[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7AP7P0SFS92RBP4TPTVVQ5B":3,"skill-same:01M7AP7P0SFS92RBP4TPTVVQ5B":89},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":26,"commitSha":27,"skillDirectory":28,"forksCount":29,"repositoryHtmlUrl":30,"repositoryHomepageUrl":6,"ownerHtmlUrl":31,"ownerBlogUrl":32,"ownerDescription":33,"updatedAt":34,"install":35,"resources":48},"01M7AP7P0SFS92RBP4TPTVVQ5B","protasis",null,"Hold a Fiat study and runbook to a content contract before any code is written: stated assumptions, testable success criteria, a chosen design with its trade named, the five discipline questions answered before any build, and steps that are discrete, green at both ends, sized for the audit loop and explicit about the gates they incur, including a source-bound known-failure inventory when audit history names one. Use when a topic is about to enter the study or runbook phase, when a requirement arrives vague or bundles several capabilities, when deciding whether a runbook is ready to build from, or when checking declared commands against source interfaces without executing them. Do not use it to run the controller or write a receipt, which belong to fiat, and do not use it to record a decision after the fact, which belongs to hypomnema.","wildcat-finance\u002Fskills","wildcat-finance","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F113041915?v=4","01M7AP0V1G8X0SWHP6C5V90RM5","The Wildcat Protocol","01M7AP0V24Y50BGVCQ4DV496VD",13,"Apache-2.0","https:\u002F\u002Fgithub.com\u002Fwildcat-finance\u002Fskills\u002Fblob\u002Fmain\u002Fplugins\u002Fhexaemeron\u002Fskills\u002Fprotasis\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: protasis\ndescription: >-\n  Hold a Fiat study and runbook to a content contract before any code is\n  written: stated assumptions, testable success criteria, a chosen design with\n  its trade named, the five discipline questions answered before any build, and\n  steps that are discrete, green at both ends, sized for the audit loop and\n  explicit about the gates they incur, including a source-bound known-failure\n  inventory when audit history names one. Use when a topic is about to enter\n  the study or runbook phase, when a requirement arrives vague or bundles several\n  capabilities, when deciding whether a runbook is ready to build from, or when\n  checking declared commands against source interfaces without executing them. Do\n  not use it to run the controller or write a receipt, which belong to fiat,\n  and do not use it to record a decision after the fact, which belongs to\n  hypomnema.\nmetadata:\n  version: \"6.20.1\"\n---\n\n\u003Cp align=\"center\">\n  \u003Cimg src=\"..\u002F..\u002Fassets\u002Fcharacters\u002Fprotasis.png\" width=\"1200\">\n\u003C\u002Fp>\n\n# Protasis\n\nFrom *protasis*, the proposition laid down before the argument runs. Nothing is\nbuilt from a topic; things are built from a proposition about a topic.\n\n## Where this sits\n\nProtasis owns the content contract for the `study` and `runbook` phases: what\nthose two documents must answer before implementation is allowed to start. It\nowns no state, writes no receipt and gates nothing itself.\n\nSurveyor may write the study packet, while Fiat keeps the artefact paths and\nreceipt commands. Mason and Warden later receive the exact source-bound\nrunbook step. Phylax, Ephoros, Metron, Elenchus, and Hypomnema answer the five\ndiscipline questions Protasis requires; Protasis cites their contracts rather\nthan copying them. A decision made after the study belongs to Hypomnema's\nrecording rules.\n\nSynkrisis proposes a next owner from a validated series of run observations.\nThat finding still needs a scoped proposition before it can become a study,\nrunbook, or framework change.\n\nIts version, held frontier, next job, and maturity state live in\n[EVOLUTION.md](EVOLUTION.md).\n\n**Current state.** Protasis checks the fixed mechanical shape of study and\nrunbook prose, one source-bound known-failure inventory, and one closed\ncandidate-by-criterion design record at the transition where each item of\nevidence becomes due. The optional command check validates literal arguments\nand finite per-file loops against registered source interfaces without\nexecuting the declared commands. Fiat's Step 5 demonstration separately\nrecords the successor controller's executed result and bounded refusal cases;\nthat report does not enlarge Protasis's declaration or claim criterion\nsufficiency.\n\n## Refuse these six\n\n1. No study, no runbook. A runbook derived from conversation instead of a\n   written study is not a runbook.\n2. No runbook, no implementation. A step that does not exist in the runbook\n   does not get built, however small it looks.\n3. No criterion, no success. A study whose success condition cannot be checked\n   by a command or a named demo path is unfinished.\n4. No stated assumption, no spec. Assumptions go on the page before the\n   content they support.\n5. No checked design record, no design lock. A preferred construction in prose\n   does not authorise implementation.\n6. No closed known-failure inventory, no production edit. Missing, stale, or\n   unassigned entries cannot be repaired by treating an empty list as clean.\n\nReport a refusal in three parts: what is missing, where you looked, and the one\naction that clears it. Say plainly that the phase is blocked rather than\nin progress. None of the four is a suggestion to proceed carefully.\n\nIf the gap is an ambiguity rather than an absence, ask one literal question\ninstead of picking a reading.\n\n## Assumptions go first\n\nList what you are assuming before writing any spec content, and say plainly\nthat you will proceed on them unless corrected.\n\n```\nAssuming, unless corrected:\n1. Foundry, not Hardhat; the repo has foundry.toml and no hardhat.config.\n2. Solidity 0.8.x with checked arithmetic; unchecked blocks are opt-in per site.\n3. The exact interpreter in the repository's `.python-version`, with stdlib unittest.\n4. An archive RPC is available for the capture step; without one, step 3 changes.\n```\n\nAn unstated assumption is the failure this phase exists to catch. On the page,\na wrong one costs a sentence. Buried in step 4, it costs the step.\n\n## Vague requirement, testable criterion\n\nRestate the request as conditions a command can check, then confirm the\nrestatement before building on it.\n\n```\nRequirement: \"make the harvester faster\"\n\nRestated:\n- A full Ethereum USDC interval harvest finishes inside 20 minutes on a warm cache.\n- The digest of the produced release is byte-identical to the current one.\n- Peak resident memory stays under 2 GB.\nMeasured by metron, before and after. Are these the right targets?\n```\n\n\"Faster\", \"cleaner\", \"more robust\" and \"production-ready\" are not criteria.\nNeither is a criterion that can only be checked by asking someone whether they\nare happy.\n\n## What a study must answer\n\n1. **Problem statement.** What is being built, for whom, and what a working\n   prototype means here. Name the demo path or the check that proves it.\n2. **Prior art.** What exists already, in this repo, in the organisation's\n   other repos, and outside. Name files, packages and standards by identifier.\n   Where the topic touches something that has shipped before, read the last two\n   merged pull requests that changed it before writing anything else. A run\n   that could not finish something records it in the body of the last pull\n   request it lands, so that is where the unfinished work of the previous run\n   is written down. Carry each item forward as content here, as a stated\n   non-goal, or as a named reason it stays open. Read the audit records of\n   every in-scope skill the same way, before design options are drawn. Every\n   discovered audit source remains authoritative. A verified synopsis is its\n   normal reading view only after\n   `python3 plugins\u002Fhexaemeron\u002Fskills\u002Ffiat\u002Fscripts\u002Faudit_synopsis.py --check\n   \u003Ctarget-root>` runs from the target root and the whole-set currency check\n   exits zero. `**\u002Faudit\u002FAUDIT.md` maps to its sibling `AUDIT_SYNOPSIS.md`. A\n   direct child `audit\u002Frounds\u002F\u003Crun>.md` maps to\n   `audit\u002Frounds\u002F\u003Crun>.synopsis.md`. The root pair covers only the root source;\n   it does not cover per-run or plugin records.\n\n   Missing, stale, unsupported, or unavailable view means: read the\n   authoritative source directly and record its source path and the reason.\n   If the whole-set check fails, do that for every in-scope source rather than\n   guessing which view remains current. A target without the renderer or a\n   committed synopsis can still be studied when its source is available. If\n   neither view nor source is available, readiness stops on that evidence gap.\n\n   Keep every finding id and status, `Covered`, `Not checked`, `Elenchus\n   verdict`, and `Leads not pursued`; `[missing legacy field: ...]` remains\n   unknown. Then name every in-scope source, which synopsis or source was\n   actually read, and the evidence for that choice. Do not claim the source\n   was read when only its synopsis was read. The run's audit file, at its\n   `config audit.log_path`, holds what each round found, what was fixed, and\n   the leads accepted with the reason nobody pursued them. Fiat derives one\n   such file per run under `audit\u002Frounds\u002F`, so the records of several runs are\n   several files; a target that kept a shared log before that change has the\n   rest of its history in `audit\u002FAUDIT.md`. A study that cannot find the real\n   reason for a decision supplies a plausible one, and the plausible one then\n   governs the build: a rejected option gets rejected for the wrong cause, an\n   accepted risk gets quietly reopened, and the round that already judged the\n   question reads afterwards as though it never happened. Found in the study\n   it costs a sentence; found in step four it costs the step.\n3. **Constraints and non-goals.** The starting ref, toolchain and version pins,\n   what the user ruled out, what is deferred past the prototype.\n4. **Design options.** Two to four candidate constructions, each with the trade\n   it makes. The prose explains the candidates; the design-evidence record\n   below selects one from checked gates and comparative measurements.\n5. **Risk register seed.** What the audit loop should look hardest at. In\n   Solidity: trust boundaries, external calls, arithmetic, upgrade paths, key\n   custody. In Python: untrusted input, subprocess and filesystem handling,\n   secret material, partial writes, and what happens when a long run is killed\n   halfway. The concerns go in a fenced block the audit loop can enumerate,\n   info string `risk-register`, one concern per line as three pipe-separated\n   fields -- a kebab-case id stable within the study, the boundary the concern\n   sits at, and what the audit loop checks:\n\n   ```risk-register\n   subprocess-input | the argv of the spawned compiler | inputs are pinned and no shell is used\n   partial-write | the release directory during a long harvest | a killed run leaves no half-written file that verifies\n   ```\n\n   A round then logs each id as reviewed or not applicable, so the look is an\n   enumerable obligation rather than a judgement call nobody can verify\n   afterwards. Prose around the block carries what a line cannot, and the ids\n   are how a round cites it.\n6. **Glossary seeds.** Terms the runbook and implementation will reuse, one\n   line each.\n7. **Sources.** Enough of a pointer to find each one again.\n\nThe five that follow are the disciplines the build will be held to. Answering\nthem here costs a sentence each; leaving them to the audit loop costs a step.\n\n8. **Signals, and the questions behind them.** Two to four questions someone\n   will ask at three in the morning once this runs unattended, and which steps\n   emit the signals that answer them.\n   [ephoros](..\u002Fephoros\u002FSKILL.md) owns what a signal must carry.\n9. **Boundaries, per capability.** Each boundary this opens, what is worth\n   taking at it, and the control that closes it. This feeds item 5 rather than\n   replacing it. [phylax](..\u002Fphylax\u002FSKILL.md) owns the boundary list and the\n   controls.\n10. **The budget, or its absence.** Any performance budget this must hold, with\n    the exact command that measures it.\n    [metron](..\u002Fmetron\u002FSKILL.md) owns what a budget carries and how it is\n    checked.\n11. **The fail-closed posture.** What stops the run, and the guard-test\n    convention a fix will follow.\n    [elenchus](..\u002Felenchus\u002FSKILL.md) owns the triage order and the guard rule.\n12. **Decisions and their homes.** The decisions expected to be expensive to\n    reverse, and the file each record will live in.\n    [hypomnema](..\u002Fhypomnema\u002FSKILL.md) owns which decisions earn a record and\n    where each one lives.\n\nFor items 8 through 12, \"none, and here is why\" is a complete answer. A lint\ninvoked from a terminal has no on-call question; a step that reads two files\nhas no budget. What is not a complete answer is silence, because silence cannot\nbe told apart from not having looked.\n\nCite those five contracts, never restate them. Each owns its own rules and each\nevolves on its own ledger, so a copy here is stale the moment it is written.\n\n### Known-failure inventory\n\nWhen audit history names failures that implementation must guard before product\nwork starts, the study carries exactly one fenced\n`known-failure-inventory` JSON object under schema\n`protasis-known-failure-inventory\u002Fv1`. Each `source_views` entry names a\nportable synopsis path and the SHA-256 of both that view and the authoritative\nsource named by its header. Each finding has one unique id, a source reference\nwhose prefix names a checked view, bounded non-empty locator text, a failure\nstatement, closed guard-path set, structural test command, report format and\npath, exact expected verdict `guarded`, fixed-tree command, and one consuming\nrunbook step. The locator text is recorded but not interpreted as a line or\nclaim proof. The inventory fence is column-zero and blank-line isolated.\n\nThe test command is a bounded six-argument invocation: `python3`, one declared\nPython reporter path, `--case`, the finding id, `--report`, and exactly one\nargv equal to `{report}`. Its fixed-tree counterpart differs only in the report\nargument. Each report and derived green path is unique under `.elenchus\u002F`.\nFilesystem-backed uniqueness uses Unicode NFC plus case folding for portable\nalias detection.\nAdmitted report formats are `unittest-json-v1`, `forge-junit-v1`, and\n`node-test-json-v1`. The caller supplies an independent expected-id set, or the\nchecker uses ids in visible assignment records; each record names a real\npositive, column-zero runbook Step, and the inventory's consuming Steps must\nmatch. Each assignment belongs to a contiguous nonblank block containing\nassignment lines only, so adjacent prose, image labels, link titles, and\nreference definitions cannot supply one. The only authoritative assignment is\na visible full line of the form\n``Known-failure assignment: `kf-453-N` -> Step S`` outside fences. Raw HTML,\ncomments, and malformed HTML-like bytes outside one-line code spans refuse at\nthe runbook boundary. Every square-bracket byte outside a complete one-line\ncode span or fenced block also refuses there. The study keeps its separate\nsurface rules. Its inventory discovery tolerates unmatched inline-backtick runs\nin ordinary prose, but never uses them to mask bytes. It processes or refuses\neach column-zero fence candidate before that tolerance. Raw HTML, image openers,\nindented fences, bad fence information, unclosed fences, and anything other\nthan one isolated inventory block still refuse. Fenced examples use column-zero\nfences. Nested or indented fence candidates refuse. At most 128 findings, 4,096\naggregate guard paths, 1,024 bytes per path, 4,096 bytes per source reference,\nfailure statement, or command, and 16 command arguments are checked. Paths are\nroot-relative portable paths without shell metacharacters or Unicode\nwhitespace. The checker opens source paths through no-follow, nonblocking\ndirectory handles. It rereads the study, runbook, views, and sources together\nbefore returning success.\n\nA non-empty finding set requires `no_known_findings: null`. An empty set needs\na non-null claim that repeats every checked source id and both digests, names\none real consuming step, and carries the exact Surveyor assertion\n`no-known-findings`. An empty array alone never establishes readiness. The\npublic `load_checked_inventory(study, runbook, repository, expected_ids=None)`\noperation returns one frozen `InventoryLoadResult`. Its `status` is exactly\n`absent`, `refused`, or `clean`. `absent` carries neither a capture nor\nfindings and is available only when the stable study has no inventory marker\nand the stable runbook has no assignment marker. Any attempted, partial, or\nmalformed marker stays on the existing K000 through K012 refusal path.\n`refused` carries a non-empty tuple of those findings and no capture. `clean`\ncarries no findings and one JSON-compatible capture with exactly `schema`,\n`study_sha256`, `runbook_sha256`, `inventory_sha256`, `source_views`,\n`findings`, `no_known_findings`, and `assignments`.\n\nThe clean schema is `protasis-known-failure-inventory-capture\u002Fv1`. Study and\nrunbook digests cover their exact stable bytes. The inventory digest covers\nthe accepted parsed object encoded by Fiat's canonical JSON form: sorted\nobject keys and compact separators. Source views, findings, and the no-known\nclaim are the complete parsed inventory values. Each assignment has exactly\n`finding_id` and `step`, ordered by Step and then finding id. The operation\nrevalidates every input before returning the capture. The compatibility\n`check` operation and command-line checker are projections of this result;\nthey never reread or reparse a clean inventory. The checker writes no\ncontroller state or ledger. It establishes structure and current bytes, not\nthat Surveyor found every failure or that any guard has run.\n\n### Design evidence and progressive gates\n\nThe study writes one closed record to `.hexaemeron\u002Fdesign-evidence.json` under\nschema `protasis-design-evidence\u002Fv1`. The record is the selection interface;\nprose is not. It contains two to four candidate ids, one to 32 criteria, the\nexact candidate-by-criterion result matrix, and one selection.\n\nEvery criterion names one of five concerns: `correctness`, `time`, `space`,\n`compatibility`, or `recovery`. The complete set must be covered. A criterion\nalso names its owner and one of two forms:\n\n- A hard gate has stage `selection` or `conformance`, compares a typed value by\n  `equals`, `at-most`, or `at-least`, and blocks `design-lock`, `step:N`, or\n  `integration`.\n- A comparative metric has stage `selection`, minimises or maximises `bytes`,\n  `count`, `milliseconds`, or `ratio`, and blocks `design-lock`.\n\nThe record uses these exact object forms; this excerpt is not a complete\nmatrix:\n\n```json\n{\n  \"schema\": \"protasis-design-evidence\u002Fv1\",\n  \"candidates\": [{\"id\": \"streaming\", \"summary\": \"Process one bounded window.\"}],\n  \"criteria\": [{\n    \"id\": \"peak-space\", \"concern\": \"space\", \"kind\": \"metric\",\n    \"stage\": \"selection\", \"owner\": \"metron\", \"unit\": \"bytes\",\n    \"comparator\": \"minimise\", \"threshold\": null, \"blocks\": \"design-lock\"\n  }],\n  \"results\": [{\n    \"candidate\": \"streaming\", \"criterion\": \"peak-space\", \"state\": \"pass\",\n    \"report\": {\"path\": \"reports\u002Fstreaming-peak-space.json\", \"sha256\": \"\u003C64 lowercase hex>\"}\n  }, {\n    \"candidate\": \"streaming\", \"criterion\": \"restart-safe\", \"state\": \"pending\",\n    \"resolver\": \"python3 tests\u002Fprove_restart.py\",\n    \"report\": \"reports\u002Fstreaming-restart-safe.json\", \"blocks\": \"step:2\"\n  }],\n  \"selection\": {\"candidate\": \"streaming\", \"rule\": \"unique-frontier\", \"policy_ref\": null}\n}\n```\n\nThe abbreviated example shows both result forms; an accepted record still has\ntwo to four candidates, all five concerns, and every matrix cell.\n\nA resolved matrix cell is `pass` or `fail` and names a report path plus its\nSHA-256. The report is one closed `protasis-design-report\u002Fv1` object containing\n`candidate`, `criterion`, typed `value`, `unit`, `command`, and integer `exit`;\nexit must be zero. Paths are relative to the record directory and may not cross\na symlink. A pending cell instead names the exact resolver command, future\nreport path, and stop point it blocks. Selection evidence may be pending while\nthe record is drafted, but not when the design locks. Conformance evidence may\nremain pending until its named step or integration transition. An unknown is\ntherefore a scheduled refusal, not a guessed score.\n\nSelection is closed and mechanical. A failed selection gate removes a\ncandidate. The checker computes the non-dominated frontier from the selection\nmetrics. `unique-frontier` requires one survivor. `exact-tie-simplicity`\napplies only when every checked comparative value is exactly equal across at\nleast two survivors. `user-policy` requires at least two survivors and a\nbounded `policy_ref`. Simplicity has no other tie-breaking authority.\n\nRun both checks before the study is receipted:\n\n```bash\npython3 \"$PLUGIN_ROOT\u002Fskills\u002Fprotasis\u002Fscripts\u002Fprotasis.py\" --study .hexaemeron\u002Fstudy.md\npython3 \"$PLUGIN_ROOT\u002Fskills\u002Fprotasis\u002Fscripts\u002Fdesign_evidence.py\" \\\n  .hexaemeron\u002Fdesign-evidence.json --transition design-lock\n```\n\nThe runbook binds the accepted bytes before Step 1:\n\n```design-lock\nschema | protasis-design-evidence\u002Fv1\nsha256 | \u003Csha256 of .hexaemeron\u002Fdesign-evidence.json>\ncandidate | \u003Cselected candidate id>\n```\n\nP007 checks the block's closed shape and position. Fiat checks that its values\nmatch the study receipt, then runs the design checker at `step:N` immediately\nbefore opening that step and at `integration` before admitting the completed\nstack. Each transition records the reports consumed there. Verification\nreplays every recorded transition against the unchanged record and report\nbytes. A run created before the contract has no `contracts.design_evidence`\nmarker and remains on the legacy path; no evidence is fabricated for it.\n\nA section reading \"TBD\" is a section to fill or cut. Where the request is\nambiguous, record the reading you chose and the reason for it. Never resolve an\nambiguity silently.\n\n## What a runbook step must contain\n\nA step is discrete: one pull request, one boundary. It is self-contained, so\nsomeone holding only the study, the runbook and the repo at that step's entry\nstate can finish it. Both ends are green; no step hands the next a broken tree.\nAnd it stays small enough to audit, because that phase dominates the clock.\n\n```markdown\n## Step N: \u003Ctitle>\n\n**Goal.** One sentence.\n**Entry.** The exact ref or state this step starts from.\n**Exit.** Deliverables, plus the command or test that proves them.\n**Files.** Paths created or changed.\n**Tests.** What gets written or extended, and the expected count if known.\n**Disciplines.** Which of the five apply to this step and why, or none with\nthe reason.\n```\n\nFor a Fiat step, `Tests` also owns the runner contract used when its audit\nclaims a fix. It names the exact Elenchus test command with one `{report}`\nargument, the report format, and the report file. Fiat sends the complete\nsource-bound step to Warden; Warden owns those three inputs and may not infer a\ncommand from `Files` or replace one with a nearby suite.\n\nThe Disciplines line carries the study's items 8 through 12 down to the step\nthat actually incurs them. Name each discipline and the reason in one clause:\n\n```text\n**Disciplines.** phylax: this step opens the ingestion path. ephoros: it runs\nunattended once deployed. metron: none, no performance claim. elenchus: none,\nno failure in hand. hypomnema: the storage format is expensive to reverse.\n```\n\nA step that names a discipline without a reason has not answered; a step that\nomits one has not been asked. Mason builds against the declared gates and\nwarden audits against them, so a gate named here is cheaper than the same gate\ndiscovered in round three.\n\nThree fixed points. Step 1 scaffolds: layout, toolchain pins, CI stub, licence,\nand committed copies of the study and runbook. The last step demonstrates, by\nrunning the demo path from the problem statement. Ordering is dependency order,\nand a step may assume every earlier step's exit state and nothing else.\n\nIf a step's exit cannot be proved by a command, it is not an exit. \"Reviewed\",\n\"working\" and \"integrated\" prove nothing on their own.\n\n### Version relations\n\nA runbook may carry one optional fenced `version-relations` block before Step\n1. Absence preserves the literal-only contract. A present block carries at\nmost 32 physical rows, with no blank row, in this closed shape:\n\n```version-relations\nprotasis | plugins\u002Fhexaemeron\u002Fskills\u002Fprotasis\u002FEVOLUTION.md | next-generation-after-integration-base\n```\n\nEach row is `skill id | EVOLUTION.md path | relation`. The skill id is\nkebab-case. The path is repository-relative, contains no empty, `.` or `..`\nsegment, backslash or control character, ends in `EVOLUTION.md`, and has the\nskill id as its parent directory. Ids and paths are unique. The sole admitted\nrelation is `next-generation-after-integration-base`; a partial target list is\nvalid because an omitted target may retain an intentional literal.\n\nA declared target has no concrete\n`\u003Cskill>-v\u003Cevolution>.\u003Cgeneration>.\u003Cepoch>` token outside the block. Prose,\nexamples, commands and amendments all count. P006\nchecks this shape and lexical identity. It does not open the ledger, decide\nwhether the relation suits the change, allocate a version, or establish what\nthe integration base will be.\n\n## When one topic is several\n\nMost topics are one capability and go straight to the study. Decompose first\nonly when a single request bundles capabilities that could ship and be verified\nseparately, or when one could be cut without rewriting the others.\n\nThe decomposition is a table and a build order, not a project plan:\n\n| Module id | Responsibility | Depends on |\n| --- | --- | --- |\n| capture | Fixed-block RPC capture, digest-keyed | none |\n| verify | Proof checks over a capture | capture |\n| replay | Local replay boundary, no fallback | capture |\n| fixture | Published fixture and its manifest | verify, replay |\n\nBuild order: capture, then verify and replay, then fixture.\n\nModule ids are kebab-case, chosen once, never renamed mid-topic. Dependencies\npoint one way; if two modules each need the other, they are one module. An\ninterface belongs in the spec of the module that provides it. Each module then\ngets its own study, and its modules become runbook steps in that order.\n\n## Boundaries the study must state\n\nThree tiers, each with concrete entries:\n\n- **Always.** Both test suites before a commit. The imprimatur lint on every\n  shipped document. A recorded measurement before any performance change.\n- **Ask first.** Adding a dependency. Changing a storage layout or a public\n  ABI. Touching CI. Widening a trust boundary. Rewriting a released digest.\n- **Never.** Commit key material or an RPC credential. Edit a vendored\n  directory. Delete a failing test to make a suite pass. Claim a command ran\n  when it did not.\n\n## The mechanical subset\n\nWhether the twelve items are present, and whether items 8 through 12 carry an\nanswer, is settled by a parser; so is the runbook step schema. Run the bundled\ncheck over each artefact and require exit 0:\n\n```bash\npython3 \"$PLUGIN_ROOT\u002Fskills\u002Fprotasis\u002Fscripts\u002Fprotasis.py\" --study \u003Cstudy>\npython3 \"$PLUGIN_ROOT\u002Fskills\u002Fprotasis\u002Fscripts\u002Fprotasis.py\" \u003Crunbook>\npython3 \"$PLUGIN_ROOT\u002Fskills\u002Fprotasis\u002Fscripts\u002Fknown_failure_inventory.py\" \u003Cstudy> \u003Crunbook> --repository \u003Croot> --expected-id \u003Cid> [--expected-id \u003Cid> ...]\npython3 \"$PLUGIN_ROOT\u002Fskills\u002Fprotasis\u002Fscripts\u002Fdesign_evidence.py\" \u003Crecord> --transition \u003Cdesign-lock|step:N|integration>\n```\n\nThe study mode reads items as `## N. Title` headings, 1 to 12, refuses\nsilence and a bare none on items 8 through 12, and reads item 5's\nrisk-register block against the shape above: S005 when no block names a\nconcern, S006 when a line does not split into the three pipe-separated\nfields, S007 when a field is malformed. S008 reports a real study amendment\nwhose heading is not a calendar date, whose four fields are missing,\nduplicated, reordered, unknown, or empty, or which is not final. A study with\nno amendment and amendment examples inside fences remain unchanged. The\nrunbook mode reads the step\nschema above, ends the last baseline step before a real amendment heading and\nreports P005 when a runbook amendment does not carry the dated four-field\nshape and complete replacement clauses below. It reports P006 when a present\nversion-relations block is open, misplaced, duplicated, oversized or malformed,\nor when a declared target is also pinned to a concrete token outside that\nblock. P007 checks a present design-lock block's position, closed rows, schema,\ndigest and selected-candidate shape. Codes P000 to P007, S000 to S008, and\nK000 to K012 are stable interfaces other tools cite. Deliberate exceptions\nstate a reason:\n`\u003C!-- protasis: allow \u003Cwhy> -->` on the heading line or the line above it.\nPresence and shape are all the parser settles; whether an answer is any good\nstays with the reviewer and the rest of this contract.\n\n## Check declared command interfaces\n\nUse `--gate-root \u003Ctarget-root>` with the runbook checker to validate the\neffective Exit and Elenchus commands against registered local CLI sources.\nUnsupported shell evaluation, argument shapes, report substitutions and source\nadapters refuse with P008. The separate\n[command reference](references\u002Fgate-commands.md) names the grammar, source\nbindings, report format, limits and replay rules.\n\nA target outside the Skills distribution may register up to 32 reviewed local\nPython interfaces in the runbook's closed `command-interfaces` block. Each row\npins the complete source SHA-256 and parser-builder name. The built-in registry\ncannot be overridden. A dated append-only amendment may replace the complete\nlocal registration set; an unregistered or changed source still refuses.\nRegistration binds the reviewed source and permits inert argument checking.\nIt grants no authority to execute the command.\n\nA row may put `step:1` in place of the digest for a runner that Step 1\ncreates. Until the caller supplies a binding, the adapter never reads or hashes\nthat path. Capture before Step 1 can require it absent, and each command naming\nit records `interface-deferred` after every check except the runner's parser\ninterface. With a binding the row is an ordinary pinned registration. The\nreference gives the placement rules, refusals and the phase record the caller\npasses.\n\nA caller may also pass `starting_bindings`, what it derived from a run's\nstarting commit: that commit's adapter digest and, for each registered module,\nthe AST digest that commit's adapter pinned with the module's complete source\ndigest. A registered module whose AST digest left the current pin is admitted\nonly when both digests equal its pair, and replay substitutes a receipt's\nadapter digest only when it equals the supplied one. Without the keyword every\nresult is unchanged. The adapter reads no Git and records no provenance; the\ncaller derives the bindings and owns that record. The reference names the\nclosed shape.\n\nThe result is `protasis-gate-commands\u002Fv1` with `operation_ran:false`. It binds\nraw commands, the captured source root, original and substituted argv, full CLI\nand adapter digests, and interface results. It establishes no execution, test success or audit verdict.\nFiat owns receipt storage, the new-init marker, legacy compatibility and the\nappend-only amendment that refreshes stale evidence.\n\n## The spec stays alive\n\nWhen a decision changes, change the study first and the code second. When scope\nmoves, say so on the page. Both documents are committed and reviewed like any\nother shipped artefact, and both go through the prose pass first.\n\nA mid-run change is an amendment, and an amendment appends rather than edits,\nso the run's earlier belief stays readable. It is a dated block at the end of\nthe study carrying four fields:\n\n```markdown\n### Amendment -- 2026-08-20\n\n**What changed.** The capture step reads the header from the fixture, not RPC.\n**Why.** The archive endpoint was withdrawn mid-run.\n**Steps touched.** Step 3's entry and step 4's files.\n**Still holding.** Step 3: entry broken; exit holds. Step 4: entry holds; exit\nholds. Step 5: entry holds; exit holds. Step 6: entry holds; exit holds.\n```\n\nEvery unbuilt step gets a verdict in the last field: its entry and exit hold,\nor they do not. Do not proceed past a step whose entry the amendment broke;\nreport it as blocked in the contract's three parts and re-derive the runbook\nfrom the amended study, or re-specify the step, before building on it. The\ndecision that forced the change is recorded where\n[hypomnema](..\u002Fhypomnema\u002FSKILL.md)'s rules put it, and the amendment points at\nthat record rather than restating it.\n\nA receipted runbook is amended by appending the same dated block, never by\nediting or repeating a numbered step. Its `What changed` value is one or more\nclauses with this exact shape:\n\n```text\nComplete replacement Exit: \u003Cthe full replacement field, including its command>\n```\n\nEach clause names one of `Goal`, `Entry`, `Exit`, `Files`, `Tests` or\n`Disciplines`, occurs once, and restates that entire field. More than one\nchanged field is written as more than one `Complete replacement \u003CField>:`\nclause. The other three amendment fields remain `Why`, `Steps touched` and\n`Still holding`, in that order. Replacement prose is still an operator claim:\nthe mechanical check establishes its shape and source bytes, not that the new\ncriterion is correct or its command will pass.\n\nFor machine-checked assignment authority, `What changed` is the first nonblank,\nunfenced record after its amendment heading. Its first replacement header\nbegins that value. Each later header begins a column-zero paragraph after a\nblank separator. Fenced and inline-code examples are not headers. A\nreplacement-like fragment at any other location refuses as ambiguous.\n\nKnown-failure assignment records outside every structurally valid complete\nreplacement `Exit` clause remain ordinary active authority and stay outside\ngeneration comparison. Records inside those clauses form source-ordered\ngenerations. Each generation first passes exact-line, assignment-only-block,\nunique-id, and real-Step checks. Leading empty generations are allowed. The\nfirst non-empty id-to-Step map locks assignment authority. Every later Exit\ngeneration, including an empty one, must equal that map; pair order does not\nmatter. The lock cannot reset or fall back. Empty, partial, extra, or reassigned\npost-lock generations refuse. A later replacement of another field creates no\ngeneration. The final Exit generation is combined with the ordinary records,\nwhile matching earlier generations remain readable history. Step headings come\nonly from the baseline before the first real amendment. Malformed amendment or\nreplacement boundaries, repeated clauses, and post-amendment Step headings\nrefuse before any generation can suppress a record.\n\nThe shared amendment scanner checks every real study or runbook amendment's\ncalendar date, four ordered non-empty fields, and final-section placement.\nFenced examples are not amendments. Runbook mode additionally checks the\ncomplete-replacement syntax described above. The runbook checker treats the\nfirst real `### Amendment -- YYYY-MM-DD`\nheading as the end of the last baseline step, so amendment fields cannot answer\nfor a missing step field. Fiat owns exact-prefix continuity,\nstep topology, touched-step verdicts, receipts, recovery and the current-study\nbinding; Protasis does not duplicate those controller gates.\n\n## Rationalisations\n\n- \"This is simple, it needs no spec.\" Simple topics need short specs, not none.\n  Two lines and a checkable criterion is a spec.\n- \"I will write the study afterwards.\" Then it is documentation. The value here\n  is forcing clarity before the code exists.\n- \"A spec slows us down.\" Fifteen minutes of study against hours of rework in\n  the audit loop, which is the phase that already dominates the clock.\n- \"Requirements will change anyway.\" Which is why the study is edited rather\n  than abandoned.\n- \"The user knows what they want.\" Every clear request carries implicit\n  assumptions. This phase exists to surface them.\n- \"It is one feature, splitting it is overhead.\" If its criteria cluster into\n  separately verifiable groups, every later step has to reason over the whole\n  contract. A four-row table is cheaper.\n- \"I will decompose while planning.\" Planning slices steps inside a study.\n  Module boundaries have to be settled before the study is written, not after.\n\n## Red flags\n\n- Code before any written requirement.\n- Asking whether to start building before \"done\" has been defined.\n- A step in flight that appears in no runbook.\n- A design decision made and not written down.\n- One study whose criteria span capabilities that could ship separately.\n- Build order settled implicitly during implementation.\n\n## Before the runbook is receipted\n\nReport the count, then name every failure. A set reported as passed without the\ncount is not a report.\n\n- [ ] The study answers all twelve items.\n- [ ] Items 8 through 12 each carry an answer or a stated none with its reason.\n- [ ] The last two merged pull requests touching the target were read, and\n      anything they carried forward is answered, refused by name, or stated as\n      still open. Say so plainly where there were none to read.\n- [ ] Every discovered audit source remains authoritative. The study must name\n      every in-scope source, which synopsis or source was actually read, and\n      the evidence for that choice. Do not claim the source was read when only\n      its synopsis was read.\n- [ ] A verified synopsis was used only after\n      `python3 plugins\u002Fhexaemeron\u002Fskills\u002Ffiat\u002Fscripts\u002Faudit_synopsis.py --check\n      \u003Ctarget-root>` ran from the target root and the whole-set currency check\n      exits zero. `**\u002Faudit\u002FAUDIT.md` maps to its sibling\n      `AUDIT_SYNOPSIS.md`. A direct child `audit\u002Frounds\u002F\u003Crun>.md` maps to\n      `audit\u002Frounds\u002F\u003Crun>.synopsis.md`. The root pair covers only the root\n      source.\n- [ ] Missing, stale, unsupported, or unavailable view means: read the\n      authoritative source directly and record its source path and the reason.\n      If the whole-set check failed, every in-scope source was read directly;\n      if the source was also unavailable, readiness stopped on that evidence\n      gap.\n- [ ] The chosen read mode retains every finding id and status, `Covered`,\n      `Not checked`, `Elenchus verdict`, and `Leads not pursued`;\n      `[missing legacy field: ...]` remains unknown.\n- [ ] No discipline core is restated where a citation belongs.\n- [ ] Assumptions are on the page and were confirmed or corrected.\n- [ ] Every success criterion names a command, a test or a demo path.\n- [ ] The final demonstration identifies its controller, source command and\n      bounded positive and negative observations without claiming sufficiency.\n- [ ] The design record covers correctness, time, space, compatibility and\n      recovery with one exact result per candidate and criterion.\n- [ ] Every resolved result binds a zero-exit report; every pending result names\n      its resolver, future report and exact blocking transition.\n- [ ] The design-lock check exits zero, and the chosen candidate is on the\n      mechanically computed frontier under its declared rule.\n- [ ] The runbook's design-lock block matches the record digest and selected\n      candidate before Step 1.\n- [ ] When known failures exist, the inventory checker exits zero against the\n      independent expected-id set, exact runbook, and current source views.\n- [ ] Always, ask-first and never each carry concrete entries.\n- [ ] Each step carries goal, entry, exit, files, tests and disciplines.\n- [ ] Every discipline a step names carries the reason it applies.\n- [ ] No exit rests on anything but a command.\n- [ ] Step 1 scaffolds and the last step demonstrates.\n- [ ] Steps are in dependency order.\n- [ ] If the topic was decomposed, every step traces to a module id.\n\n## Hand back\n\nLead with the state: ready to build, or blocked on a named gap. Then give the\ncount of checks passed out of the total, and name each one that failed.\n\nKeep three things apart. What the study establishes, what it assumes, and what\ncould not be settled. An assumption that changes the build order or the chosen\ndesign gets said out loud, not buried in a section.\n\nEnd with one action, and make it something the reader can do in a couple of\nminutes: confirm an assumption, answer one question, or approve the runbook.\nName the open question rather than closing it with a guess. Corrected here, an\nassumption costs a sentence. Found in the audit loop, it costs a step.\n\n## Promise Machine contract\n\n### protasis-known-failure-inventory\n\n- Promise: A clean inventory result establishes that one closed `protasis-known-failure-inventory\u002Fv1` object matches an independent exact id set, current source and synopsis bytes, portable guard paths, structurally bound commands, admitted reports, and real consuming runbook steps, or carries one digest-bound no-known-findings claim.\n- Evidence: The exact study and runbook bytes, caller or effective runbook expected ids, source and synopsis paths and SHA-256 values, finding fields and assignments, parsed argv, path and count caps, locked replacement-Exit map, stable K000 to K012 result, and focused hostile cases.\n- Evidence classes: checked\n- Boundary: This result establishes bounded structure and byte currency. A `source_ref` prefix selects a checked source view, but its locator text is opaque and does not prove that a cited line, range, or statement supports the failure. The result does not discover unknown failures, judge Surveyor's completeness, execute a guard, classify Elenchus output, write a receipt, or authorise a product edit by itself.\n- Authorises: Carrying the exact checked inventory into runbook derivation and the later Fiat inoculation boundary without inferring an omitted entry or an empty success.\n- Consequence: 1\n- Refuses: An absent, duplicated, open, malformed, non-finite, oversized, stale, unsafe, unassigned, command-drifted, id-mismatched, bracket-bearing runbook, replacement-ambiguous, post-lock generation-drifted, or unbound empty inventory.\n- Recovery: Repair the named K000 to K012 cause in the study, runbook, expected-id set, or source view, then rerun the complete read-only check.\n- Exceptions: none\n\n### protasis-study-readiness\n\n- Promise: A study accepted by Protasis states the problem, assumptions, current state, boundaries, failure and recovery model, affected versions and success criteria, and locks one candidate through a complete checked `protasis-design-evidence\u002Fv1` matrix without predicting evidence that is not yet available.\n- Evidence: The exact study, source inventory, explicit unknowns and exclusions, answered discipline questions, design record and report digests, mechanical frontier result, named future resolvers and completed study checklist.\n- Evidence classes: checked, inferred, recorded\n- Boundary: Readiness establishes the closed record, the reports checked at design lock and the declared stop points for later evidence. It does not establish that the selected design is correct, that pending conformance will pass, that implementation exists, or that later receipts are true.\n- Authorises: Derivation of a discrete runbook from the accepted study without silently adding a new design decision.\n- Consequence: 1\n- Refuses: Hidden assumptions, an unstated boundary, solution-first prose, an incomplete candidate matrix, missing required concern, unresolved selection evidence, a selected dominated candidate, unsupported tie-break, untestable success language or a topic still containing several independent deliveries.\n- Recovery: Name the missing question or decision, gather the required evidence, amend the study and repeat the complete Protasis review.\n- Exceptions: none\n\n### protasis-runbook-readiness\n\n- Promise: A runbook accepted by Protasis binds the exact selected design and decomposes the study into ordered steps whose entry, modules, exit commands, files, tests and discipline effects are discrete and provable, and any optional governed-skill version relation has one closed source without a competing concrete target token. When the study carries a known-failure inventory, Fiat must join this result to the clean `protasis-known-failure-inventory` result before product work.\n- Evidence: The accepted study, design-evidence digest and selected candidate, exact runbook and matching design-lock block, `protasis.py` structural result, per-step commands and files, dependency order, optional version-relations block, version boundary, completed pre-receipt checklist and, when present, the clean known-failure inventory result.\n- Evidence classes: checked, inferred, recorded\n- Boundary: Runbook readiness establishes buildable specification content, the design lock, and the lexical shape of a declared version relation. It does not establish correct implementation, later conformance evidence, command success, relation suitability, a selected version or integration base, audit closure or delivery completion. When an inventory is present, runbook readiness alone never authorises a product edit.\n- Authorises: Entering controller-governed runbook execution at the first step and, when an inventory is present, its required inoculation gate.\n- Consequence: 1\n- Refuses: A missing or mismatched design lock, a step with no executable exit, mixed independent outcomes, missing affected files or tests, forward references to an undecided design, receipt language with no evidence command, an applicable inventory without a clean joined result, or a malformed, ambiguous or concretely contradicted version relation.\n- Recovery: Split or reorder the failing step, supply its exact evidence and rerun both the mechanical check and the full content review.\n- Exceptions: none\n\n### protasis-gate-command-validation\n\n- Promise: A successful command check establishes that each effective declared invocation fits the registered CLI interface, with exact source and report bindings, or, while a runner that Step 1 creates has no binding, passes every check except that runner's parser interface without reading it; a registered module whose AST digest left the current pin fits only when the caller's starting-commit bindings carry that path's AST and source digest pair and both equal what was read; no declared command is executed and no target module imported.\n- Evidence: Captured runbook SHA-256 and source root, raw command text and UTF-8 offsets, command digests, superseded-source records, original and substituted argv, full CLI and adapter digests, deferred records naming the creating step, the caller's binding and phase record, the caller's starting-commit bindings when supplied, reviewed adapter, supplied starting-commit adapter or exact runner source-pair compatibility when applicable, declaration digests, interface results, current-root report resolution and `operation_ran:false` under `protasis-gate-commands\u002Fv1`.\n- Evidence classes: checked, recorded\n- Boundary: Interface validity establishes supported argument and source shape. It does not establish command execution, executable behavior, test success, report truth, audit judgement or atomic protection against later source changes. Historical absolute argv records nonexecuted derivation and grants no authority to execute at an old root; checkpoint relocation belongs to Fiat. An `interface-deferred` result establishes no parser fit, execution or success; a runner binding pins that file only, not the modules it imports, and the adapter takes the run's phase and binding from its caller. The adapter derives no starting-commit binding and reads no Git: it admits only the pair its caller supplies, and its result records no provenance for that admission. This Consequence 1 operation has no native runtime binding. Its actual interface result retains operation_ran:false because the declared commands were not executed.\n- Authorises: Supplying the exact interface result to Fiat as command-validation evidence for its separately governed runbook or amendment receipt.\n- Consequence: 1\n- Refuses: Unregistered commands, malformed or misplaced local registrations, a deferred row naming another step, added after Step 1 starts or repeated after binding, a present or unsafe unbound deferred path when absence is required, an inconsistent binding or phase record, registry overrides, unsupported parser declarations or converters, private worker arguments, unsupported shell evaluation, invalid literal or loop argv, malformed command fences, missing or unsafe report declarations, CLI source drift outside the reviewed timestamp pair, a starting-commit pair that does not equal the module read or a bindings value outside its closed shape, unknown adapter digests, bound violations or replay mismatch.\n- Recovery: Preserve the original command and refusal, correct the declared interface or provide a reviewed adapter, then repeat the check. For a deferred row, remove what occupies its path before capture, or supply the recorded binding and phase. For a receipted run, use Fiat's permitted append-only amendment when the source or bindings changed. A released adapter listed in the [command reference](references\u002Fgate-commands.md) may replay unchanged evidence without amendment, as does the adapter at the run's starting commit when the caller supplies its digest. That reference also permits one exact runner timestamp transition when every invocation uses the reviewed source pair and every other receipt field matches; earlier receipts remain intact.\n- Exceptions: none\n","\u003Cp align=\"center\">\n  \u003Cimg src=\"..\u002F..\u002Fassets\u002Fcharacters\u002Fprotasis.png\" width=\"1200\">\n\u003C\u002Fp>\n\n# Protasis\n\nFrom *protasis*, the proposition laid down before the argument runs. Nothing is\nbuilt from a topic; things are built from a proposition about a topic.\n\n## Where this sits\n\nProtasis owns the content contract for the `study` and `runbook` phases: what\nthose two documents must answer before implementation is allowed to start. It\nowns no state, writes no receipt and gates nothing itself.\n\nSurveyor may write the study packet, while Fiat keeps the artefact paths and\nreceipt commands. Mason and Warden later receive the exact source-bound\nrunbook step. Phylax, Ephoros, Metron, Elenchus, and Hypomnema answer the five\ndiscipline questions Protasis requires; Protasis cites their contracts rather\nthan copying them. A decision made after the study belongs to Hypomnema's\nrecording rules.\n\nSynkrisis proposes a next owner from a validated series of run observations.\nThat finding still needs a scoped proposition before it can become a study,\nrunbook, or framework change.\n\nIts version, held frontier, next job, and maturity state live in\n[EVOLUTION.md](EVOLUTION.md).\n\n**Current state.** Protasis checks the fixed mechanical shape of study and\nrunbook prose, one source-bound known-failure inventory, and one closed\ncandidate-by-criterion design record at the transition where each item of\nevidence becomes due. The optional command check validates literal arguments\nand finite per-file loops against registered source interfaces without\nexecuting the declared commands. Fiat's Step 5 demonstration separately\nrecords the successor controller's executed result and bounded refusal cases;\nthat report does not enlarge Protasis's declaration or claim criterion\nsufficiency.\n\n## Refuse these six\n\n1. No study, no runbook. A runbook derived from conversation instead of a\n   written study is not a runbook.\n2. No runbook, no implementation. A step that does not exist in the runbook\n   does not get built, however small it looks.\n3. No criterion, no success. A study whose success condition cannot be checked\n   by a command or a named demo path is unfinished.\n4. No stated assumption, no spec. Assumptions go on the page before the\n   content they support.\n5. No checked design record, no design lock. A preferred construction in prose\n   does not authorise implementation.\n6. No closed known-failure inventory, no production edit. Missing, stale, or\n   unassigned entries cannot be repaired by treating an empty list as clean.\n\nReport a refusal in three parts: what is missing, where you looked, and the one\naction that clears it. Say plainly that the phase is blocked rather than\nin progress. None of the four is a suggestion to proceed carefully.\n\nIf the gap is an ambiguity rather than an absence, ask one literal question\ninstead of picking a reading.\n\n## Assumptions go first\n\nList what you are assuming before writing any spec content, and say plainly\nthat you will proceed on them unless corrected.\n\n```\nAssuming, unless corrected:\n1. Foundry, not Hardhat; the repo has foundry.toml and no hardhat.config.\n2. Solidity 0.8.x with checked arithmetic; unchecked blocks are opt-in per site.\n3. The exact interpreter in the repository's `.python-version`, with stdlib unittest.\n4. An archive RPC is available for the capture step; without one, step 3 changes.\n```\n\nAn unstated assumption is the failure this phase exists to catch. On the page,\na wrong one costs a sentence. Buried in step 4, it costs the step.\n\n## Vague requirement, testable criterion\n\nRestate the request as conditions a command can check, then confirm the\nrestatement before building on it.\n\n```\nRequirement: \"make the harvester faster\"\n\nRestated:\n- A full Ethereum USDC interval harvest finishes inside 20 minutes on a warm cache.\n- The digest of the produced release is byte-identical to the current one.\n- Peak resident memory stays under 2 GB.\nMeasured by metron, before and after. Are these the right targets?\n```\n\n\"Faster\", \"cleaner\", \"more robust\" and \"production-ready\" are not criteria.\nNeither is a criterion that can only be checked by asking someone whether they\nare happy.\n\n## What a study must answer\n\n1. **Problem statement.** What is being built, for whom, and what a working\n   prototype means here. Name the demo path or the check that proves it.\n2. **Prior art.** What exists already, in this repo, in the organisation's\n   other repos, and outside. Name files, packages and standards by identifier.\n   Where the topic touches something that has shipped before, read the last two\n   merged pull requests that changed it before writing anything else. A run\n   that could not finish something records it in the body of the last pull\n   request it lands, so that is where the unfinished work of the previous run\n   is written down. Carry each item forward as content here, as a stated\n   non-goal, or as a named reason it stays open. Read the audit records of\n   every in-scope skill the same way, before design options are drawn. Every\n   discovered audit source remains authoritative. A verified synopsis is its\n   normal reading view only after\n   `python3 plugins\u002Fhexaemeron\u002Fskills\u002Ffiat\u002Fscripts\u002Faudit_synopsis.py --check\n   \u003Ctarget-root>` runs from the target root and the whole-set currency check\n   exits zero. `**\u002Faudit\u002FAUDIT.md` maps to its sibling `AUDIT_SYNOPSIS.md`. A\n   direct child `audit\u002Frounds\u002F\u003Crun>.md` maps to\n   `audit\u002Frounds\u002F\u003Crun>.synopsis.md`. The root pair covers only the root source;\n   it does not cover per-run or plugin records.\n\n   Missing, stale, unsupported, or unavailable view means: read the\n   authoritative source directly and record its source path and the reason.\n   If the whole-set check fails, do that for every in-scope source rather than\n   guessing which view remains current. A target without the renderer or a\n   committed synopsis can still be studied when its source is available. If\n   neither view nor source is available, readiness stops on that evidence gap.\n\n   Keep every finding id and status, `Covered`, `Not checked`, `Elenchus\n   verdict`, and `Leads not pursued`; `[missing legacy field: ...]` remains\n   unknown. Then name every in-scope source, which synopsis or source was\n   actually read, and the evidence for that choice. Do not claim the source\n   was read when only its synopsis was read. The run's audit file, at its\n   `config audit.log_path`, holds what each round found, what was fixed, and\n   the leads accepted with the reason nobody pursued them. Fiat derives one\n   such file per run under `audit\u002Frounds\u002F`, so the records of several runs are\n   several files; a target that kept a shared log before that change has the\n   rest of its history in `audit\u002FAUDIT.md`. A study that cannot find the real\n   reason for a decision supplies a plausible one, and the plausible one then\n   governs the build: a rejected option gets rejected for the wrong cause, an\n   accepted risk gets quietly reopened, and the round that already judged the\n   question reads afterwards as though it never happened. Found in the study\n   it costs a sentence; found in step four it costs the step.\n3. **Constraints and non-goals.** The starting ref, toolchain and version pins,\n   what the user ruled out, what is deferred past the prototype.\n4. **Design options.** Two to four candidate constructions, each with the trade\n   it makes. The prose explains the candidates; the design-evidence record\n   below selects one from checked gates and comparative measurements.\n5. **Risk register seed.** What the audit loop should look hardest at. In\n   Solidity: trust boundaries, external calls, arithmetic, upgrade paths, key\n   custody. In Python: untrusted input, subprocess and filesystem handling,\n   secret material, partial writes, and what happens when a long run is killed\n   halfway. The concerns go in a fenced block the audit loop can enumerate,\n   info string `risk-register`, one concern per line as three pipe-separated\n   fields -- a kebab-case id stable within the study, the boundary the concern\n   sits at, and what the audit loop checks:\n\n   ```risk-register\n   subprocess-input | the argv of the spawned compiler | inputs are pinned and no shell is used\n   partial-write | the release directory during a long harvest | a killed run leaves no half-written file that verifies\n   ```\n\n   A round then logs each id as reviewed or not applicable, so the look is an\n   enumerable obligation rather than a judgement call nobody can verify\n   afterwards. Prose around the block carries what a line cannot, and the ids\n   are how a round cites it.\n6. **Glossary seeds.** Terms the runbook and implementation will reuse, one\n   line each.\n7. **Sources.** Enough of a pointer to find each one again.\n\nThe five that follow are the disciplines the build will be held to. Answering\nthem here costs a sentence each; leaving them to the audit loop costs a step.\n\n8. **Signals, and the questions behind them.** Two to four questions someone\n   will ask at three in the morning once this runs unattended, and which steps\n   emit the signals that answer them.\n   [ephoros](..\u002Fephoros\u002FSKILL.md) owns what a signal must carry.\n9. **Boundaries, per capability.** Each boundary this opens, what is worth\n   taking at it, and the control that closes it. This feeds item 5 rather than\n   replacing it. [phylax](..\u002Fphylax\u002FSKILL.md) owns the boundary list and the\n   controls.\n10. **The budget, or its absence.** Any performance budget this must hold, with\n    the exact command that measures it.\n    [metron](..\u002Fmetron\u002FSKILL.md) owns what a budget carries and how it is\n    checked.\n11. **The fail-closed posture.** What stops the run, and the guard-test\n    convention a fix will follow.\n    [elenchus](..\u002Felenchus\u002FSKILL.md) owns the triage order and the guard rule.\n12. **Decisions and their homes.** The decisions expected to be expensive to\n    reverse, and the file each record will live in.\n    [hypomnema](..\u002Fhypomnema\u002FSKILL.md) owns which decisions earn a record and\n    where each one lives.\n\nFor items 8 through 12, \"none, and here is why\" is a complete answer. A lint\ninvoked from a terminal has no on-call question; a step that reads two files\nhas no budget. What is not a complete answer is silence, because silence cannot\nbe told apart from not having looked.\n\nCite those five contracts, never restate them. Each owns its own rules and each\nevolves on its own ledger, so a copy here is stale the moment it is written.\n\n### Known-failure inventory\n\nWhen audit history names failures that implementation must guard before product\nwork starts, the study carries exactly one fenced\n`known-failure-inventory` JSON object under schema\n`protasis-known-failure-inventory\u002Fv1`. Each `source_views` entry names a\nportable synopsis path and the SHA-256 of both that view and the authoritative\nsource named by its header. Each finding has one unique id, a source reference\nwhose prefix names a checked view, bounded non-empty locator text, a failure\nstatement, closed guard-path set, structural test command, report format and\npath, exact expected verdict `guarded`, fixed-tree command, and one consuming\nrunbook step. The locator text is recorded but not interpreted as a line or\nclaim proof. The inventory fence is column-zero and blank-line isolated.\n\nThe test command is a bounded six-argument invocation: `python3`, one declared\nPython reporter path, `--case`, the finding id, `--report`, and exactly one\nargv equal to `{report}`. Its fixed-tree counterpart differs only in the report\nargument. Each report and derived green path is unique under `.elenchus\u002F`.\nFilesystem-backed uniqueness uses Unicode NFC plus case folding for portable\nalias detection.\nAdmitted report formats are `unittest-json-v1`, `forge-junit-v1`, and\n`node-test-json-v1`. The caller supplies an independent expected-id set, or the\nchecker uses ids in visible assignment records; each record names a real\npositive, column-zero runbook Step, and the inventory's consuming Steps must\nmatch. Each assignment belongs to a contiguous nonblank block containing\nassignment lines only, so adjacent prose, image labels, link titles, and\nreference definitions cannot supply one. The only authoritative assignment is\na visible full line of the form\n``Known-failure assignment: `kf-453-N` -> Step S`` outside fences. Raw HTML,\ncomments, and malformed HTML-like bytes outside one-line code spans refuse at\nthe runbook boundary. Every square-bracket byte outside a complete one-line\ncode span or fenced block also refuses there. The study keeps its separate\nsurface rules. Its inventory discovery tolerates unmatched inline-backtick runs\nin ordinary prose, but never uses them to mask bytes. It processes or refuses\neach column-zero fence candidate before that tolerance. Raw HTML, image openers,\nindented fences, bad fence information, unclosed fences, and anything other\nthan one isolated inventory block still refuse. Fenced examples use column-zero\nfences. Nested or indented fence candidates refuse. At most 128 findings, 4,096\naggregate guard paths, 1,024 bytes per path, 4,096 bytes per source reference,\nfailure statement, or command, and 16 command arguments are checked. Paths are\nroot-relative portable paths without shell metacharacters or Unicode\nwhitespace. The checker opens source paths through no-follow, nonblocking\ndirectory handles. It rereads the study, runbook, views, and sources together\nbefore returning success.\n\nA non-empty finding set requires `no_known_findings: null`. An empty set needs\na non-null claim that repeats every checked source id and both digests, names\none real consuming step, and carries the exact Surveyor assertion\n`no-known-findings`. An empty array alone never establishes readiness. The\npublic `load_checked_inventory(study, runbook, repository, expected_ids=None)`\noperation returns one frozen `InventoryLoadResult`. Its `status` is exactly\n`absent`, `refused`, or `clean`. `absent` carries neither a capture nor\nfindings and is available only when the stable study has no inventory marker\nand the stable runbook has no assignment marker. Any attempted, partial, or\nmalformed marker stays on the existing K000 through K012 refusal path.\n`refused` carries a non-empty tuple of those findings and no capture. `clean`\ncarries no findings and one JSON-compatible capture with exactly `schema`,\n`study_sha256`, `runbook_sha256`, `inventory_sha256`, `source_views`,\n`findings`, `no_known_findings`, and `assignments`.\n\nThe clean schema is `protasis-known-failure-inventory-capture\u002Fv1`. Study and\nrunbook digests cover their exact stable bytes. The inventory digest covers\nthe accepted parsed object encoded by Fiat's canonical JSON form: sorted\nobject keys and compact separators. Source views, findings, and the no-known\nclaim are the complete parsed inventory values. Each assignment has exactly\n`finding_id` and `step`, ordered by Step and then finding id. The operation\nrevalidates every input before returning the capture. The compatibility\n`check` operation and command-line checker are projections of this result;\nthey never reread or reparse a clean inventory. The checker writes no\ncontroller state or ledger. It establishes structure and current bytes, not\nthat Surveyor found every failure or that any guard has run.\n\n### Design evidence and progressive gates\n\nThe study writes one closed record to `.hexaemeron\u002Fdesign-evidence.json` under\nschema `protasis-design-evidence\u002Fv1`. The record is the selection interface;\nprose is not. It contains two to four candidate ids, one to 32 criteria, the\nexact candidate-by-criterion result matrix, and one selection.\n\nEvery criterion names one of five concerns: `correctness`, `time`, `space`,\n`compatibility`, or `recovery`. The complete set must be covered. A criterion\nalso names its owner and one of two forms:\n\n- A hard gate has stage `selection` or `conformance`, compares a typed value by\n  `equals`, `at-most`, or `at-least`, and blocks `design-lock`, `step:N`, or\n  `integration`.\n- A comparative metric has stage `selection`, minimises or maximises `bytes`,\n  `count`, `milliseconds`, or `ratio`, and blocks `design-lock`.\n\nThe record uses these exact object forms; this excerpt is not a complete\nmatrix:\n\n```json\n{\n  \"schema\": \"protasis-design-evidence\u002Fv1\",\n  \"candidates\": [{\"id\": \"streaming\", \"summary\": \"Process one bounded window.\"}],\n  \"criteria\": [{\n    \"id\": \"peak-space\", \"concern\": \"space\", \"kind\": \"metric\",\n    \"stage\": \"selection\", \"owner\": \"metron\", \"unit\": \"bytes\",\n    \"comparator\": \"minimise\", \"threshold\": null, \"blocks\": \"design-lock\"\n  }],\n  \"results\": [{\n    \"candidate\": \"streaming\", \"criterion\": \"peak-space\", \"state\": \"pass\",\n    \"report\": {\"path\": \"reports\u002Fstreaming-peak-space.json\", \"sha256\": \"\u003C64 lowercase hex>\"}\n  }, {\n    \"candidate\": \"streaming\", \"criterion\": \"restart-safe\", \"state\": \"pending\",\n    \"resolver\": \"python3 tests\u002Fprove_restart.py\",\n    \"report\": \"reports\u002Fstreaming-restart-safe.json\", \"blocks\": \"step:2\"\n  }],\n  \"selection\": {\"candidate\": \"streaming\", \"rule\": \"unique-frontier\", \"policy_ref\": null}\n}\n```\n\nThe abbreviated example shows both result forms; an accepted record still has\ntwo to four candidates, all five concerns, and every matrix cell.\n\nA resolved matrix cell is `pass` or `fail` and names a report path plus its\nSHA-256. The report is one closed `protasis-design-report\u002Fv1` object containing\n`candidate`, `criterion`, typed `value`, `unit`, `command`, and integer `exit`;\nexit must be zero. Paths are relative to the record directory and may not cross\na symlink. A pending cell instead names the exact resolver command, future\nreport path, and stop point it blocks. Selection evidence may be pending while\nthe record is drafted, but not when the design locks. Conformance evidence may\nremain pending until its named step or integration transition. An unknown is\ntherefore a scheduled refusal, not a guessed score.\n\nSelection is closed and mechanical. A failed selection gate removes a\ncandidate. The checker computes the non-dominated frontier from the selection\nmetrics. `unique-frontier` requires one survivor. `exact-tie-simplicity`\napplies only when every checked comparative value is exactly equal across at\nleast two survivors. `user-policy` requires at least two survivors and a\nbounded `policy_ref`. Simplicity has no other tie-breaking authority.\n\nRun both checks before the study is receipted:\n\n```bash\npython3 \"$PLUGIN_ROOT\u002Fskills\u002Fprotasis\u002Fscripts\u002Fprotasis.py\" --study .hexaemeron\u002Fstudy.md\npython3 \"$PLUGIN_ROOT\u002Fskills\u002Fprotasis\u002Fscripts\u002Fdesign_evidence.py\" \\\n  .hexaemeron\u002Fdesign-evidence.json --transition design-lock\n```\n\nThe runbook binds the accepted bytes before Step 1:\n\n```design-lock\nschema | protasis-design-evidence\u002Fv1\nsha256 | \u003Csha256 of .hexaemeron\u002Fdesign-evidence.json>\ncandidate | \u003Cselected candidate id>\n```\n\nP007 checks the block's closed shape and position. Fiat checks that its values\nmatch the study receipt, then runs the design checker at `step:N` immediately\nbefore opening that step and at `integration` before admitting the completed\nstack. Each transition records the reports consumed there. Verification\nreplays every recorded transition against the unchanged record and report\nbytes. A run created before the contract has no `contracts.design_evidence`\nmarker and remains on the legacy path; no evidence is fabricated for it.\n\nA section reading \"TBD\" is a section to fill or cut. Where the request is\nambiguous, record the reading you chose and the reason for it. Never resolve an\nambiguity silently.\n\n## What a runbook step must contain\n\nA step is discrete: one pull request, one boundary. It is self-contained, so\nsomeone holding only the study, the runbook and the repo at that step's entry\nstate can finish it. Both ends are green; no step hands the next a broken tree.\nAnd it stays small enough to audit, because that phase dominates the clock.\n\n```markdown\n## Step N: \u003Ctitle>\n\n**Goal.** One sentence.\n**Entry.** The exact ref or state this step starts from.\n**Exit.** Deliverables, plus the command or test that proves them.\n**Files.** Paths created or changed.\n**Tests.** What gets written or extended, and the expected count if known.\n**Disciplines.** Which of the five apply to this step and why, or none with\nthe reason.\n```\n\nFor a Fiat step, `Tests` also owns the runner contract used when its audit\nclaims a fix. It names the exact Elenchus test command with one `{report}`\nargument, the report format, and the report file. Fiat sends the complete\nsource-bound step to Warden; Warden owns those three inputs and may not infer a\ncommand from `Files` or replace one with a nearby suite.\n\nThe Disciplines line carries the study's items 8 through 12 down to the step\nthat actually incurs them. Name each discipline and the reason in one clause:\n\n```text\n**Disciplines.** phylax: this step opens the ingestion path. ephoros: it runs\nunattended once deployed. metron: none, no performance claim. elenchus: none,\nno failure in hand. hypomnema: the storage format is expensive to reverse.\n```\n\nA step that names a discipline without a reason has not answered; a step that\nomits one has not been asked. Mason builds against the declared gates and\nwarden audits against them, so a gate named here is cheaper than the same gate\ndiscovered in round three.\n\nThree fixed points. Step 1 scaffolds: layout, toolchain pins, CI stub, licence,\nand committed copies of the study and runbook. The last step demonstrates, by\nrunning the demo path from the problem statement. Ordering is dependency order,\nand a step may assume every earlier step's exit state and nothing else.\n\nIf a step's exit cannot be proved by a command, it is not an exit. \"Reviewed\",\n\"working\" and \"integrated\" prove nothing on their own.\n\n### Version relations\n\nA runbook may carry one optional fenced `version-relations` block before Step\n1. Absence preserves the literal-only contract. A present block carries at\nmost 32 physical rows, with no blank row, in this closed shape:\n\n```version-relations\nprotasis | plugins\u002Fhexaemeron\u002Fskills\u002Fprotasis\u002FEVOLUTION.md | next-generation-after-integration-base\n```\n\nEach row is `skill id | EVOLUTION.md path | relation`. The skill id is\nkebab-case. The path is repository-relative, contains no empty, `.` or `..`\nsegment, backslash or control character, ends in `EVOLUTION.md`, and has the\nskill id as its parent directory. Ids and paths are unique. The sole admitted\nrelation is `next-generation-after-integration-base`; a partial target list is\nvalid because an omitted target may retain an intentional literal.\n\nA declared target has no concrete\n`\u003Cskill>-v\u003Cevolution>.\u003Cgeneration>.\u003Cepoch>` token outside the block. Prose,\nexamples, commands and amendments all count. P006\nchecks this shape and lexical identity. It does not open the ledger, decide\nwhether the relation suits the change, allocate a version, or establish what\nthe integration base will be.\n\n## When one topic is several\n\nMost topics are one capability and go straight to the study. Decompose first\nonly when a single request bundles capabilities that could ship and be verified\nseparately, or when one could be cut without rewriting the others.\n\nThe decomposition is a table and a build order, not a project plan:\n\n| Module id | Responsibility | Depends on |\n| --- | --- | --- |\n| capture | Fixed-block RPC capture, digest-keyed | none |\n| verify | Proof checks over a capture | capture |\n| replay | Local replay boundary, no fallback | capture |\n| fixture | Published fixture and its manifest | verify, replay |\n\nBuild order: capture, then verify and replay, then fixture.\n\nModule ids are kebab-case, chosen once, never renamed mid-topic. Dependencies\npoint one way; if two modules each need the other, they are one module. An\ninterface belongs in the spec of the module that provides it. Each module then\ngets its own study, and its modules become runbook steps in that order.\n\n## Boundaries the study must state\n\nThree tiers, each with concrete entries:\n\n- **Always.** Both test suites before a commit. The imprimatur lint on every\n  shipped document. A recorded measurement before any performance change.\n- **Ask first.** Adding a dependency. Changing a storage layout or a public\n  ABI. Touching CI. Widening a trust boundary. Rewriting a released digest.\n- **Never.** Commit key material or an RPC credential. Edit a vendored\n  directory. Delete a failing test to make a suite pass. Claim a command ran\n  when it did not.\n\n## The mechanical subset\n\nWhether the twelve items are present, and whether items 8 through 12 carry an\nanswer, is settled by a parser; so is the runbook step schema. Run the bundled\ncheck over each artefact and require exit 0:\n\n```bash\npython3 \"$PLUGIN_ROOT\u002Fskills\u002Fprotasis\u002Fscripts\u002Fprotasis.py\" --study \u003Cstudy>\npython3 \"$PLUGIN_ROOT\u002Fskills\u002Fprotasis\u002Fscripts\u002Fprotasis.py\" \u003Crunbook>\npython3 \"$PLUGIN_ROOT\u002Fskills\u002Fprotasis\u002Fscripts\u002Fknown_failure_inventory.py\" \u003Cstudy> \u003Crunbook> --repository \u003Croot> --expected-id \u003Cid> [--expected-id \u003Cid> ...]\npython3 \"$PLUGIN_ROOT\u002Fskills\u002Fprotasis\u002Fscripts\u002Fdesign_evidence.py\" \u003Crecord> --transition \u003Cdesign-lock|step:N|integration>\n```\n\nThe study mode reads items as `## N. Title` headings, 1 to 12, refuses\nsilence and a bare none on items 8 through 12, and reads item 5's\nrisk-register block against the shape above: S005 when no block names a\nconcern, S006 when a line does not split into the three pipe-separated\nfields, S007 when a field is malformed. S008 reports a real study amendment\nwhose heading is not a calendar date, whose four fields are missing,\nduplicated, reordered, unknown, or empty, or which is not final. A study with\nno amendment and amendment examples inside fences remain unchanged. The\nrunbook mode reads the step\nschema above, ends the last baseline step before a real amendment heading and\nreports P005 when a runbook amendment does not carry the dated four-field\nshape and complete replacement clauses below. It reports P006 when a present\nversion-relations block is open, misplaced, duplicated, oversized or malformed,\nor when a declared target is also pinned to a concrete token outside that\nblock. P007 checks a present design-lock block's position, closed rows, schema,\ndigest and selected-candidate shape. Codes P000 to P007, S000 to S008, and\nK000 to K012 are stable interfaces other tools cite. Deliberate exceptions\nstate a reason:\n`\u003C!-- protasis: allow \u003Cwhy> -->` on the heading line or the line above it.\nPresence and shape are all the parser settles; whether an answer is any good\nstays with the reviewer and the rest of this contract.\n\n## Check declared command interfaces\n\nUse `--gate-root \u003Ctarget-root>` with the runbook checker to validate the\neffective Exit and Elenchus commands against registered local CLI sources.\nUnsupported shell evaluation, argument shapes, report substitutions and source\nadapters refuse with P008. The separate\n[command reference](references\u002Fgate-commands.md) names the grammar, source\nbindings, report format, limits and replay rules.\n\nA target outside the Skills distribution may register up to 32 reviewed local\nPython interfaces in the runbook's closed `command-interfaces` block. Each row\npins the complete source SHA-256 and parser-builder name. The built-in registry\ncannot be overridden. A dated append-only amendment may replace the complete\nlocal registration set; an unregistered or changed source still refuses.\nRegistration binds the reviewed source and permits inert argument checking.\nIt grants no authority to execute the command.\n\nA row may put `step:1` in place of the digest for a runner that Step 1\ncreates. Until the caller supplies a binding, the adapter never reads or hashes\nthat path. Capture before Step 1 can require it absent, and each command naming\nit records `interface-deferred` after every check except the runner's parser\ninterface. With a binding the row is an ordinary pinned registration. The\nreference gives the placement rules, refusals and the phase record the caller\npasses.\n\nA caller may also pass `starting_bindings`, what it derived from a run's\nstarting commit: that commit's adapter digest and, for each registered module,\nthe AST digest that commit's adapter pinned with the module's complete source\ndigest. A registered module whose AST digest left the current pin is admitted\nonly when both digests equal its pair, and replay substitutes a receipt's\nadapter digest only when it equals the supplied one. Without the keyword every\nresult is unchanged. The adapter reads no Git and records no provenance; the\ncaller derives the bindings and owns that record. The reference names the\nclosed shape.\n\nThe result is `protasis-gate-commands\u002Fv1` with `operation_ran:false`. It binds\nraw commands, the captured source root, original and substituted argv, full CLI\nand adapter digests, and interface results. It establishes no execution, test success or audit verdict.\nFiat owns receipt storage, the new-init marker, legacy compatibility and the\nappend-only amendment that refreshes stale evidence.\n\n## The spec stays alive\n\nWhen a decision changes, change the study first and the code second. When scope\nmoves, say so on the page. Both documents are committed and reviewed like any\nother shipped artefact, and both go through the prose pass first.\n\nA mid-run change is an amendment, and an amendment appends rather than edits,\nso the run's earlier belief stays readable. It is a dated block at the end of\nthe study carrying four fields:\n\n```markdown\n### Amendment -- 2026-08-20\n\n**What changed.** The capture step reads the header from the fixture, not RPC.\n**Why.** The archive endpoint was withdrawn mid-run.\n**Steps touched.** Step 3's entry and step 4's files.\n**Still holding.** Step 3: entry broken; exit holds. Step 4: entry holds; exit\nholds. Step 5: entry holds; exit holds. Step 6: entry holds; exit holds.\n```\n\nEvery unbuilt step gets a verdict in the last field: its entry and exit hold,\nor they do not. Do not proceed past a step whose entry the amendment broke;\nreport it as blocked in the contract's three parts and re-derive the runbook\nfrom the amended study, or re-specify the step, before building on it. The\ndecision that forced the change is recorded where\n[hypomnema](..\u002Fhypomnema\u002FSKILL.md)'s rules put it, and the amendment points at\nthat record rather than restating it.\n\nA receipted runbook is amended by appending the same dated block, never by\nediting or repeating a numbered step. Its `What changed` value is one or more\nclauses with this exact shape:\n\n```text\nComplete replacement Exit: \u003Cthe full replacement field, including its command>\n```\n\nEach clause names one of `Goal`, `Entry`, `Exit`, `Files`, `Tests` or\n`Disciplines`, occurs once, and restates that entire field. More than one\nchanged field is written as more than one `Complete replacement \u003CField>:`\nclause. The other three amendment fields remain `Why`, `Steps touched` and\n`Still holding`, in that order. Replacement prose is still an operator claim:\nthe mechanical check establishes its shape and source bytes, not that the new\ncriterion is correct or its command will pass.\n\nFor machine-checked assignment authority, `What changed` is the first nonblank,\nunfenced record after its amendment heading. Its first replacement header\nbegins that value. Each later header begins a column-zero paragraph after a\nblank separator. Fenced and inline-code examples are not headers. A\nreplacement-like fragment at any other location refuses as ambiguous.\n\nKnown-failure assignment records outside every structurally valid complete\nreplacement `Exit` clause remain ordinary active authority and stay outside\ngeneration comparison. Records inside those clauses form source-ordered\ngenerations. Each generation first passes exact-line, assignment-only-block,\nunique-id, and real-Step checks. Leading empty generations are allowed. The\nfirst non-empty id-to-Step map locks assignment authority. Every later Exit\ngeneration, including an empty one, must equal that map; pair order does not\nmatter. The lock cannot reset or fall back. Empty, partial, extra, or reassigned\npost-lock generations refuse. A later replacement of another field creates no\ngeneration. The final Exit generation is combined with the ordinary records,\nwhile matching earlier generations remain readable history. Step headings come\nonly from the baseline before the first real amendment. Malformed amendment or\nreplacement boundaries, repeated clauses, and post-amendment Step headings\nrefuse before any generation can suppress a record.\n\nThe shared amendment scanner checks every real study or runbook amendment's\ncalendar date, four ordered non-empty fields, and final-section placement.\nFenced examples are not amendments. Runbook mode additionally checks the\ncomplete-replacement syntax described above. The runbook checker treats the\nfirst real `### Amendment -- YYYY-MM-DD`\nheading as the end of the last baseline step, so amendment fields cannot answer\nfor a missing step field. Fiat owns exact-prefix continuity,\nstep topology, touched-step verdicts, receipts, recovery and the current-study\nbinding; Protasis does not duplicate those controller gates.\n\n## Rationalisations\n\n- \"This is simple, it needs no spec.\" Simple topics need short specs, not none.\n  Two lines and a checkable criterion is a spec.\n- \"I will write the study afterwards.\" Then it is documentation. The value here\n  is forcing clarity before the code exists.\n- \"A spec slows us down.\" Fifteen minutes of study against hours of rework in\n  the audit loop, which is the phase that already dominates the clock.\n- \"Requirements will change anyway.\" Which is why the study is edited rather\n  than abandoned.\n- \"The user knows what they want.\" Every clear request carries implicit\n  assumptions. This phase exists to surface them.\n- \"It is one feature, splitting it is overhead.\" If its criteria cluster into\n  separately verifiable groups, every later step has to reason over the whole\n  contract. A four-row table is cheaper.\n- \"I will decompose while planning.\" Planning slices steps inside a study.\n  Module boundaries have to be settled before the study is written, not after.\n\n## Red flags\n\n- Code before any written requirement.\n- Asking whether to start building before \"done\" has been defined.\n- A step in flight that appears in no runbook.\n- A design decision made and not written down.\n- One study whose criteria span capabilities that could ship separately.\n- Build order settled implicitly during implementation.\n\n## Before the runbook is receipted\n\nReport the count, then name every failure. A set reported as passed without the\ncount is not a report.\n\n- [ ] The study answers all twelve items.\n- [ ] Items 8 through 12 each carry an answer or a stated none with its reason.\n- [ ] The last two merged pull requests touching the target were read, and\n      anything they carried forward is answered, refused by name, or stated as\n      still open. Say so plainly where there were none to read.\n- [ ] Every discovered audit source remains authoritative. The study must name\n      every in-scope source, which synopsis or source was actually read, and\n      the evidence for that choice. Do not claim the source was read when only\n      its synopsis was read.\n- [ ] A verified synopsis was used only after\n      `python3 plugins\u002Fhexaemeron\u002Fskills\u002Ffiat\u002Fscripts\u002Faudit_synopsis.py --check\n      \u003Ctarget-root>` ran from the target root and the whole-set currency check\n      exits zero. `**\u002Faudit\u002FAUDIT.md` maps to its sibling\n      `AUDIT_SYNOPSIS.md`. A direct child `audit\u002Frounds\u002F\u003Crun>.md` maps to\n      `audit\u002Frounds\u002F\u003Crun>.synopsis.md`. The root pair covers only the root\n      source.\n- [ ] Missing, stale, unsupported, or unavailable view means: read the\n      authoritative source directly and record its source path and the reason.\n      If the whole-set check failed, every in-scope source was read directly;\n      if the source was also unavailable, readiness stopped on that evidence\n      gap.\n- [ ] The chosen read mode retains every finding id and status, `Covered`,\n      `Not checked`, `Elenchus verdict`, and `Leads not pursued`;\n      `[missing legacy field: ...]` remains unknown.\n- [ ] No discipline core is restated where a citation belongs.\n- [ ] Assumptions are on the page and were confirmed or corrected.\n- [ ] Every success criterion names a command, a test or a demo path.\n- [ ] The final demonstration identifies its controller, source command and\n      bounded positive and negative observations without claiming sufficiency.\n- [ ] The design record covers correctness, time, space, compatibility and\n      recovery with one exact result per candidate and criterion.\n- [ ] Every resolved result binds a zero-exit report; every pending result names\n      its resolver, future report and exact blocking transition.\n- [ ] The design-lock check exits zero, and the chosen candidate is on the\n      mechanically computed frontier under its declared rule.\n- [ ] The runbook's design-lock block matches the record digest and selected\n      candidate before Step 1.\n- [ ] When known failures exist, the inventory checker exits zero against the\n      independent expected-id set, exact runbook, and current source views.\n- [ ] Always, ask-first and never each carry concrete entries.\n- [ ] Each step carries goal, entry, exit, files, tests and disciplines.\n- [ ] Every discipline a step names carries the reason it applies.\n- [ ] No exit rests on anything but a command.\n- [ ] Step 1 scaffolds and the last step demonstrates.\n- [ ] Steps are in dependency order.\n- [ ] If the topic was decomposed, every step traces to a module id.\n\n## Hand back\n\nLead with the state: ready to build, or blocked on a named gap. Then give the\ncount of checks passed out of the total, and name each one that failed.\n\nKeep three things apart. What the study establishes, what it assumes, and what\ncould not be settled. An assumption that changes the build order or the chosen\ndesign gets said out loud, not buried in a section.\n\nEnd with one action, and make it something the reader can do in a couple of\nminutes: confirm an assumption, answer one question, or approve the runbook.\nName the open question rather than closing it with a guess. Corrected here, an\nassumption costs a sentence. Found in the audit loop, it costs a step.\n\n## Promise Machine contract\n\n### protasis-known-failure-inventory\n\n- Promise: A clean inventory result establishes that one closed `protasis-known-failure-inventory\u002Fv1` object matches an independent exact id set, current source and synopsis bytes, portable guard paths, structurally bound commands, admitted reports, and real consuming runbook steps, or carries one digest-bound no-known-findings claim.\n- Evidence: The exact study and runbook bytes, caller or effective runbook expected ids, source and synopsis paths and SHA-256 values, finding fields and assignments, parsed argv, path and count caps, locked replacement-Exit map, stable K000 to K012 result, and focused hostile cases.\n- Evidence classes: checked\n- Boundary: This result establishes bounded structure and byte currency. A `source_ref` prefix selects a checked source view, but its locator text is opaque and does not prove that a cited line, range, or statement supports the failure. The result does not discover unknown failures, judge Surveyor's completeness, execute a guard, classify Elenchus output, write a receipt, or authorise a product edit by itself.\n- Authorises: Carrying the exact checked inventory into runbook derivation and the later Fiat inoculation boundary without inferring an omitted entry or an empty success.\n- Consequence: 1\n- Refuses: An absent, duplicated, open, malformed, non-finite, oversized, stale, unsafe, unassigned, command-drifted, id-mismatched, bracket-bearing runbook, replacement-ambiguous, post-lock generation-drifted, or unbound empty inventory.\n- Recovery: Repair the named K000 to K012 cause in the study, runbook, expected-id set, or source view, then rerun the complete read-only check.\n- Exceptions: none\n\n### protasis-study-readiness\n\n- Promise: A study accepted by Protasis states the problem, assumptions, current state, boundaries, failure and recovery model, affected versions and success criteria, and locks one candidate through a complete checked `protasis-design-evidence\u002Fv1` matrix without predicting evidence that is not yet available.\n- Evidence: The exact study, source inventory, explicit unknowns and exclusions, answered discipline questions, design record and report digests, mechanical frontier result, named future resolvers and completed study checklist.\n- Evidence classes: checked, inferred, recorded\n- Boundary: Readiness establishes the closed record, the reports checked at design lock and the declared stop points for later evidence. It does not establish that the selected design is correct, that pending conformance will pass, that implementation exists, or that later receipts are true.\n- Authorises: Derivation of a discrete runbook from the accepted study without silently adding a new design decision.\n- Consequence: 1\n- Refuses: Hidden assumptions, an unstated boundary, solution-first prose, an incomplete candidate matrix, missing required concern, unresolved selection evidence, a selected dominated candidate, unsupported tie-break, untestable success language or a topic still containing several independent deliveries.\n- Recovery: Name the missing question or decision, gather the required evidence, amend the study and repeat the complete Protasis review.\n- Exceptions: none\n\n### protasis-runbook-readiness\n\n- Promise: A runbook accepted by Protasis binds the exact selected design and decomposes the study into ordered steps whose entry, modules, exit commands, files, tests and discipline effects are discrete and provable, and any optional governed-skill version relation has one closed source without a competing concrete target token. When the study carries a known-failure inventory, Fiat must join this result to the clean `protasis-known-failure-inventory` result before product work.\n- Evidence: The accepted study, design-evidence digest and selected candidate, exact runbook and matching design-lock block, `protasis.py` structural result, per-step commands and files, dependency order, optional version-relations block, version boundary, completed pre-receipt checklist and, when present, the clean known-failure inventory result.\n- Evidence classes: checked, inferred, recorded\n- Boundary: Runbook readiness establishes buildable specification content, the design lock, and the lexical shape of a declared version relation. It does not establish correct implementation, later conformance evidence, command success, relation suitability, a selected version or integration base, audit closure or delivery completion. When an inventory is present, runbook readiness alone never authorises a product edit.\n- Authorises: Entering controller-governed runbook execution at the first step and, when an inventory is present, its required inoculation gate.\n- Consequence: 1\n- Refuses: A missing or mismatched design lock, a step with no executable exit, mixed independent outcomes, missing affected files or tests, forward references to an undecided design, receipt language with no evidence command, an applicable inventory without a clean joined result, or a malformed, ambiguous or concretely contradicted version relation.\n- Recovery: Split or reorder the failing step, supply its exact evidence and rerun both the mechanical check and the full content review.\n- Exceptions: none\n\n### protasis-gate-command-validation\n\n- Promise: A successful command check establishes that each effective declared invocation fits the registered CLI interface, with exact source and report bindings, or, while a runner that Step 1 creates has no binding, passes every check except that runner's parser interface without reading it; a registered module whose AST digest left the current pin fits only when the caller's starting-commit bindings carry that path's AST and source digest pair and both equal what was read; no declared command is executed and no target module imported.\n- Evidence: Captured runbook SHA-256 and source root, raw command text and UTF-8 offsets, command digests, superseded-source records, original and substituted argv, full CLI and adapter digests, deferred records naming the creating step, the caller's binding and phase record, the caller's starting-commit bindings when supplied, reviewed adapter, supplied starting-commit adapter or exact runner source-pair compatibility when applicable, declaration digests, interface results, current-root report resolution and `operation_ran:false` under `protasis-gate-commands\u002Fv1`.\n- Evidence classes: checked, recorded\n- Boundary: Interface validity establishes supported argument and source shape. It does not establish command execution, executable behavior, test success, report truth, audit judgement or atomic protection against later source changes. Historical absolute argv records nonexecuted derivation and grants no authority to execute at an old root; checkpoint relocation belongs to Fiat. An `interface-deferred` result establishes no parser fit, execution or success; a runner binding pins that file only, not the modules it imports, and the adapter takes the run's phase and binding from its caller. The adapter derives no starting-commit binding and reads no Git: it admits only the pair its caller supplies, and its result records no provenance for that admission. This Consequence 1 operation has no native runtime binding. Its actual interface result retains operation_ran:false because the declared commands were not executed.\n- Authorises: Supplying the exact interface result to Fiat as command-validation evidence for its separately governed runbook or amendment receipt.\n- Consequence: 1\n- Refuses: Unregistered commands, malformed or misplaced local registrations, a deferred row naming another step, added after Step 1 starts or repeated after binding, a present or unsafe unbound deferred path when absence is required, an inconsistent binding or phase record, registry overrides, unsupported parser declarations or converters, private worker arguments, unsupported shell evaluation, invalid literal or loop argv, malformed command fences, missing or unsafe report declarations, CLI source drift outside the reviewed timestamp pair, a starting-commit pair that does not equal the module read or a bindings value outside its closed shape, unknown adapter digests, bound violations or replay mismatch.\n- Recovery: Preserve the original command and refusal, correct the declared interface or provide a reviewed adapter, then repeat the check. For a deferred row, remove what occupies its path before capture, or supply the recorded binding and phase. For a receipted run, use Fiat's permitted append-only amendment when the source or bindings changed. A released adapter listed in the [command reference](references\u002Fgate-commands.md) may replay unchanged evidence without amendment, as does the adapter at the run's starting commit when the caller supplies its digest. That reference also permits one exact runner timestamp transition when every invocation uses the reviewed source pair and every other receipt field matches; earlier receipts remain intact.\n- Exceptions: none\n",{"name":5,"metadata":24,"description":7},{"version":25},"6.20.1","main","6b6406ecd922c4127bc688e0b5122bbe395be0eb","plugins\u002Fhexaemeron\u002Fskills\u002Fprotasis",6,"https:\u002F\u002Fgithub.com\u002Fwildcat-finance\u002Fskills","https:\u002F\u002Fgithub.com\u002Fwildcat-finance","https:\u002F\u002Fwildcat.finance","Banking, But Worse","2026-11-07 17:14:07",{"source":8,"skillName":5,"commands":36,"prompt":46,"zipAvailable":47},[37,40,43],{"tool":38,"command":39},"npx","npx skills add wildcat-finance\u002Fskills --skill protasis",{"tool":41,"command":42},"bunx","bunx skills add wildcat-finance\u002Fskills --skill protasis",{"tool":44,"command":45},"pnpm","pnpm dlx skills add wildcat-finance\u002Fskills --skill protasis","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fwildcat-finance\u002Fskills\u002Fblob\u002Fmain\u002Fplugins\u002Fhexaemeron\u002Fskills\u002Fprotasis\u002FSKILL.md",true,[49,55,59,63,67,73,77,81,85],{"path":50,"resourceType":51,"resourceTypeDesc":52,"sizeBytes":53,"sourceHtmlUrl":54},"agents\u002Fopenai.yaml","REFERENCE","参考文档",216,"https:\u002F\u002Fgithub.com\u002Fwildcat-finance\u002Fskills\u002Fblob\u002Fmain\u002Fplugins\u002Fhexaemeron\u002Fskills\u002Fprotasis\u002Fagents\u002Fopenai.yaml",{"path":56,"resourceType":51,"resourceTypeDesc":52,"sizeBytes":57,"sourceHtmlUrl":58},"DEMONSTRATION.md",2692,"https:\u002F\u002Fgithub.com\u002Fwildcat-finance\u002Fskills\u002Fblob\u002Fmain\u002Fplugins\u002Fhexaemeron\u002Fskills\u002Fprotasis\u002FDEMONSTRATION.md",{"path":60,"resourceType":51,"resourceTypeDesc":52,"sizeBytes":61,"sourceHtmlUrl":62},"EVOLUTION.md",23522,"https:\u002F\u002Fgithub.com\u002Fwildcat-finance\u002Fskills\u002Fblob\u002Fmain\u002Fplugins\u002Fhexaemeron\u002Fskills\u002Fprotasis\u002FEVOLUTION.md",{"path":64,"resourceType":51,"resourceTypeDesc":52,"sizeBytes":65,"sourceHtmlUrl":66},"references\u002Fgate-commands.md",25632,"https:\u002F\u002Fgithub.com\u002Fwildcat-finance\u002Fskills\u002Fblob\u002Fmain\u002Fplugins\u002Fhexaemeron\u002Fskills\u002Fprotasis\u002Freferences\u002Fgate-commands.md",{"path":68,"resourceType":69,"resourceTypeDesc":70,"sizeBytes":71,"sourceHtmlUrl":72},"scripts\u002Fdesign_evidence.py","SCRIPT","脚本",30436,"https:\u002F\u002Fgithub.com\u002Fwildcat-finance\u002Fskills\u002Fblob\u002Fmain\u002Fplugins\u002Fhexaemeron\u002Fskills\u002Fprotasis\u002Fscripts\u002Fdesign_evidence.py",{"path":74,"resourceType":69,"resourceTypeDesc":70,"sizeBytes":75,"sourceHtmlUrl":76},"scripts\u002Fgate_commands.py",49755,"https:\u002F\u002Fgithub.com\u002Fwildcat-finance\u002Fskills\u002Fblob\u002Fmain\u002Fplugins\u002Fhexaemeron\u002Fskills\u002Fprotasis\u002Fscripts\u002Fgate_commands.py",{"path":78,"resourceType":69,"resourceTypeDesc":70,"sizeBytes":79,"sourceHtmlUrl":80},"scripts\u002Fknown_failure_inventory.py",58800,"https:\u002F\u002Fgithub.com\u002Fwildcat-finance\u002Fskills\u002Fblob\u002Fmain\u002Fplugins\u002Fhexaemeron\u002Fskills\u002Fprotasis\u002Fscripts\u002Fknown_failure_inventory.py",{"path":82,"resourceType":69,"resourceTypeDesc":70,"sizeBytes":83,"sourceHtmlUrl":84},"scripts\u002Fprotasis.py",42184,"https:\u002F\u002Fgithub.com\u002Fwildcat-finance\u002Fskills\u002Fblob\u002Fmain\u002Fplugins\u002Fhexaemeron\u002Fskills\u002Fprotasis\u002Fscripts\u002Fprotasis.py",{"path":86,"resourceType":69,"resourceTypeDesc":70,"sizeBytes":87,"sourceHtmlUrl":88},"scripts\u002Fsuccess_criteria.py",20519,"https:\u002F\u002Fgithub.com\u002Fwildcat-finance\u002Fskills\u002Fblob\u002Fmain\u002Fplugins\u002Fhexaemeron\u002Fskills\u002Fprotasis\u002Fscripts\u002Fsuccess_criteria.py",[90,97,103,110,116,122],{"skillNo":91,"name":92,"nameZh":6,"description":93,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":94,"categories":95,"keywords":96,"publishedAt":19},"01M7APAWXHFSKVQEJBAHKASH25","example","A deliberately unsupported evidence class fixture.","https:\u002F\u002Fgithub.com\u002Fwildcat-finance\u002Fskills\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fpromise-machine\u002Funsupported-evidence-class\u002Fplugins\u002Fexample\u002Fskills\u002Fexample\u002FSKILL.md",[],[],{"skillNo":98,"name":92,"nameZh":6,"description":99,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":100,"categories":101,"keywords":102,"publishedAt":19},"01M7APATBK93107VF5NMPA3RJH","A router fixture target.","https:\u002F\u002Fgithub.com\u002Fwildcat-finance\u002Fskills\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fpromise-machine\u002Funresolved-router\u002Fplugins\u002Fexample\u002Fskills\u002Fexample\u002FSKILL.md",[],[],{"skillNo":104,"name":105,"nameZh":6,"description":106,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":107,"categories":108,"keywords":109,"publishedAt":19},"01M7APARV9C0MQB8WTB0GGAAMG","promise-machine","A deliberately unresolved router.","https:\u002F\u002Fgithub.com\u002Fwildcat-finance\u002Fskills\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fpromise-machine\u002Funresolved-router\u002F.agents\u002Fskills\u002Fpromise-machine\u002FSKILL.md",[],[],{"skillNo":111,"name":92,"nameZh":6,"description":112,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":113,"categories":114,"keywords":115,"publishedAt":19},"01M7APAQMDG4977HXGSKMB29KV","A deliberately unclassified skill fixture.","https:\u002F\u002Fgithub.com\u002Fwildcat-finance\u002Fskills\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fpromise-machine\u002Funclassified-skill\u002Fplugins\u002Fexample\u002Fskills\u002Fexample\u002FSKILL.md",[],[],{"skillNo":117,"name":92,"nameZh":6,"description":118,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":119,"categories":120,"keywords":121,"publishedAt":19},"01M7APAPE6DE05A24J4BYPK6WY","A deliberately unattributed exception fixture.","https:\u002F\u002Fgithub.com\u002Fwildcat-finance\u002Fskills\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fpromise-machine\u002Funattributed-exception\u002Fplugins\u002Fexample\u002Fskills\u002Fexample\u002FSKILL.md",[],[],{"skillNo":123,"name":92,"nameZh":6,"description":124,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":125,"categories":126,"keywords":127,"publishedAt":19},"01M7APAKYXRTABBSN7Q637CC6M","A deliberately conflated version fixture.","https:\u002F\u002Fgithub.com\u002Fwildcat-finance\u002Fskills\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fpromise-machine\u002Fpackage-as-skill-version\u002Fplugins\u002Fexample\u002Fskills\u002Fexample\u002FSKILL.md",[],[]]