v0.8.0

Full implementation of the five-tier roadmap from docs/cc_genui_20260420_200730_face_importance_ranking.html. Closes every Fixed-status row in the skill's collision table and adds foundation-precedence detection, directional resonance, and control-type composition reporting.

Added — v1 · Layer B (skill: task-aware interpretation)

• Prompt Refiner skill v0.6.1 bundled with the plugin (was user-profile only).
• Step 2.5 (Classify task type) — establishes task type and loads a per-task face priority template (11 categories including "Execution prompt inheriting plan" from field observations).
• Step 2.6 (Compute priority gaps) — replaces undifferentiated "neglected dimensions" with three lists: critical gaps (H priority, low activation), acceptable neglect (L priority, low activation), diffusion candidates (L priority, high activation). Working thresholds Low<0.2 / Mid 0.2-0.55 / High>0.55.
• Step 2.6b (Routing-collision suspicion) — detects inhabitation-isn't-moving-the-needle cases. Status table tracks fixed vs pending collisions.
• Step 2.7 (Pattern interventions) — library of 12+ prompt-engineering patterns mapped to primary face activation, cost-ordered.
• Task-Type Priority Reference and Pattern Intervention Reference tables.

Added — v2 · Layer A (engine: vocabulary disambiguation)

• 10 new DISAMBIGUATION_ENTRIES routing duty-bearing vocabulary (duty, owe, moral, virtue, culpable, forbidden, warrant, obligation, ought) to ethics when morally framed, and purpose to teleology when grounded in aim/goal/end context.
• Semantic bridge artifact rebuilt. Disambiguation triggers: 13 → 23; senses: 13 → 25.

Added — v3 · Foundation-precedence flag

• New math/precedence.py surfacing incoherent-stance patterns:
  • foundation_missing — evaluative face (ethics/axiology/teleology) dominant while both ontology and epistemology are below the Low threshold.
  • triad_cascade — within the classical triad (teleology → ethics → axiology), a downstream face dominates while an upstream face is at floor.
• Flags surfaced in construction_basis.precedence_flags and guidance.precedence_flags for both compact and focused outputs.
• 14 unit tests in tests/test_math/test_precedence.py.

Added — v4 · Directional resonance (ADR-014)

• CUBE_PAIR_DIRECTIONS in graph/schema.py — grounding face → grounded face mapping for all 6 cube pairs (ontology→praxeology, epistemology→methodology, ethics→axiology, teleology→heuristics, phenomenology→aesthetics, semiotics→hermeneutics).
• directional_resonance metric in math/harmonization.py — credits grounding face's coverage independent of grounded face density, resolving the field-observation "inherited plan" case where symmetric resonance under-reported legitimate grounding.
• Symmetric resonance retained unchanged for backward compatibility.
• grounding_face and grounded_face fields per harmonization pair.
• 10 new tests across schema integrity and metric behavior, including the inherited-plan pattern verification.
• docs/adr/014-directional-resonance.md.

Added — v5 · Control-type classification

• FACE_CONTROL_TYPE in graph/schema.py — each of the 12 faces classified as structural / bias / mixed:
  • structural (4): semiotics, methodology, ontology, praxeology.
  • bias (5): ethics, axiology, aesthetics, hermeneutics, heuristics.
  • mixed (3): epistemology, phenomenology, teleology.
• New math/control_type.py computes per-prompt composition (structural_share / bias_share / mixed_share / dominant_control_type).
• Per-face control_type annotation added to coordinate output.
• Aggregate control_type_composition surfaced at top level and in guidance.
• 15 unit tests in tests/test_math/test_control_type.py.

Changed

• Compact output extended with control_type_composition, precedence_flags, per-face control_type, and directional harmonization fields.
• Focused output includes precedence_flags and control_type_composition via guidance.
• Step 3, 5, 6 of the skill augmented with priority-aware reading, overlay checks, and new termination conditions.
• New "Vocabulary routing collisions are real" limitation in the skill.

Measurement

• 8-text literary benchmark: 17/20 → 18/20. Marx Communist Manifesto 2/3 → 3/3 (ethics rank #8 → #5). MLK I Have a Dream gained axiology in top-6.
• Field-observation validation (synthetic ethics-heavy prompt): ethics 0.10 → 0.702, axiology 0.60 → 0.352. The four-iteration routing collision is resolved on its specific class of prompts.
• Integration test: engineering prompt reports dominant_control_type: structural with 0 precedence flags; ethics-heavy prompt reports dominant_control_type: bias with a foundation_missing flag.
• Test suite: 327 → 366 passing (+39 new tests across precedence, directional resonance, and control type).

Context

• v1+v2 close the routing-collision and task-awareness gaps the external (ChatGPT) review and four-iteration field observation surfaced. v3 adds the foundation-precedence check (Principle 3). v4 makes cube pairs directional (Principle 4, ADR-014). v5 adds the structural-vs-bias distinction (Section 4 of the report) as explicit output.
• Remaining known limitations (documented in the skill's status table): semiotics↔hermeneutics vocabulary lockstep, aesthetics/semiotics "form/shape" ambiguity. Not addressed in 0.8.0; each needs its own disambiguation-entry pass with benchmark verification.

v0.7.0

[0.7.0] - 2026-04-16

Added

• BGE-large-en-v1.5 as sole embedding source at native 1024d (ADR-013). Replaces GloVe + Model2Vec paths. Runtime remains numpy-only; BGE is a build-time dependency under a new [build] optional extra (sentence-transformers, torch, wordfreq, nltk).
• wordfreq for frequency ordering and top-N vocabulary selection — no more GloVe dependency of any kind.
• FACE_VERNACULAR dictionary — targeted object-level vocabulary for underperforming faces (ethics, axiology, methodology) only. Added at 4x weight in face centroid construction.
• scripts/expand_pole_synonyms.py — WordNet-based pole-synonym expansion proposals for human review.
• ADR-013: rationale for BGE as the sole build-time embedding source.

Changed

• Build script rewritten (~2080 → ~770 lines). Deleted Model2Vec path, face-informed QR+PCA reduction, 400K→15K vocab trimming, OOV expansion, counter-fitting machinery. Single unified vocab.
• Questions and phrases are now encoded as full sentences through BGE, not as IDF-weighted word-vector averages.
• face_relevance uses IDF² weighting — rare face-specific vocabulary dominates the weighted average; common tokens no longer drown signal.
• 20% per-face column centering applied at artifact load — corrects systematic bias where "specific" face centroids (ethics, methodology) accumulated negative bias from generic tokens.
• Cube-pair contrast dampening disabled: the mechanism punished legitimate co-activations (ethics+axiology on moral rhetoric, epistemology+methodology on scientific texts). Re-enable only if evidence supports it.
• docs/specs/semantic-bridge-algorithms.md updated for BGE pipeline.

Measurement

• Benchmark: scripts/benchmark_8texts.py score improved from 13/20 → 17/20 (+31%).
• Aesthetics moved from rank #10 to #1 on Aristotle's Poetics; methodology now hits top-6 on Newton; phenomenology-dominates-everything ceiling broken.

Removed

• GloVe download/loading code paths.
• Model2Vec integration and PCA reduction machinery.
• select_runtime_vocab 400K→15K trimming (vocab is assembled directly from what's needed).
• expand_vocab_by_question_proximity OOV expansion pass.
• Counter-fitting SGD (retired pending evidence that contextual embeddings still benefit).

v0.6.0

Added

• Synthesis/guidance layer in construction bridge (Stage 8): dominant dimensions, neglected dimensions with gap statements, strongest harmonization pair with cube pair concern, plain-language summary.
• Focused output mode: create_prompt_basis(focused=true) returns ~3KB guidance-centric output instead of ~695KB full pipeline data.
• interpret_basis MCP tool: takes construction basis JSON and returns markdown-formatted interpretation with dominant dimensions, gaps, and strongest resonance.
• ape://examples MCP resource: 4 annotated example intents demonstrating engine output interpretation.
• Per-face action guidance in ape://axiom_manifest resource: "If strongly activated: X. If weak: consider Y."
• docs/GEOMETRY-NOTES.md: latent polyhedral properties and structural capacities.
• 27 new tests: phase weighting, question position, disambiguation, phrase detection, MCP tool handlers.

Changed

• MCP prompts rewritten with face definitions, interpretation guidance, and usage instructions.
• External surface: 4 tools (was 3), 4 resources (was 3), 4 prompts.
• 327 tests (up from 300).

Fixed

• Dead causal propagation code removed (~40 lines).
• Float equality guard in gem computation (== 0 → < 1e-10).
• Raw discriminative scores captured before phase mutation in intent parser.
• Test fixture expanded SYMMETRIC_RELATIONS for production graph parity.
• Consistent ALL_FACES fallback in face_relevance no-match paths.

v0.5.0

Added

• Phase-aware face weighting: 3 phase centroids (comprehension/evaluation/application) provide 30% modulation on face relevance scores. Pre-computed per-word phase similarity matrix.
• Expanded contextual disambiguation: 15 trigger words with 15 context-aware senses (up from 8/10). Covers physics→methodology, drama→aesthetics, rhetoric→semiotics contexts.
• Per-face question position matching: pre-computed per-word best-matching question within each face. Phase 2 blends axis projection (40%) with question-matched position (60%).
• Question-guided vocabulary expansion: 15K→20K words selected by proximity to 1728 construction question embeddings (not generic frequency). Covers literary/archaic terms like "imitation", "catharsis", "sovereignty".
• Benchmark script (scripts/benchmark_8texts.py): 8-text literary benchmark across Shakespeare, Genesis, Marx, MLK, Newton, Aristotle, Tao Te Ching, Descartes.
• Model2Vec integration (conditional): if distilled sentence transformer exists at build time, uses Model2Vec vectors instead of GloVe. Currently disabled (GloVe outperforms for this domain).
• Semantic bridge algorithm specification (docs/specs/semantic-bridge-algorithms.md).
• /pre-release and /release project commands.
• /research-semantic-improvement research command.

Changed

• Contrastive cube-pair dampening: 30% score transfer within each complementary pair, enforcing theoretical/applied distinction.
• Synonym decontamination: refined pole synonyms for Ethics, Aesthetics, Teleology, Axiology, Phenomenology, Praxeology to reduce cross-face vocabulary overlap.

Performance

• Literary text benchmark: 18/20 expected faces in top 6 (up from 14/20 at v0.4.0 keyword-only baseline).
• Remaining 2 misses (MLK teleology, Newton methodology) are at the fundamental limit of word-level static embeddings.
• 300 tests passing (up from 261).

v0.4.0

Added

• Geometry-integral intent parser: the parser IS the Construct's inference machinery running in reverse. Phase 1 projects intent onto face centroids (discriminative cosine similarity). Phase 2 projects onto axis direction vectors (high_pole - low_pole). Phase 3 maps scalars to grid via polarity convention.
• GloVe 6B 100d semantic bridge: 15K-word vocabulary with pre-computed face similarity and axis projection matrices. Built at dev time, ships as ~2MB numpy artifacts. Zero ML runtime dependencies.
• Contextual disambiguation table: 8 polysemous trigger words (state, compelled, right, deep, forces, heaven, tragedy, action) with 10 context-aware senses. Overrides face/axis scores when >= 2 context indicator words are present.
• N-gram phrase embeddings: 92 curated phrases (domain replacements, pole pair bigrams, philosophical key phrases). Greedy longest-match tokenizer with surface-to-canonical mapping.
• Contrastive cube-pair dampening: within each complementary pair, transfers 30% of score difference from weaker to stronger face.
• 46 pole synonym clusters (~320 curated words) for GloVe centroid construction.
• GeometricBridge class (replaces SemanticBridge): face_relevance() + axis_projection() with disambiguation and phrase support.
• 39 new tests for GeometricBridge and geometry-integral parser.
• Semantic bridge algorithm specification (docs/specs/semantic-bridge-algorithms.md).
• Build script (scripts/build_semantic_bridge.py) with pole self-test validation (all 24 axes pass).

Changed

• Intent parser no longer uses TF-IDF or keyword matching for face/position selection — replaced by geometric projection.
• Tokenizer is now unstemmed (GloVe needs word forms) with expanded stop-word list (~95 words).
• Face centroids built from authored layers only (core questions + sub-dimension labels + pole synonyms) — NOT from derived question templates.

Removed

• Keyword-based face matching (_FACE_KEYWORDS dictionary).
• TF-IDF dependency in Stage 1 (intent parser no longer queries construct questions).
• Stemming in intent parser tokenizer.

Performance

• Literary text benchmark: 15/20 expected faces in top 6 across 8 benchmark texts (Shakespeare, Bible, Marx, MLK, Newton, Aristotle, Tao Te Ching, Descartes).
• 300 tests passing.

v0.3.0

Changed (breaking)

• 12-face rebuild: Complete reconstruction from 10-branch/10×10 to 12-face/12×12 architecture.
• 12 philosophical domains (added Ethics and Aesthetics as distinct faces; Axiology reframed as evaluative theory).
• 12×12 grids (144 points per face, 1728 total constructs, up from 1000).
• Invariant axis meta-meaning: x = constitutive character, y = dispositional orientation.
• Polarity convention: low = constrained/foundational, high = expansive/exploratory on all axes.
• "Face" terminology replaces "branch" throughout (backward-compat aliases retained).
• Graph: 1873 nodes, 2279 edges (up from 1101/1696).
• 66 unique nexus pairs (up from 45), 132 directional gems (up from 90), 12 spokes (up from 10).

Added

• Cube pairing model: 6 complementary pairs with harmonization through shared surfaces.
• Nexus stratification: 6 paired + 48 adjacent + 12 opposite = 66, with tier-modulated computation.
• Positional correspondence: inter-face relationships via shared coordinate system (replaces 237 declared cross-face edges).
• Harmonization module: paired face resonance scoring with bidirectional positional alignment.
• Weight-modulated activation radius in construct resolver (face relevance affects activation footprint).
• Gem magnitude incorporates positional correspondence + cube tier modulation.
• Central gem coherence uses coefficient of variation for spoke differentiation.
• Meaning hierarchy in output: corners → integration, edges → demarcation, midpoints → axial_balance, center → composition.
• Causal phase classification: comprehension (1-5), evaluation (6-7), application (8-12).
• 144 authored question templates across 5 zones (corners, edge midpoints, other edge, near-edge interior, deep interior).
• 66 nexus definitions (21 new for Ethics/Aesthetics pairs).
• 66 first-principles compliance tests covering all 12 architectural principles.
• CONSTRUCT-v2.md specification (18 core commitments, 15 sections).
• CONSTRUCT-v2-questions.md (all 144 templates).
• Work effort triad (Roadmap, WorkEffort, Results).

Removed

• 237 declared cross-face point-to-point edges (replaced by positional correspondence).
• 5 graph algorithm modules: spectral embedding, community detection, centrality, CSP, distance (v2 geometry is regular — computation is coordinate math).
• v1 specification files: CONSTRUCT.md, CONSTRUCT-INTEGRATION.md, 12 specs/ files.
• Spectrum questions and revisited questions (v2 derives spectrum meaning from sub-dimensions).