, its figure and its verdict. ``` ## The two modes, one protocol `review` is this skill's present behaviour, unchanged: an independent reviewer READS the targets and returns findings, and the act publishes `REVIEW-{SCOPE}-{NNN}` with a work order. `verify` is the same protocol over a MEASUREMENT: the runner hands the target and the brief, the measuring context runs what the brief names and returns what it read, and the act publishes `VERIFICATION-{SCOPE}-{NNN}` and lands a record carrying `producedBy: "independent"` and `evidenceAsset` set to that asset id. The mode is a mode and never a sibling skill: an independent act is one genus, and a second skill beside this one would be a second chance to forget a step, which is the whole history this file carries. Where a step below names a difference, it names the mode with it; everything else is common to both. ## Step 0 - Which copy, which Mission, which line `mc17 version`, exact-title resolution of the Mission's `MANIFEST-{SCOPE}-001`, pin gate as everywhere (a v1.6 Mission has no review law in this line). Settle the mode first, because the mode settles the title: resolve the next `REVIEW-{SCOPE}-{NNN}` in review mode and the next `VERIFICATION-{SCOPE}-{NNN}` in verify mode, by listing the collection's titles under that prefix. **Those two title forms are the only names the evidence join reads as evidence** (IP-MC17-011 settlement 2): an asset published under a third name is invisible to the half of the join that hunts an evidence asset no record cites, so the form is load-bearing. Record who ordered the act, in the authority's words verbatim; an act nobody ordered is this skill refusing to invent authority - the runner may order one and the record then says so. ## Step 0b - The Mission's parameters, read first Run `mc17 params {manifest_asset_id} --ledger {ledger_asset_id}` and take two values, each from the row in force or from the default the plugin ships: `review.triage-before-obligation`, on by default, which decides the shape of Step 4, and `review.refs-scope`, target by default, which decides the refs the record carries. Say in the report which value each read and where it came from. ## Step 1 - Export the targets For each target: `mc17 hash` (or `binding --class artifact` for a governed artifact, both where useful), and for a repository target the tip sha. The export is what the reviewer reads, with each target named by id and figure, so the review binds what was actually reviewed and a later reader can tell whether it moved. In verify mode the figures do the same work for a measurement: they are what the provenance header will carry, and they are how a later reader tells whether the thing measured is the thing standing now. ## Step 2 - The brief, in the standard form The reviewer's brief states: the reviewer is INDEPENDENT and wrote none of the targets; it reads the exports whole; it cites evidence by file and line or section; it owes NO deference to approvals or to the authors' standing; it returns a verdict in its own words and a NUMBERED work order, each item carrying a severity (`critical`, `high`, `medium`, `low`) and an owner: an authority's word where only a ruling can close it, the pen where an edit can. The brief carries the targets and their figures, and nothing of the runner's opinions. The brief states the figure rule of Profile 7.12: a figure marked as a bound with its reason and its measurement or record is graded; an unmarked figure in prose is an estimate, a finding for the author and no block. In verify mode the brief carries the measurement: the measuring context is INDEPENDENT and wrote none of the targets; it runs each command the brief names and returns every command with its output tail; it names as unrun whatever it could not run; it invents no figure and it softens no failure; it returns a verdict per item in its own words. The runner's expectation of the result stays out of the brief, for the reason the reviewer's brief keeps the runner's opinions out. ## Step 3 - The independent agent Spawn the reviewing agent in a SEPARATE context that wrote none of the documents under review: start this package's custom agent mc-independent as a subagent, which holds nothing of this conversation: in the Copilot CLI through the tool that starts a subagent, naming the agent mc-independent, and in VS Code through the tool runSubagent with its agentName set to mc-independent (where VS Code offers no runSubagent, ask the developer to turn on Run Subagent under Configure Tools and continue once it is on; where neither client starts it, take the fallback below), hand it nothing but what this step names, and wait for its return: start it with the brief and the exports, no shared session memory with the author. The runner never reviews its own writing through this skill; where the runner authored a target, the separation is the whole point. In verify mode the same separation holds, and the reason is sharper: the recorder is a lossy channel by construction however honest, so the measuring context speaks for itself through the published asset. Where this host starts no agent in a context of its own, write the brief and the exports to files, hand the developer the brief to run in a new session of the host, and stop. Publish nothing and append nothing until the developer brings back that session's return; then continue at Step 4 with it. ## Step 4 - Transcribe and publish, one act 1. Transcribe the returned text lint-clean, keeping every claim; where the lint forces a rewording, the claim is kept and the provenance header says the transcription repaired lint and changed no finding. Compose the provenance header: who ordered it (verbatim), what was read (each target by id and figure, or the repository tip sha), the independence statement, and the transcription note. The header is the same in both modes, and it is what makes the published asset stand on its own before any record cites it. 2. `mc17 create --title REVIEW-{SCOPE}-{NNN} --collection {id} --file {f} --type REVIEW --authority A3 --lifecycle current` with the Mission's tags and domains; in verify mode the same guarded create with `--title VERIFICATION-{SCOPE}-{NNN} --type VERIFICATION`. `mc17 create` refuses a governed title without proof of the plain-language pass (`AC-24`). Here the reviewer's return is the draft and the transcription is the rewrite: run `mc17 plain --draft --pass {f} --out proof.json` and add `--plain proof.json` to the create. Where that command names something dropped, the transcription lost it, and it is repaired before anything is published. **The returned asset id with its hash proof is the only ground for any claim of publication** (Profile 4.6). No id, no claim: on a failed create this skill reports FAILURE with the error, in either mode, and nothing else it did counts as done. **The publication comes BEFORE the record that cites it**, because a record citing an asset that was never created is the citation of nothing the evidence join hunts. 3. In review mode, in the same act, the ledger record. Its shape is one in both branches below: kind `review`, body carrying the verdict in the reviewer's own words, `workOrder` carrying every item as `{id, severity, owner, disposition}` under the ledger schema (the shape of `harness/schemas/review-result-v1.schema.json`; the command validates the work order and refuses an entry outside it before the send), appended with `mc17 append {ledger} --record `, the record written to a file first. The refs of the review record: where `review.refs-scope` reads target, the review asset and the documents the brief named as targets; a document handed to the reviewer as context stands in the provenance header and in no ref; where it reads target-and-context, every document read. A document of another Mission is named in the body with its asset id and in no ref, because the evidence join reads a ref that carries a review or verification title as this collection's own evidence. The branch is the value `review.triage-before-obligation` read in Step 0b. **Where it reads off,** append the record now, every item `open`, and take the number the command prints. **Where it reads on,** the act forks after point 2: publish the review; run `mc17 published {target_asset_id} --ledger {ledger_asset_id} --manifest {manifest_asset_id}` over each target that is a document of the six classes, MANIFEST, SPEC, ARCH, IP, TP and WAVE-PLAN, and where the command refuses, or where the binding it prints differs from the binding the provenance header names for that target, stop before the triage and name the run owed, because the authority triages findings over the version the review read; put the reviewer's return with every finding before the approving authority, take his triage, and, where the session has announced no speaker for the words, ask who speaks them, in which role and under which name, before the record is written. The record carries `spokenBy`, one entry `{role} {name}` for each passage of `authorityWords`, in the order the passages stand. Then append his triage record first, kind `ruling`, his words verbatim, refs naming the review asset, and in its body the one line the gate reads, `Triage of REVIEW-{SCOPE}-{NNN}: sets aside {ids}`, the ids of the entries he set aside separated by commas, or `sets aside none`, and each line `Read: {asset id} version {n}, binding {hex}, verified at record {m}` the command printed, a target of another class and a repository tip carrying none, since the provenance header names them by id and figure; then append the review record, every finding he admitted `open` and every finding he set aside `ruled-away`, with `record` naming the triage record and `note` carrying his words for that item. The review record lands after his triage, and nothing lands on the ledger before he has read the findings. The gate admits a set-aside entry only where the triage record is a ruling with authority words, precedes the review record and carries the mark naming this review with the entry's id; an entry of any other disposition names a record that follows the review record, as before. The approving authority is the one the target's approval skill names: the Blast Radius for an IP, the Technical Owner for an ARCH or a Test Plan, the Business Owner for a SPEC; where the targets of one review answer to different authorities, each triages the findings on his own target. 4. In verify mode, the record of the same act, appended the same way, carries the kind the commissioning act calls for, `verification` where the measurement stands on its own, with refs carrying the published asset id AND every measured target by id, `producedBy: "independent"` and `evidenceAsset` set to that asset id, and the body carrying the measuring context's verdict in its own words. **The evidence field MAY summarize and it never replaces the asset**: a summary that names no asset is the exact failure L8 exists to end. Where this act publishes the report an EARLIER record owed, the same act appends one more record, kind `note`, whose text opens `RESOLVED {n}`, which declares `repairs: [n]`, the earlier record numbers this asset now carries the citation for, and names this asset in its refs: the ledger is append-only, the earlier record gains no field, and the declaration is what the join reads. A body that merely mentions the earlier number repairs nothing. Read `n` back by pointer. 5. Track the published asset beside the Manifest's table like every artifact that arises, where the Mission tracks rows. ## Step 5 - Report, the contract's out The published asset id, the ledger record number, and the structured result: in review mode the work order with every item standing `open`, or `ruled-away` citing the triage record where the triage parameter read on, in verify mode the measurement's items with their commands, figures and verdicts. That is the contract's out in both modes, and it is what the sweeps now watch: `mc17 evidence {collection}` fails while an item stands open, while an evidence asset stands in the collection that no record cites, while a citation resolves to no asset, while a record declaring `producedBy: independent` names no `evidenceAsset`, while a `producedBy` stands outside `independent|recorder`, and while a declared repair repairs nothing. The Move sweep records that verdict and the closure act refuses over a failure. Dispositions land as later records: `applied` with the rework's record, `argued` with the reason, `ruled-away` with the authority's words, each written into the record's workOrder entry (`disposition`, `record`, `note`) by the act that dispatches it. That write is the ONE the law admits inside a record already appended (Definition, *The review record*, L17), and `mc17 dispatch {ledger} --record {n} --item {id} --disposition {applied|argued|ruled-away} --by {n} [--note "..."]` composes and proves it: it refuses an item the record does not carry, a disposition that dispatches nothing, an `argued` or `ruled-away` with no note, and a dispatching record that does not exist or does not follow the review record, and prints the RFC 6902 operation array to send through the patch door. Append the dispatching record first, then name it. Where the entry's own dispatch names a record from before the review record, a later re-dispatching ruling dispatches it from outside the entry, as step 6 says. ## Step 6 - The dispatch, later A review's act ends with every item standing `open`, or, where the triage parameter reads on, with the entries the authority set aside standing `ruled-away` citing his triage record, which precedes the review record and which the gate admits as the dispatching record, as the Definition's dispatch check says since its amendment at `VF-MC171-001` record 490. The gate reads the parameter as it stood at the review record's number: a row whose record follows the review record does not govern it, and the default applies to that review (the Technical Owner's ruling, `VF-MC171-001` record 500). Each open item is dispatched afterwards, by the act that applies it, argues it or carries the authority's ruling, and that dispatch is the ONE write the law admits inside a record already appended (Definition, *The review record*, L17). 1. Append the dispatching record first: the rework's record, the argument, or the authority's own words. The dispatch names it, so it exists before it is named. The record the dispatch names under `--by` names the review `REVIEW-{SCOPE}-{NNN}` in its refs or its body. 2. Compose the write: `mc17 dispatch {ledger} --record {n} --item {id} --disposition {applied|argued|ruled-away} --by {the record from step 1} [--note "..."]`. It refuses an item the record does not carry, a disposition that dispatches nothing, an entry already dispatched, a ledger holding a closure record, an `argued` or `ruled-away` with no note, a dispatching record that does not exist or does not follow the review record, a dispatching record of verdict FAIL (`record {n} carries verdict FAIL: a failed act dispatches nothing`), and, where the review record names a REVIEW asset in its refs, a dispatching record that names that review in neither its refs nor its body (`record {n} names the review REVIEW-{SCOPE}-{NNN} in neither its refs nor its body: a dispatching record names the review it dispatches`). It prints the RFC 6902 operation array it composed, sends it through the platform's content endpoint, reads the entry back at its pointer and prints the entry as stored; `--dry` prints the array and sends nothing. 3. The command sent the array and read the entry back at its own pointer, requiring the disposition, the record and the note to read as sent. Where it printed a difference, the dispatch did not take, and the entry is read before anything is sent again. An entry is dispatched once: a disposition an authority gave is changed by a later record that says what changed, never by a second write to the entry. An entry carrying no dispatching record stands dispatched and is named by `mc17 evidence` and `mc17 reviewed` as naming none, which is what every dispatch written before this rule reads as. Where an entry's dispatch names a record that does not follow the review record, as when the review was recorded after the records that applied it, first take the re-citation, `mc17 dispatch --recite`, where the review record names its REVIEW asset and the disposition stays the same; otherwise a later ruling dispatches the entry: the authority's own words with his speaker line, a first line that reads exactly `REDISPATCH {n}` with the review record's number, and one line `Entry {id}: {applied|argued|ruled-away}` for each entry, appended with `mc17 append`; `mc17 evidence` and `mc17 reviewed` read that ruling as the entry's dispatch, the entry stays as written, and an entry no such ruling names stays open. ## Rules this skill cannot check - Whether the review's findings are right: the authority who reads them, and the dispositions are where the answer lands. - Whether the reviewing or measuring agent was truly independent: the spawn is this skill's act and the header says what was done; a reader judges the claim against the session record. - Whether a HAND-WRITTEN record declares its independence: no machine reads which context took a measurement (IP-MC17-011 settlement 3). Verify mode sets `producedBy` and `evidenceAsset` by construction, so an act carried through this protocol is declared without anyone remembering to; a record written by hand outside it stands on the runner's honesty, and the Rulings register carries that residual on its face. - Whether to order a review at all: the authority's judgment, and the culture this law carries without replacing.