elevanaltd/octave-mcp/octave-compression

===OCTAVE_COMPRESSION=== META: TYPE::SKILL VERSION::"4.1.1" STATUS::ACTIVE PURPOSE::"Decision rules and workflow for turning prose into OCTAVE at the right fidelity tier" REQUIRES::octave-literacy PATTERNS::[octave-tool-reference] SPEC_REFERENCE::octave-data-spec.oct.md

§1::TIER_SELECTION WHEN::"choose the tier before reading the source; the tier drives every later decision" TIERS: LOSSLESS: TARGET::"100 percent fidelity" PRESERVE::everything DROP::nothing USE::[ "critical reasoning", "legal documents", "safety analysis", "audit trails", "syntax references" ] CONSERVATIVE: TARGET::"85 to 90 percent fidelity" PRESERVE::"explanatory depth" DROP::redundancy LOSS::"10 to 15 percent: repetition, verbose phrasing, some edge cases" USE::[ "research summaries", "design decisions", "technical analysis" ] AGGRESSIVE: TARGET::"70 percent fidelity" PRESERVE::"core thesis and conclusions" DROP::"nuance or narrative" LOSS::"30 percent: explanatory depth, trade-off narratives, edge-case exploration" USE::[ "context window scarcity", "quick reference", "decision support" ] ULTRA: TARGET::"50 percent fidelity" PRESERVE::"facts and structure" DROP::"all narrative" LOSS::"50 percent: almost all explanatory content, trade-off reasoning" USE::[ "extreme scarcity", "embedding generation", "dense reference" ] DECISION_RULES::[ "Reconstruction accuracy critical: CONSERVATIVE or LOSSLESS.", "Context window scarce and loss acceptable: AGGRESSIVE or ULTRA.", "Decision-relevant facts must survive: CONSERVATIVE.", "Source is a rule or syntax reference: LOSSLESS; a reference compressed below its rules is a broken reference.", "Otherwise: CONSERVATIVE." ] EVIDENCE::"stated both ways. octave-mcp study (docs/research/compression-fidelity-round-trip-study.md): CONSERVATIVE with mythology domain labels as key prefixes kept 11 of 11 decision-relevant facts at 15 percent fewer tokens than the prose, because each labelled field was translated separately, which prevents fact merging. A later bench of this document language: mythology as writing escalates when misapplied and, under exact fit, is harmless and does no work plain words do not (confidence High); myth names as instruction vocabulary carried no readable weight over plain names (confidence Medium: never had headroom to read). Rule: mythology is not instruction vocabulary; write labels in plain words." §2::LOSS_ACCOUNTING RECEIPT_RULE::"North Star I4 (transform auditability): if bits were dropped, the output carries a receipt; no silent loss" REQUIRED_META_FIELDS::[COMPRESSION_TIER,LOSS_PROFILE] SCOPE::"the pair belongs in the META of compressed outputs only, as the named exception; skills and other documents do not carry it" LOSS_PROFILE_FORM::"a note-form sentence saying what was kept and what was dropped" EXAMPLES: CONSERVATIVE::"kept causal chains; dropped verbose phrasing" AGGRESSIVE::"kept core thesis and conclusions; dropped explanatory depth and edge cases" ULTRA::"kept facts and structure; dropped all narrative and trade-off reasoning" META_BLOCK_TEMPLATE: `` META: TYPE::DECISION VERSION::"1.0.0" COMPRESSION_TIER::CONSERVATIVE LOSS_PROFILE::"kept causal chains and numbers; dropped verbose phrasing" ` KNOWN_GAP::"under schema=META STRICT both keys give E007 today (#514); the validator still pairs them: W_META_001 when COMPRESSION_TIER has no LOSS_PROFILE, W_META_002 when LOSS_PROFILE is none on a tier other than LOSSLESS (src/octave_mcp/core/validator.py, _check_meta_warnings)" CONSISTENCY_RULE::"LOSSLESS with a non-empty dropped list is a contradiction; if LOSS_PROFILE names anything dropped, the tier is not LOSSLESS" §3::TRANSFORMATION_WORKFLOW PHASE_1_READ: MAP::"find every causal chain (A causes B, X requires Y)" IDENTIFY::[ redundancy, verbosity, "conditional qualifiers" ] NOTE::"conditional qualifiers (when X, if Y, unless Z) carry material risk; never drop them" PHASE_2_EXTRACT: PRESERVE::[ "causal chains", "BECAUSE statements", metrics, "conditional qualifiers", "identifiers, copied exactly", "hedges, copied verbatim" ] COMPRESS::[ stopwords, repetition, "verbose phrasing", "narrative connectives" ] ANCHOR::"one concrete example per major abstraction anchors reconstruction fidelity" PHASE_3_COMPRESS: HIERARCHY::"group related concepts under parent block keys" ARRAYS::"turn parallel items into a list of quoted sentences or atoms" RELATIONS::"write relations in words, per the literacy §4b rules list" CONTRADICTIONS::"a source that contradicts itself is kept both ways and flagged" PHASE_4_VALIDATE: FIDELITY::"are all causal chains intact?" LOSS_RECEIPT::"does META carry COMPRESSION_TIER and LOSS_PROFILE?" GROUNDING::"is there at least one concrete example per major abstraction?" RECEIPTS::"read corrections[] and warnings[] from octave_write and triage them per octave-tool-reference §3::TRIAGE; repairs[] comes only from octave_validate; W_BARE_LINE_DROPPED and W_NUMERIC_KEY_DROPPED are silent data loss; empty warnings alone are not a clean receipt" §4::COMPRESSION_RULES RULES::[ "Keep causality: X leads to Y because Z; never flatten it to X leads to Y alone.", "Keep conditional qualifiers (when X, if Y, unless Z); they carry material risk.", "Keep trade-offs in words: gain versus loss.", "Keep one concrete example per major abstraction as the minimum reconstruction anchor.", "Never drop numbers, identifiers, thresholds or named entities; they cannot be rebuilt.", "Never place a value on a section header; it is dropped silently (literacy §3, the rule that a section header carries no value); content goes on child lines.", "Never invent to fill a slot or field: write none or not known`." ] §5::ANCHOR_KERNEL TARGET::"prose to OCTAVE at a declared tier, written in PITH" NEVER::["compressing before choosing the tier","dropping causal chains, conditions or numbers"] MUST::[ "Note form", "Exact identifiers (copied exactly, in backticks; never shortened or merged)", "Hedges verbatim", "Relations in words (no operators in prose or values)", "Nothing invented (none or not known)", "No writer-made ids, tags or counts", "One calibration sentence", "Judgement unattributed", "Code in fences", "Contradictions kept and flagged", "Plain vocabulary (no blobs, no mythology as instruction words)", "every line keyed, values with spaces or punctuation quoted", "no unkeyed lines", "no value on a section header", "contents indented under the header", "edits addressed as section dot key (§N.KEY)", "receipt fields read, not the status", "tier first; default conservative; references lossless", "compressed output keeps COMPRESSION_TIER and LOSS_PROFILE in META", "a LOSSLESS tier drops nothing", "one example per abstraction kept" ] GATE::"Could a reader rebuild every causal chain, condition and number, and does META say what was dropped?" §7::ANTI_PATTERNS ANTI_PATTERNS::[ "Markdown inside OCTAVE: no bold, no headings, no bullet hyphens.", "JSON or YAML syntax: no curly braces, no trailing commas, no YAML bullet hyphens.", "Nesting deeper than 3 levels: restructure with parent block keys.", "Silent loss: a compressed output always declares COMPRESSION_TIER and LOSS_PROFILE in META.", "Paradigm drift: validate for LLM parse efficiency, not human readability.", "Compressing the rulebook: a syntax or governance reference at AGGRESSIVE drops the examples that make it a reference; references are LOSSLESS." ] ===END===