" type: source depth: research confidence: high # authoritative if peer-reviewed/published; see CLAUDE.md §4.6 audited: <YYYY-MM-DD> # today — assignment with the source in context is the check (§4.6) tags: [paper, <field>] authors: [<First Author>, <…>] year: <YYYY> venue: "<journal / conference>" doi: "<DOI or stable URL>" sources: [raw/2-papers/<file>.pdf, raw/2-papers/<file>.md] # original + converted created: <YYYY-MM-DD> updated: <YYYY-MM-DD> --- ## Citation <full reference string> ## Research Question ## Methodology ## Key Findings - <finding with exact figures, e.g. "+12.3 BLEU, p<0.01"> ## Data / Setup ## Contributions ## Limitations & Threats to Validity ## Relation to Wiki - Confirms [[…]]; contradicts [[…]] → flag a `## Conflicts / Open Questions` block on that page ## Key Quotes > "<verbatim>" (§/p.) ## Open Questions / Follow-ups ## Related - [[…]] ``` **Dual provenance:** for a converted source, `sources:` lists **both** the converted `.md` and the original (file path, or the URL for a web or video source). Where the head sorts the raw file after you close, predict the post-sort path so the link does not break — your brief names it. **Research depth is not paper-only.** T1 ∧ T2 fire on any primary source whose exact wording matters (a regulation, an institutional handbook, a contract, a policy document). Keep the frontmatter fields that exist (no DOI → omit it; `venue:` takes the issuing body and version) and replace the paper-shaped middle sections with genre-fitting equivalents (Scope & Applicability · Requirements & Thresholds · Deadlines & Milestones · Governance). The obligations never flex — exact figures, verbatim quotes with locators, `## Relation to Wiki`, `## Related` — and you declare the substitution on the page and in your report. ## 3 · Confidence Assign every page's `confidence:` from `contract/confidence-rubric.md`, on source authority × verification × derivation, and stamp `audited:` with today's date in the same pass — the stamp is the audit trail, so never stamp a badge you did not check. Judge by **grounding strength, not source count**: one source primary for the page's own claims grounds `high` alone; a strong source primary only for something adjacent, or reputable-secondary grounding, is `medium`; a single paragraph in a single witness is `low`. Compiled pages cap at `high`, the cap being a ceiling and never a demotion. On a tie take the lower tier, and never raise an existing page's tier — you may propose a raise with the evidence, and the head reads the source before any raise stands. Report each page's tier with the reason that decided it. ## 4 · Network the knowledge Add the links the source warrants, and only the pages the schema warrants. - **Models and benchmarks are first-class.** Any LLM named gets a page under `wiki/models/`, any evaluation dataset a page under `wiki/benchmarks/`, and the link runs **both ways**: the source under the model or benchmark page's `## Appears in`, the models and benchmarks under the source page's `## Related`. - **A product name is a model mention.** A source that names an LLM product or family even once, even as a licence or plan name (ChatGPT Edu, GPT-4o, Claude, Qwen), links the family's model page from `## Related` — ChatGPT and every GPT-n fold into `[[GPT]]` — and says on the page what the source names; the model page's `## Appears in` gains the source. The link does not require an individually named model (schema §4.5). - **Reuse, never duplicate.** Fold a variant into its existing page (a point release into the family page, a subset benchmark into its parent) rather than creating a near-twin. Canonicalise every name against `wiki/index.md` in `name` mode (`grep -o '^- \[\[[^]|]*' wiki/index.md | sed 's/^- \[\[//'`) before you write it — never a whole read. - **A page exists** → read it and merge incrementally. Never clobber, and never silently overwrite a contradiction: keep both statements under `## Conflicts / Open Questions` and report it. A conflict never pauses you. - **The stub rule.** A tool or an entity that the source gives one paragraph or one row, with no second witness, is **a fact on the source page — never a link and never a page**. A link forces a page, and a one-paragraph page carries no behaviour the fact would not. Create a tool, entity or concept page only where it has enough substance for the `## Definition` and `## Key Points` the schema requires. Models and benchmarks keep their rule above: they get pages either way. - Every page you write ends with a `## Related` section carrying at least one link (no orphans), and every link you write resolves to a real page or to a target in your emitted claims. ## 5 · Registries: propose, do not write A log entry stays concise (the §5 shape: the title line, `- **Changed**:`, `- **Conflicts**:`); detail belongs on the pages, never in the entry. Frontmatter carries values only: no run id, lane id or evidence in a comment — your depth and confidence evidence goes in the report. `wiki/index.md` and `wiki/log.md` are the head agent's to write. Return, in your report: - the `index.md` line for each page, under the heading it belongs to, one line each: `- [[Page]] — one-line description.` - the `log.md` entry the run would carry, in the schema's shape. Only where your brief explicitly grants the ingest exception do you write them yourself, and then by anchored per-heading `Edit` for the index and a shell append for the log — never a whole-file read-modify-write, which commits a stale snapshot and erases a concurrent writer's entry. Verify either write back with a grep and show the grep. Any model or effort value you write into a registry entry is copied from your brief's first line or omitted — never inferred. **Shared-type pages under a parallel run.** Where your brief says other lanes are compiling alongside you, do not create or edit entity, concept, model or benchmark pages. Emit a claim and an anchored unified diff per target into the diff directory your brief explicitly grants. The head plans and applies the accepted diffs. Mirror each complete claim under `claims` in the `claims_emitted` ledger checkpoint as well as your report: `{name · type · kind: create|update · target (canonicalised against index.md — a variant arrives as an update to the family page) · facts[] with a per-fact source locator · links[] · appears_in[] · confidence · from: source-page · source: raw identity · diff: absolute path · base: sha256 prefix · diff_hash: sha256 prefix}` An update diff is against the page exactly as read, with `base` the first 16 sha256 hex digits (the `source_hash` length used by de-dup). A create is a new-file diff carrying the schema's frontmatter and sections and `aliases:` with acronym and expansion wherever the source gives both. Record the diff's own sha256 prefix in `diff_hash`. Never apply the diff yourself. The lane's `lane_open` carries `role: compile` and the same lane id as the spawn record for metering. Checkpoint key: `claims`: the list of claim objects, never a count; an optional count goes under `claims_count`. Each member is a complete object with a non-empty `name`; an empty list means no claims. Never use `claim_objects` as a substitute key. Shape-only example for a source with no claims (add the run and write-time timestamp): `{"event":"checkpoint","stage":"claims_emitted","lane":"<lane>","source":"<raw identity>","claims":[],"claims_count":0}` **Mandatory post-checkpoint assertion.** After every `claims_emitted` append, run the following read-only Python through `python3 -B -` with three quoted arguments: the literal ledger path, this lane id, and this checkpoint's exact source identity. Its stdin is the code below. Record the `claims-shape:` and `controls:` lines in the lane's final report under a `## Controls` heading, with one unquoted line `claims-shape: checkpoints N · objects M · invalid 0`. This evidence is required before `registered` or `lane_close`; a failed or missing assertion is a failed output gate. The verifier consumes these report lines. The check reads all matching checkpoints, so a later good append cannot conceal an earlier malformed one; it checks list shape, not factual accuracy or the full claim schema. ```python # claims-shape assertion: kept executable for the brief suite. import json import sys def stop(message): print("PROBE FAILED: " + message) raise SystemExit(2) def valid(event): claims = event.get("claims") return (isinstance(claims, list) and all(isinstance(c, dict) and isinstance(c.get("name"), str) and c["name"].strip() for c in claims) and ("claims_count" not in event or (type(event["claims_count"]) is int and event["claims_count"] == len(claims)))) if (valid({"claims": 1, "claim_objects": [{"name": "Example"}]}) or valid({"claims": [{}]}) or not valid({"claims": [{"name": "Example"}], "claims_count": 1}) or not valid({"claims": [], "claims_count": 0})): stop("claims-shape controls did not discriminate") print("controls: integer claims rejected; unnamed object rejected; object list and empty list accepted") if len(sys.argv) != 4 or not all(sys.argv[1:]): stop("claims-shape needs ledger, lane and source") ledger, lane, source = sys.argv[1:] matched = objects = 0 try: with open(ledger, encoding="utf-8") as handle: for number, line in enumerate(handle, 1): event = json.loads(line) if not isinstance(event, dict): stop("ledger line %d is not an object" % number) if (event.get("event") == "checkpoint" and event.get("stage") == "claims_emitted" and event.get("lane") == lane and event.get("source") == source): if not valid(event): stop("line %d: claims must be a list of named objects; claims_count must equal its length" % number) matched += 1 objects += len(event["claims"]) except (OSError, ValueError) as error: stop("cannot check ledger: %s" % error) if not matched: stop("no claims_emitted checkpoint for this lane and source") print("claims-shape: checkpoints %d · objects %d · invalid 0" % (matched, objects)) ``` Mark a shaky fact `unverified`. A claim's facts each carry the locator they came from; a fact without one does not go in the claim. ## 6 · Context notes, plants and anomalies Your brief's CONTEXT NOTES are pointers, never sources. A note the raw does not bear out is not compiled: list it under `## Anomalies` as "not in the raw". One such note may be a deliberate plant testing exactly this, and the rule is the same either way. Your report ends with `## Anomalies`: everything the rubric does not cover, each with a locator — a source-internal inconsistency (kept on the page, never resolved), a raw-versus-wiki conflict and the block you wrote for it, a provenance gap, a context note not borne out, a thin or truncated source, anything left unsettled. `none` is a claim, and names what you checked. ## Hard constraints - **Never modify or delete the text inside a raw file.** Raw content is immutable; you neither move nor sort raw files unless your brief says so. - Every page carries the frontmatter and the sections `contract/schema-s4.md` specifies for its type, a `confidence:` with today's `audited:`, and a `## Related` section. - Every source page carries `depth:`, decided after the read, inside the authorised range. Never backfill `depth:` onto a page compiled before the rule. - Write everything in British/UK English, translating a non-English source; US spelling survives only inside a verbatim quote, a proper noun or code. - Never fabricate. Where you cannot ground a claim, mark it `unverified` and cite what you have, or leave it out and report the gap. - Report every page you created or updated with its depth locator and its confidence reason. That report is the completion gate: no done-declaration without it.

--- name: compile-core description: > The compile lane's half of the ingest skill: choose a source's depth after reading it, write the source page from the depth's template, assign confidence, network the knowledge, and propose the registry lines rather than writing them. Preloaded into a wiki-compile lane home; the head keeps the orchestration half (conversion, de-dup, pacing, parallel merge, the verify step, sorting, the run ledger). Slice source: the ingest skill, sections Depth, Step 3, Step 4 and Hard constraints. user-invocable: false --- # compile-core — compile an assigned source into wiki pages You compile the sources your brief assigns and nothing else. A gap in them is reported, never filled from your own knowledge. The page schema, its frontmatter fields and the confidence ordinal live in `contract/schema-s4.md` and `contract/confidence-rubric.md` beside this slice — read them rather than recalling them. ## 1 · Read, then choose the depth Read the source in full first. **Depth is decided after the read — never from a filename, folder, length or marker**, none of which separate a research-worthy source from an ordinary one. Your brief names the run's **authorised range** (default: all three). Never choose outside it. If a source genuinely needs a depth the range excludes, compile at the nearest authorised depth, say so, and report it as a gap. Three triggers, judged on the source you have just read: - **T1 · Reuse** — original figures, a reimplementable method, or wording where a paraphrase would falsify the claim. - **T2 · Proximity** — it bears on a topic the owner's own pages name, or on an open question the wiki already holds (a `## Conflicts / Open Questions` block, a `flagged:` page). Where those pages are outside your reading list, read T2 as "a topic with at least three source pages in `wiki/index.md`"; where the wiki is too small for that, say so once and work in concise or standard only. - **T3 · Novelty** — it adds something no existing page holds. **`research` = T1 ∧ T2 · `concise` = ¬T3 · `standard` = everything else.** Overrides, which fire regardless of the triggers: a source that is the owner's own work is never below standard; a source that corrects an existing wiki page is always research, because a correction has to be exact; a source whose conversion collapsed word boundaries can never be research (verbatim quoting is impossible on it) — compile at standard and say so. **Name the evidence or drop a rung.** Every depth you record carries a locator: `research — T1 (Table 3 ablation) · T2 (calibration, an owner page)`, `concise — ¬T3 (link post, all three tools already have pages)`. A trigger letter with empty parentheses is a defect, not a record. Where you cannot name the evidence, take the lower rung. What each depth changes: - **standard** (the usual outcome) — articles, blogs, posts, docs. Sections 2–4 as written. - **concise** — thin or fully redundant sources: a concise summary, minimal bullets, new pages only for genuinely new entities or concepts. - **research** — primary material the owner will need to reuse exactly. Preserve exact figures (no rounding), quote critical claims verbatim with section or page references, mark anything not directly stated as `unverified`, never infer a number, use the research template below, add the academic frontmatter, and cross-check findings against existing pages, flagging confirmations and contradictions explicitly. Long block quotes stay in the raw file — link, never transplant. Record it always: the source page carries `depth:`, and your report carries the depth with its locator for every source. ## 2 · The source page Write to `wiki/sources/.md`, kebab-case, inside your file whitelist. Both templates are copied verbatim from the ingest skill, comments included. Where a comment cites `CLAUDE.md §4.x`, the text it points at is in the `contract/` slices you hold: §4.1–§4.6 in `contract/schema-s4.md`, the confidence rubric in `contract/confidence-rubric.md`. **Standard and concise depth:** ```markdown --- title: "Source: " type: source depth: concise | standard | research # the depth you chose (Depth) — required on every source page confidence: medium # per CLAUDE.md §4.6 — reflects the source: peer-reviewed/expert→authoritative · preprint/official-doc/owner-work/faithful-summary(default)→high · secondary or adjacent-only grounding→medium · promo/social/transcript→low (grounding strength, not source count; a user instruction can override the tier) audited: # today — assignment with the source in context is the check (§4.6); on updating an existing page, re-check its badge and re-stamp tags: [topic] sources: [raw/2-papers/report.md, raw/2-papers/report.pdf] # converted .md AND original; one entry if native .md or URL source_url: "" # de-dup source_hash: "" # de-dup — the length the pre-flight greps created: updated: --- ## Summary [The core summary, with the detail its claims need.] ## Key Takeaways - ... ## Related - [[EntityName]] — why related - [[Concept Name]] — why related ``` **Research depth (replaces the template above):** ```markdown --- title: "Paper: " type: source depth: research confidence: high # authoritative if peer-reviewed/published; see CLAUDE.md §4.6 audited: <YYYY-MM-DD> # today — assignment with the source in context is the check (§4.6) tags: [paper, <field>] authors: [<First Author>, <…>] year: <YYYY> venue: "<journal / conference>" doi: "<DOI or stable URL>" sources: [raw/2-papers/<file>.pdf, raw/2-papers/<file>.md] # original + converted created: <YYYY-MM-DD> updated: <YYYY-MM-DD> --- ## Citation <full reference string> ## Research Question ## Methodology ## Key Findings - <finding with exact figures, e.g. "+12.3 BLEU, p<0.01"> ## Data / Setup ## Contributions ## Limitations & Threats to Validity ## Relation to Wiki - Confirms [[…]]; contradicts [[…]] → flag a `## Conflicts / Open Questions` block on that page ## Key Quotes > "<verbatim>" (§/p.) ## Open Questions / Follow-ups ## Related - [[…]] ``` **Dual provenance:** for a converted source, `sources:` lists **both** the converted `.md` and the original (file path, or the URL for a web or video source). Where the head sorts the raw file after you close, predict the post-sort path so the link does not break — your brief names it. **Research depth is not paper-only.** T1 ∧ T2 fire on any primary source whose exact wording matters (a regulation, an institutional handbook, a contract, a policy document). Keep the frontmatter fields that exist (no DOI → omit it; `venue:` takes the issuing body and version) and replace the paper-shaped middle sections with genre-fitting equivalents (Scope & Applicability · Requirements & Thresholds · Deadlines & Milestones · Governance). The obligations never flex — exact figures, verbatim quotes with locators, `## Relation to Wiki`, `## Related` — and you declare the substitution on the page and in your report. ## 3 · Confidence Assign every page's `confidence:` from `contract/confidence-rubric.md`, on source authority × verification × derivation, and stamp `audited:` with today's date in the same pass — the stamp is the audit trail, so never stamp a badge you did not check. Judge by **grounding strength, not source count**: one source primary for the page's own claims grounds `high` alone; a strong source primary only for something adjacent, or reputable-secondary grounding, is `medium`; a single paragraph in a single witness is `low`. Compiled pages cap at `high`, the cap being a ceiling and never a demotion. On a tie take the lower tier, and never raise an existing page's tier — you may propose a raise with the evidence, and the head reads the source before any raise stands. Report each page's tier with the reason that decided it. ## 4 · Network the knowledge Add the links the source warrants, and only the pages the schema warrants. - **Models and benchmarks are first-class.** Any LLM named gets a page under `wiki/models/`, any evaluation dataset a page under `wiki/benchmarks/`, and the link runs **both ways**: the source under the model or benchmark page's `## Appears in`, the models and benchmarks under the source page's `## Related`. - **A product name is a model mention.** A source that names an LLM product or family even once, even as a licence or plan name (ChatGPT Edu, GPT-4o, Claude, Qwen), links the family's model page from `## Related` — ChatGPT and every GPT-n fold into `[[GPT]]` — and says on the page what the source names; the model page's `## Appears in` gains the source. The link does not require an individually named model (schema §4.5). - **Reuse, never duplicate.** Fold a variant into its existing page (a point release into the family page, a subset benchmark into its parent) rather than creating a near-twin. Canonicalise every name against `wiki/index.md` in `name` mode (`grep -o '^- \[\[[^]|]*' wiki/index.md | sed 's/^- \[\[//'`) before you write it — never a whole read. - **A page exists** → read it and merge incrementally. Never clobber, and never silently overwrite a contradiction: keep both statements under `## Conflicts / Open Questions` and report it. A conflict never pauses you. - **The stub rule.** A tool or an entity that the source gives one paragraph or one row, with no second witness, is **a fact on the source page — never a link and never a page**. A link forces a page, and a one-paragraph page carries no behaviour the fact would not. Create a tool, entity or concept page only where it has enough substance for the `## Definition` and `## Key Points` the schema requires. Models and benchmarks keep their rule above: they get pages either way. - Every page you write ends with a `## Related` section carrying at least one link (no orphans), and every link you write resolves to a real page or to a target in your emitted claims. ## 5 · Registries: propose, do not write A log entry stays concise (the §5 shape: the title line, `- **Changed**:`, `- **Conflicts**:`); detail belongs on the pages, never in the entry. Frontmatter carries values only: no run id, lane id or evidence in a comment — your depth and confidence evidence goes in the report. `wiki/index.md` and `wiki/log.md` are the head agent's to write. Return, in your report: - the `index.md` line for each page, under the heading it belongs to, one line each: `- [[Page]] — one-line description.` - the `log.md` entry the run would carry, in the schema's shape. Only where your brief explicitly grants the ingest exception do you write them yourself, and then by anchored per-heading `Edit` for the index and a shell append for the log — never a whole-file read-modify-write, which commits a stale snapshot and erases a concurrent writer's entry. Verify either write back with a grep and show the grep. Any model or effort value you write into a registry entry is copied from your brief's first line or omitted — never inferred. **Shared-type pages under a parallel run.** Where your brief says other lanes are compiling alongside you, do not create or edit entity, concept, model or benchmark pages. Emit a claim and an anchored unified diff per target into the diff directory your brief explicitly grants. The head plans and applies the accepted diffs. Mirror each complete claim under `claims` in the `claims_emitted` ledger checkpoint as well as your report: `{name · type · kind: create|update · target (canonicalised against index.md — a variant arrives as an update to the family page) · facts[] with a per-fact source locator · links[] · appears_in[] · confidence · from: source-page · source: raw identity · diff: absolute path · base: sha256 prefix · diff_hash: sha256 prefix}` An update diff is against the page exactly as read, with `base` the first 16 sha256 hex digits (the `source_hash` length used by de-dup). A create is a new-file diff carrying the schema's frontmatter and sections and `aliases:` with acronym and expansion wherever the source gives both. Record the diff's own sha256 prefix in `diff_hash`. Never apply the diff yourself. The lane's `lane_open` carries `role: compile` and the same lane id as the spawn record for metering. Checkpoint key: `claims`: the list of claim objects, never a count; an optional count goes under `claims_count`. Each member is a complete object with a non-empty `name`; an empty list means no claims. Never use `claim_objects` as a substitute key. Shape-only example for a source with no claims (add the run and write-time timestamp): `{"event":"checkpoint","stage":"claims_emitted","lane":"<lane>","source":"<raw identity>","claims":[],"claims_count":0}` **Mandatory post-checkpoint assertion.** After every `claims_emitted` append, run the following read-only Python through `python3 -B -` with three quoted arguments: the literal ledger path, this lane id, and this checkpoint's exact source identity. Its stdin is the code below. Record the `claims-shape:` and `controls:` lines in the lane's final report under a `## Controls` heading, with one unquoted line `claims-shape: checkpoints N · objects M · invalid 0`. This evidence is required before `registered` or `lane_close`; a failed or missing assertion is a failed output gate. The verifier consumes these report lines. The check reads all matching checkpoints, so a later good append cannot conceal an earlier malformed one; it checks list shape, not factual accuracy or the full claim schema. ```python # claims-shape assertion: kept executable for the brief suite. import json import sys def stop(message): print("PROBE FAILED: " + message) raise SystemExit(2) def valid(event): claims = event.get("claims") return (isinstance(claims, list) and all(isinstance(c, dict) and isinstance(c.get("name"), str) and c["name"].strip() for c in claims) and ("claims_count" not in event or (type(event["claims_count"]) is int and event["claims_count"] == len(claims)))) if (valid({"claims": 1, "claim_objects": [{"name": "Example"}]}) or valid({"claims": [{}]}) or not valid({"claims": [{"name": "Example"}], "claims_count": 1}) or not valid({"claims": [], "claims_count": 0})): stop("claims-shape controls did not discriminate") print("controls: integer claims rejected; unnamed object rejected; object list and empty list accepted") if len(sys.argv) != 4 or not all(sys.argv[1:]): stop("claims-shape needs ledger, lane and source") ledger, lane, source = sys.argv[1:] matched = objects = 0 try: with open(ledger, encoding="utf-8") as handle: for number, line in enumerate(handle, 1): event = json.loads(line) if not isinstance(event, dict): stop("ledger line %d is not an object" % number) if (event.get("event") == "checkpoint" and event.get("stage") == "claims_emitted" and event.get("lane") == lane and event.get("source") == source): if not valid(event): stop("line %d: claims must be a list of named objects; claims_count must equal its length" % number) matched += 1 objects += len(event["claims"]) except (OSError, ValueError) as error: stop("cannot check ledger: %s" % error) if not matched: stop("no claims_emitted checkpoint for this lane and source") print("claims-shape: checkpoints %d · objects %d · invalid 0" % (matched, objects)) ``` Mark a shaky fact `unverified`. A claim's facts each carry the locator they came from; a fact without one does not go in the claim. ## 6 · Context notes, plants and anomalies Your brief's CONTEXT NOTES are pointers, never sources. A note the raw does not bear out is not compiled: list it under `## Anomalies` as "not in the raw". One such note may be a deliberate plant testing exactly this, and the rule is the same either way. Your report ends with `## Anomalies`: everything the rubric does not cover, each with a locator — a source-internal inconsistency (kept on the page, never resolved), a raw-versus-wiki conflict and the block you wrote for it, a provenance gap, a context note not borne out, a thin or truncated source, anything left unsettled. `none` is a claim, and names what you checked. ## Hard constraints - **Never modify or delete the text inside a raw file.** Raw content is immutable; you neither move nor sort raw files unless your brief says so. - Every page carries the frontmatter and the sections `contract/schema-s4.md` specifies for its type, a `confidence:` with today's `audited:`, and a `## Related` section. - Every source page carries `depth:`, decided after the read, inside the authorised range. Never backfill `depth:` onto a page compiled before the rule. - Write everything in British/UK English, translating a non-English source; US spelling survives only inside a verbatim quote, a proper noun or code. - Never fabricate. Where you cannot ground a claim, mark it `unverified` and cite what you have, or leave it out and report the gap. - Report every page you created or updated with its depth locator and its confidence reason. That report is the completion gate: no done-declaration without it.