diff --git a/docs/audit/2026-05-stage-label-reconciliation.md b/docs/audit/2026-05-stage-label-reconciliation.md
index 7d8f67e4..b464af10 100644
--- a/docs/audit/2026-05-stage-label-reconciliation.md
+++ b/docs/audit/2026-05-stage-label-reconciliation.md
@@ -19,7 +19,12 @@ forward.
 transport refactor · `S3` Guitar Pro import · `S4` phrase boundary detection ·
 `S5` corpus + schema · `S6` rule generator v0 · `S7` graph layer · `S8`
 preview app · `S9` feedback layer · `S10` CLAP MVP · `S11` region
-regeneration · `S12` neural assistance.
+regeneration · `S12` neural assistance · `S13` complementary part generation ·
+`S14` structure controls and metrics · `S15` tonal context and harmonic control.
+
+S13 and later stages are append-only additions. Their logical dependency order
+may place them beside or before an earlier-numbered future stage; existing stage
+numbers are never repurposed or renumbered to make the list look prettier.
 
 ## Mislabel mapping
 
@@ -43,14 +48,17 @@ baseline**. It is the input to:
 
 ## Rules from here
 
-1. New commits use canonical `S0…S12` numbering per the glossary only.
+1. New commits use canonical stage numbers per the glossary and this roadmap.
 2. A stage is "closed" only when its acceptance criterion (its stage doc) is
    met, tested, and documented.
 3. Earlier mislabeled work is **not** re-closed under its old number; relevant
    pieces are re-credited to their true stage in the stage docs' "See also".
+4. New cross-cutting stages take the next free number. Existing S-numbers are
+   never reassigned, even when the logical execution order differs.
 
 ## See also
 
 - [`../glossary.md`](../glossary.md) §0, §17
 - [`../SPEC.md`](../SPEC.md)
 - [`../stages/S0-baseline-and-tests.md`](../stages/S0-baseline-and-tests.md)
+- [`../stages/S15-tonal-context-and-harmonic-control.md`](../stages/S15-tonal-context-and-harmonic-control.md)
diff --git a/docs/audit/2026-07-symbolic-harmony-and-evolution-research.md b/docs/audit/2026-07-symbolic-harmony-and-evolution-research.md
new file mode 100644
index 00000000..fad8ff38
--- /dev/null
+++ b/docs/audit/2026-07-symbolic-harmony-and-evolution-research.md
@@ -0,0 +1,238 @@
+# Research: symbolic harmony, global paths, and interactive evolution (2026-07)
+
+Status: accepted as roadmap input; no dependency adoption
+Decision: map the useful ideas onto S7/S8/S9/S12/S15 rather than create a
+repository-per-stage roadmap
+
+## Executive verdict
+
+The strongest common pattern is:
+
+```text
+enumerate feasible local possibilities
+→ assign explainable local / transition costs
+→ optimise globally
+→ return diverse top-k paths
+→ let the human choose
+→ retain feedback and lineage
+```
+
+The responsibilities are intentionally split:
+
+- **S15** defines tonal/harmonic meaning and uncertainty;
+- **S7** optimises multi-step paths;
+- **S8** displays alternatives and provenance;
+- **S9** captures selection and experiments with structural evolution;
+- **S12** remains the owner of future neural assistance.
+
+No reviewed repository is adopted as a production dependency by this decision.
+
+## Source assessment
+
+| Source | Value now | Later | What to take | Owner |
+|---|---:|---:|---|---|
+| [`ekzhang/harmony`](https://github.com/ekzhang/harmony) | 8/10 | 8/10 | layered DP, transition costs, path reconstruction | S7 |
+| [`napulen/romanyh`](https://github.com/napulen/romanyh) | 8/10 | 9/10 | alternative harmonisations, k-best global paths, RomanText fixtures | S7 + S15 |
+| [`napulen/AugmentedNet`](https://github.com/napulen/AugmentedNet) | 4/10 | 9/10 | decomposed harmonic targets, uncertainty, synthetic labelled examples | S15 + S12 radar |
+| [`perfect-shuffle-music`](https://github.com/kroger66/perfect-shuffle-music) | 6/10 concept | 7/10 | human-in-the-loop generations, crossover concept, lineage | S9, surfaced by S8 |
+| [`napulen/harmalysis`](https://github.com/napulen/harmalysis) | 4/10 | 7/10 | external harmonic fixture DSL inspiration | S15 |
+| [`ekzhang/composing.studio`](https://github.com/ekzhang/composing.studio) | 3/10 | 6/10 | textual playground, live preview/playback, inspectable edits | S8 |
+| [`ekzhang/crepe`](https://github.com/ekzhang/crepe) | 5/10 | 6/10 | possible declarative rule engine when rule closure/explanations justify it | S7/S15 radar |
+
+The ratings describe architectural relevance to `griff`, not general project
+quality.
+
+## 1. `ekzhang/harmony`: global path shape
+
+The useful lesson is not classical four-part harmony itself. It is the layered
+optimisation shape:
+
+1. enumerate valid states for each musical position;
+2. calculate local and adjacent-state costs;
+3. use dynamic programming to select a globally coherent path;
+4. reconstruct the selected path with an inspectable total cost.
+
+This is directly relevant to S7's multi-bar candidate chain. Potential Griff
+transition terms include:
+
+- phrase / contour continuity;
+- register continuity;
+- rhythm continuity or complement;
+- repeated-technique penalties;
+- playability / fret travel;
+- harmonic fit supplied by S15;
+- repetition and mud penalties.
+
+Do **not** copy SATB-specific prohibitions or doubling/resolution rules as
+swancore policy. Transfer the algorithmic form; make costs Griff-specific and
+explainable.
+
+The earlier idea of making a new generic `RegisterPlanner` the first client is
+superseded: the register track is already accepted and closed after generator
+semantics were repaired. Reopen register planning only when a new measured
+multi-bar problem justifies it.
+
+## 2. `napulen/romanyh`: k-best paths
+
+The important extension is returning several ranked global alternatives rather
+than one optimum. For Griff this supports:
+
+- globally coherent alternatives without seed-only noise;
+- several complementary-guitar trajectories;
+- controlled register/harmonic alternatives;
+- human selection feeding S9;
+- explicit diversity constraints and fixed tie-breaking.
+
+S7 should first extract a concrete layered-path contract from a real multi-bar
+client, then add deterministic k-best enumeration. Avoid a universal
+`MusicDPGodObject`; specialised state/cost modules share only the small engine.
+
+## 3. Nápoles ecosystem: harmonic decomposition and fixtures
+
+### `AugmentedNet`
+
+Do not introduce its Python/TensorFlow/MusicXML runtime into Griff. Useful ideas:
+
+- decompose harmonic analysis into tonal centre, mode, chord root, quality,
+  inversion, and related targets;
+- expose distributions/alternatives rather than one unquestioned label;
+- create labelled synthetic examples and vary their texture.
+
+S15 should grow incrementally from its accepted tonic/mode estimate. Full Roman
+numeral analysis is not the next task.
+
+### `harmalysis` and RomanText
+
+Use a small text language only as an external fixture/debugging surface:
+
+```text
+C: I | vi | IV | V
+C: I | V/V | V | I
+a: i | VI | III | VII
+```
+
+Core representation stays typed. The fixture pipeline can generate
+transpositions, omissions, inversions, passing tones, pedal textures, and
+modulations with known labels for calibration tests.
+
+### Audio key detection
+
+Audio chroma/HMM approaches are deferred. Griff currently receives symbolic
+MIDI/Guitar Pro evidence; rendering it to audio/chroma and guessing the lost
+symbolic information back would be an avoidable information-loss loop. Revisit
+only with a real audio-input stage.
+
+## 4. `perfect-shuffle-music`: concept, not code
+
+The implementation is too note-array-centric for Griff. Do not port its genome
+or crossover code.
+
+The useful experiment is structural evolution over Griff objects:
+
+```text
+candidate set
+→ human selects parents
+→ bar/motif/parameter crossover
+→ mutation
+→ normalisation and validators
+→ existing rerank as a safety/quality guard
+→ next generation
+```
+
+Candidate operators for an S9 experiment:
+
+```rust
+pub enum EvolutionOperator {
+    AlternateBars,
+    AlternateMotifs,
+    RhythmFromAContourFromB,
+    PrefixSuffixCrossover,
+    ParameterBlend,
+}
+```
+
+Lineage must record parents, operator, mutations, generation, and session. Track
+population collapse, strategy diversity, operator survival, and repeated-parent
+dominance. This remains an **Evolution Lab** experiment under S9; create a new
+stage only if it later becomes a standalone persistent workflow.
+
+## 5. `composing.studio`: S8 playground direction
+
+The transferable UI idea is editable text plus immediate visual/audio feedback.
+A Griff playground should expose:
+
+- textual request / constraints / future harmonic fixture input;
+- piano roll / tab / playback;
+- candidate list and score axes;
+- register, novelty, playability, and tonal diagnostics;
+- path and transition explanations;
+- like/dislike/favourite controls;
+- parent/child lineage for the Evolution Lab.
+
+S8 supplies the surface; it does not own tonal inference, path optimisation, or
+preference semantics.
+
+## 6. `crepe`: parked implementation option
+
+A Datalog-like rule engine may become useful when Griff has all of:
+
+- many independently maintained rules;
+- recursive graph closure;
+- a need to explain derivations;
+- frequent rule additions that make imperative orchestration brittle.
+
+Until then, typed Rust remains simpler. A pleasant macro is not evidence that the
+generator needs an expert-system runtime.
+
+## Roadmap mapping
+
+### S7 — graph layer
+
+- concrete layered-path contract from the first multi-bar client;
+- global candidate-chain DP/Viterbi;
+- deterministic k-best diverse paths;
+- later harmonic/complement/cadence clients using S15 states.
+
+### S8 — preview app / cockpit
+
+- textual playground;
+- candidate, tonal, path, and lineage inspectors;
+- UI for S9 feedback/evolution experiments.
+
+### S9 — human feedback
+
+- feedback capture and preference reranking first;
+- Evolution Lab second;
+- diversity/collapse controls before productisation.
+
+### S12 — neural assistance
+
+- AugmentedNet/MiniBach remain decomposition and synthetic-data research inputs;
+- no runtime model before the existing corpus, baseline, and feedback gates.
+
+### S15 — tonal context and harmonic control
+
+- Phase 0 evidence audit: accepted/closed;
+- Phase 1 shared tonal core: accepted/closed;
+- Phase 2 explicit scoped context: next;
+- later calibration, fixture DSL, soft harmonic generation, local context and
+  cadence.
+
+## Immediate actions
+
+1. Maintain S15 as the owner of tonal meaning and confidence.
+2. Implement S15 Phase 2 without changing generated output.
+3. Write the S7 layered-path design against a concrete multi-bar candidate-chain
+   client before extracting a generic engine.
+4. Add harmonic synthetic-fixture planning to S15 Phase 3/4.
+5. Add Evolution Lab to the S9 backlog, not the production API.
+6. Add the textual/candidate/tonal/lineage playground to the S8 backlog.
+
+## Explicit non-decisions
+
+- no automatic highest-margin track selection;
+- no hard inferred-scale pitch whitelist;
+- no generic register planner without new evidence;
+- no classical harmony rule transplant;
+- no neural or Datalog dependency;
+- no new Evolution stage yet.
diff --git a/docs/stages/S12-neural-assistance.md b/docs/stages/S12-neural-assistance.md
index 6b705934..83532c24 100644
--- a/docs/stages/S12-neural-assistance.md
+++ b/docs/stages/S12-neural-assistance.md
@@ -22,17 +22,38 @@ from scratch".
 - Models suited to fixed-region control / infilling (e.g. anticipatory /
   masked-LM style on MIDI tokens).
 - Inference in Rust via `candle` or ONNX (`ort`); training offline.
+- Harmonic-analysis research such as AugmentedNet is useful for **task
+  decomposition** (tonal centre, mode, chord root, quality, inversion) and
+  synthetic labelled examples, not as a Phase-1 runtime dependency.
+- Symbolic generation projects such as MiniBach remain research references only;
+  they do not replace the S6/S7 deterministic baseline or S15 calibrated tonal
+  context.
 
 ## Preconditions (hard gate)
 
 - Corpus ≥ ~100 phrases; working S6 baseline; S9 feedback integration.
 - Not started before all three hold (glossary §17.5).
+- Any harmonic neural proposal must also compare against the accepted S15
+  symbolic estimator/fixtures and preserve uncertainty/abstention rather than
+  emit one unquestioned label.
 
 ## Acceptance criteria
 
 - Neural suggestions beat S6/S7 on the quality score in a blind comparison.
 - Deterministic given a fixed seed and model checkpoint.
+- Harmonic assistance, when attempted, beats or complements the S15 symbolic
+  baseline on a labelled test set and reports calibrated uncertainty.
+
+## Non-goals
+
+- No Python/TensorFlow/MusicXML service in the core runtime merely because a
+  research implementation uses that stack.
+- No rendering symbolic input to audio/chroma and guessing its lost semantics
+  back while native symbolic evidence is available.
 
 ## See also
 
+- [`../audit/2026-07-symbolic-harmony-and-evolution-research.md`](../audit/2026-07-symbolic-harmony-and-evolution-research.md)
+- [`S9-feedback-layer.md`](S9-feedback-layer.md)
+- [`S15-tonal-context-and-harmonic-control.md`](S15-tonal-context-and-harmonic-control.md)
 - [`../glossary.md`](../glossary.md) §8, §17.5
diff --git a/docs/stages/S13-complementary-part-generation.md b/docs/stages/S13-complementary-part-generation.md
index a1fdadc3..2fda7157 100644
--- a/docs/stages/S13-complementary-part-generation.md
+++ b/docs/stages/S13-complementary-part-generation.md
@@ -22,6 +22,13 @@ ADRs: ADR-0012, ADR-0011
 > `ConstrainedRandomWalk` line over a request derived from A; typed
 > `NonUniformTimeline` on mid-score meter changes). `ModeNotImplemented` is
 > gone — every `RelationMode` arranges.
+>
+> Ownership update (2026-07): the original private winner-only key estimate was
+> generalised into the accepted shared `core::tonal` layer. S13 keeps the
+> backwards-compatible `PartProfile::harmony` projection and complement-specific
+> use of it; **S15** owns scoped tonal evidence, uncertainty, calibration,
+> harmonic fixtures, and future soft harmonic/cadence policy. **S7** owns any
+> later multi-bar global optimisation or k-best complementary trajectories.
 
 ## Remaining work (follow-up increments)
 
@@ -42,7 +49,8 @@ ADRs: ADR-0012, ADR-0011
       `PartProfile::harmony` (tonic, major/natural minor, `scale_fit` as a
       fact); B's substitution material becomes A's literal pitch classes
       *plus* the inferred key's scale, so a sparse part no longer collapses
-      B onto one pitch per band (2026-06-11).
+      B onto one pitch per band (2026-06-11). The estimator now delegates to
+      S15's shared `core::tonal` implementation.
 
 > Roadmap note: appended as the next free stage number (append-only, per the
 > stage-label audit). Logically it sits between the single-part generator (S6)
@@ -103,7 +111,7 @@ Generative-first: B is derived from A by rule; no corpus pair mining. `ChunkMeta
 ## Open questions
 
 - Harmonic-compatibility thresholds (allowed coincident intervals) — calibrate
-  on the corpus.
+  on the corpus and S15 labelled fixtures.
 - Default per-axis ratios (e.g. support-layer density factor) before S9 feedback
   exists to tune them.
 - Where the part profile lives: extend the feature layer vs a dedicated
@@ -111,6 +119,8 @@ Generative-first: B is derived from A by rule; no corpus pair mining. `ChunkMeta
 
 ## See also
 
+- [`S15-tonal-context-and-harmonic-control.md`](S15-tonal-context-and-harmonic-control.md)
+- [`../audit/2026-07-symbolic-harmony-and-evolution-research.md`](../audit/2026-07-symbolic-harmony-and-evolution-research.md)
 - [`../glossary.md`](../glossary.md) §8 (Complementary part, Complement relation,
   Relation mode), §9 (Complement hyperedge), §10 (Relation preference)
 - [`../adr/0012-complementary-part-generation.md`](../adr/0012-complementary-part-generation.md)
diff --git a/docs/stages/S15-tonal-context-and-harmonic-control.md b/docs/stages/S15-tonal-context-and-harmonic-control.md
new file mode 100644
index 00000000..4f32abde
--- /dev/null
+++ b/docs/stages/S15-tonal-context-and-harmonic-control.md
@@ -0,0 +1,214 @@
+# S15: Tonal context and harmonic control
+
+Status: in progress — Phase 0 (evidence audit) and Phase 1 (shared tonal core)
+accepted and closed on 2026-07-12; Phase 2 is next
+Depends on: S1 (canonical score), S5 (corpus), S6 (rule generator)
+Builds on: S13 harmonic-context analysis
+Feeds: S6 generation, S7 graph costs, S11 regeneration, S13 complement
+
+## Goal
+
+Make tonal and harmonic context explicit, uncertain, scoped, and reusable before
+it is allowed to influence generation. Separate observed evidence from inferred
+musical meaning; preserve an honest abstention path; never turn one best guess
+into a hard seven-note whitelist.
+
+S15 owns the meaning of tonal/harmonic states. S7 owns global path optimisation
+over such states. S8 displays the alternatives and provenance. S9 learns from
+human choices among them.
+
+## Guardrails
+
+- A real-song key result is a tonal hypothesis for one `EvidenceScope`, not
+  verified ground truth.
+- Confidence thresholds and automatic scope selection require calibration; a
+  larger margin alone does not prove that one track is the correct reference.
+- The observed `PitchClassSet` remains distinct from an inferred scale or tonal
+  hierarchy. Chromatic passing tones, borrowed notes, and tensions are not
+  automatically errors.
+- `None` / ambiguous context is a valid result. No silent fallback to C major,
+  the lowest pitch, or the highest-margin track.
+- Cadence and generation integration remain frozen until their phase-specific
+  acceptance gates are met.
+
+## Phase 0 — evidence audit and diagnostics ✅
+
+Accepted and closed.
+
+- Audited the existing generation input and established that
+  `PitchMaterial.root` was an anchor derived from the minimum input pitch, not a
+  tonic.
+- Measured `WholeScore`, `Track`, and `Voice` evidence on real and synthetic
+  inputs.
+- Established raw evidence vocabulary: onset counts, duration mass, note count,
+  and observed pitch range.
+- Demonstrated that scope can change the winning tonal hypothesis (including the
+  Wolf & Bear whole-score vs track conflict).
+- Rejected confidence cut-offs inferred from the small diagnostic fixture set.
+
+Primary record:
+[`../audit/2026-07-tonal-context-phase0.md`](../audit/2026-07-tonal-context-phase0.md).
+
+## Phase 1 — shared evidence/inference core ✅
+
+Accepted and closed.
+
+`core/src/tonal.rs` now provides:
+
+- `EvidenceScope::{WholeScore, Track, Voice}`;
+- `PitchEvidence::measure` with raw integer `onset_counts`, `duration_mass`,
+  `note_count`, and observed `feature::PitchRange`;
+- `estimate_key` returning all 24 major/natural-minor `TonalCandidate`s,
+  best-first, with per-candidate correlation and `scale_fit`;
+- `TonalEstimate::confidence_margin` as winner minus runner-up;
+- duration-only KS v1, with onset-count fallback only when total duration mass is
+  zero.
+
+`complement::estimate_harmony` delegates to the shared inference core and keeps
+its public winner-only projection. Focused validation proved:
+
+- `HarmonicContext`: 16/16 exact, 0 changed;
+- structure consumer: 7/7 byte-identical;
+- evidence mapping: 39/39, 0 mismatches;
+- histogram additivity: pass;
+- 24 finite candidates per non-empty scope: pass;
+- generation smoke: 30/30 byte-identical.
+
+Cloud implementation: `6f9114d` (red), `184b586` (green), `e2c9c7f` (docs),
+`af26206` (accepted/closed). Local validation: `bd2c7c8`; archival:
+`3993bb0`.
+
+## Phase 2 — explicit scoped context contract (next)
+
+Allow generation-facing requests and provenance to carry an optional, explicit
+scoped tonal estimate **without changing note selection yet**.
+
+The exact type is a design output, not pre-decided, but it must preserve:
+
+- the chosen `EvidenceScope`;
+- the ranked estimate or an intentionally compact immutable projection;
+- absence / ambiguity;
+- deterministic serialisation and replay;
+- provenance identifying how the estimate was measured.
+
+Acceptance:
+
+- requests without tonal context remain byte-identical to the Phase-1 baseline;
+- context is optional and scope is explicit;
+- no automatic whole-score/track/voice selection;
+- no pitch restriction, reranker-weight change, cadence, or production behaviour
+  change;
+- round-trip and deterministic replay tests cover the new contract.
+
+## Phase 3 — scope policy and confidence calibration
+
+### Phase 3A — scope-selection experiments
+
+Compare explicit policies rather than silently choosing the largest margin:
+
+- selected/reference track;
+- whole score;
+- guitar-only subsets;
+- weighted combinations of tracks;
+- multiple competing scope estimates carried together.
+
+A repetitive bass pedal or ostinato may yield a strong margin while describing
+only one layer, so `argmax(margin)` is not an approved policy.
+
+### Phase 3B — confidence calibration and synthetic controls
+
+Build labelled, programmatic controls covering:
+
+- exactly flat chromatic material;
+- diatonic and pentatonic material;
+- pedal tones and omitted tones;
+- borrowed notes and secondary dominants;
+- modal ambiguity;
+- modulations / tonicisations;
+- transpositions and alternative textures.
+
+Report error and abstention behaviour by scope/material class. Only then may a
+stable confidence vocabulary or threshold be proposed.
+
+## Phase 4 — harmonic fixture DSL
+
+Introduce a small external fixture language inspired by RomanText / harmonic
+analysis DSLs, for example:
+
+```text
+C: I | vi | IV | V
+a: i | VI | III | VII
+C: I | V/V | V | I
+```
+
+The text format is for fixtures, debugging, and synthetic corpus generation.
+Core uses typed structures (`Degree`, `HarmonicFunction`, chord quality,
+inversion, modulation/tonicisation); parser strings are not the domain model.
+
+Acceptance:
+
+- scripts transpose deterministically;
+- scripts generate labelled symbolic fixtures with multiple textures;
+- parser failures are typed and localised;
+- no runtime dependency on Python, TensorFlow, MusicXML, or an external analysis
+  service.
+
+## Phase 5 — soft harmonic generation policy
+
+Allow calibrated tonal/harmonic information to influence candidate generation or
+ranking as a **soft preference**:
+
+- observed pitch classes stay legal unless an explicit user constraint says
+  otherwise;
+- inferred hierarchy may reward chord/scale tones and controlled resolutions;
+- ambiguous estimates abstain;
+- borrowed/chromatic colour tones remain representable;
+- A/B evaluation covers harmonic fit, closure, novelty, rhythm, register, and
+  playability together.
+
+No hard `inferred scale == allowed notes` shortcut.
+
+## Phase 6 — local context and cadence
+
+Move the remaining S6 cadence-aware-ending backlog here. Cadence requires local
+section context, phrase boundaries, calibrated confidence, and an abstention
+path; a global winner plus `last_note = tonic` is not a cadence model.
+
+Candidate ending states and explainable resolution costs may be optimised through
+the S7 layered-path engine once both stages provide stable contracts.
+
+Acceptance:
+
+- section-local context beats global-only context on defined fixtures;
+- ambiguous/modulating regions may decline to force a cadence;
+- endings are returned as ranked alternatives with explanations;
+- context-free generation remains unchanged.
+
+## Research inputs
+
+See
+[`../audit/2026-07-symbolic-harmony-and-evolution-research.md`](../audit/2026-07-symbolic-harmony-and-evolution-research.md).
+The main inputs are:
+
+- `ekzhang/harmony` and `napulen/romanyh`: layered DP, transition costs, and
+  k-best global alternatives (algorithmic shape shared with S7);
+- `napulen/AugmentedNet`: decomposed harmonic targets and synthetic labelled
+  examples, not a runtime dependency;
+- `napulen/harmalysis` / RomanText: fixture-language inspiration.
+
+## Non-goals
+
+- No classical SATB rules copied wholesale into swancore guitar generation.
+- No neural runtime dependency in S15; neural assistance remains S12.
+- No audio chroma/HMM round-trip while the source is already symbolic.
+- No generic `MusicDPGodObject`; S7 owns a small path engine with separate
+  state/cost clients.
+
+## See also
+
+- [`S6-rule-generator-v0.md`](S6-rule-generator-v0.md)
+- [`S7-graph-layer.md`](S7-graph-layer.md)
+- [`S8-preview-app.md`](S8-preview-app.md)
+- [`S9-feedback-layer.md`](S9-feedback-layer.md)
+- [`S13-complementary-part-generation.md`](S13-complementary-part-generation.md)
+- [`../audit/2026-07-tonal-context-phase0.md`](../audit/2026-07-tonal-context-phase0.md)
diff --git a/docs/stages/S6-rule-generator-v0.md b/docs/stages/S6-rule-generator-v0.md
index 842598b1..82a60f3e 100644
--- a/docs/stages/S6-rule-generator-v0.md
+++ b/docs/stages/S6-rule-generator-v0.md
@@ -10,9 +10,21 @@ ADRs: ADR-0005, ADR-0010
 > policy (ADR-0017; melodic-closure note §7.2/§7.3). `griff generate` uses it
 > by default, and `--corpus <dir>` feeds rhythm templates, novelty
 > references, and the burst/rest gesture ask from curated chunks
-> (decisions.log 2026-07-11). Still open from the list below: cadence-aware
-> endings inside the strategies, anchor preservation, the string/fret
-> playability filter, and the density/syncopation corpus gates.
+> (decisions.log 2026-07-11).
+>
+> Post-acceptance hardening (2026-07-12): the register track is **accepted and
+> closed**. `ScaleLadder` provides full-range reachability; Shuffle uses an
+> unbiased ≤-octave `LadderWindow`; RhythmCopy traverses reflectively without
+> wrapping; RepeatVariation chooses an endpoint local to the bar's actual
+> penultimate degree. Focused corpus/synthetic validation found no remaining
+> target-strategy >12-semitone jumps. A generic register rerank axis and global
+> `RegisterPlan` were not justified and remain rejected/not started.
+>
+> Remaining follow-ups from the list below: anchor preservation, the string/fret
+> playability filter, and the density/syncopation corpus gates. Cadence-aware
+> endings are no longer owned as an unscoped S6 patch: they move to **S15 Phase
+> 6**, after explicit tonal context, scope/confidence calibration, and an
+> abstention path exist.
 
 ## Goal
 
@@ -33,7 +45,8 @@ swancore-like riffs.
    off-style density).
 4. Shuffle motifs grouped by tags.
 5. Repeat + variation (call/response — replace last beat).
-- Cadence-aware endings; anchor preservation; string/fret playability filter.
+- Anchor preservation; string/fret playability filter.
+- Cadence-aware endings are deferred to S15 Phase 6.
 
 ## Acceptance criteria
 
@@ -50,5 +63,6 @@ swancore-like riffs.
 
 ## See also
 
+- [`S15-tonal-context-and-harmonic-control.md`](S15-tonal-context-and-harmonic-control.md)
 - [`../glossary.md`](../glossary.md) §8
 - [`../fuzzing.md`](../fuzzing.md) (`generation_request`, P2)
diff --git a/docs/stages/S7-graph-layer.md b/docs/stages/S7-graph-layer.md
index 8f9cc412..83d50820 100644
--- a/docs/stages/S7-graph-layer.md
+++ b/docs/stages/S7-graph-layer.md
@@ -21,6 +21,13 @@ multi-technique with evidence; supersedes ADR-0014)
 > idea (a)). Nodes, transition / co-occurrence edges, complement hyperedges,
 > and the DP/Viterbi traversal remain gated on S6 acceptance and corpus
 > scale.
+>
+> Research update (2026-07): `ekzhang/harmony` and `napulen/romanyh` reinforce
+> the stage's existing architecture: enumerate feasible states per layer,
+> calculate explainable local/transition costs, optimise globally, reconstruct
+> the path, and later return deterministic k-best alternatives. The algorithmic
+> form is adopted as roadmap input; their classical SATB rules and runtime code
+> are not.
 
 ## Goal
 
@@ -33,7 +40,8 @@ connected / possible); DP/Viterbi is the *route* (which sequence is best).
 - In: corpus chunks + features (≥ ~100 phrases recommended before this pays
   off).
 - Out: a (hyper)graph (nodes + edges) and a DP/Viterbi traversal producing the
-  optimal candidate chain.
+  optimal candidate chain and, after the first path is validated, ranked k-best
+  alternatives.
 
 ## Approach
 
@@ -48,11 +56,47 @@ connected / possible); DP/Viterbi is the *route* (which sequence is best).
   approximation for graphs too large for exact DP.
 - DP state carries running context: current candidate, fretboard position and
   last technique (ADR-0018 — the rich note model makes both expressible),
-  `EnergyState`, rhythmic similarity to part A.
+  `EnergyState`, rhythmic similarity to part A, and optional S15 harmonic state
+  once that contract is calibrated.
 - Cost function (inspectable, the same weights S9 later tunes):
   `harmonic_fit + rhythm_complement + style_fit + playability + phrase_continuity
   − mud_penalty − repetition_penalty − fret_jump_penalty`.
 
+## Planned slices
+
+### Slice A — concrete layered-path contract
+
+Extract the smallest reusable path contract from a real multi-bar client, not
+from a speculative universal framework. A layer exposes feasible states; the
+client supplies local and transition costs plus explanations; the engine returns
+the deterministic best path.
+
+The first preferred client is a multi-bar `GenerationCandidate` chain. The
+already-accepted register track is **not** reopened merely to manufacture a
+first generic client.
+
+### Slice B — multi-bar global candidate chain
+
+For each bar/phrase layer, enumerate candidate states and optimise the whole
+sequence using continuity, rhythm, register, technique, playability, style, and
+available harmonic costs. Compare against S6's locally ranked output.
+
+### Slice C — deterministic k-best alternatives
+
+Return several ranked global paths with:
+
+- fixed tie-breaking;
+- complete total/local/transition explanations;
+- an explicit diversity rule so alternatives are not path clones;
+- stable provenance for S8 display and S9 feedback.
+
+### Slice D — specialised clients
+
+After the engine and first client are accepted, consider complementary-guitar,
+harmonic, and cadence planners. Reuse existing fretboard DP rather than rewrite
+it only for abstraction symmetry. Register planning requires new measured
+counterevidence before reopening.
+
 ## Acceptance criteria
 
 - Recombined chains beat S6 single-strategy output on a defined quality score.
@@ -60,15 +104,22 @@ connected / possible); DP/Viterbi is the *route* (which sequence is best).
   by a fixed documented rule).
 - Multi-bar output shows a global arc (e.g. no 4 identical-technique bars in a
   row), not a chain of locally-best fragments.
+- Every selected path exposes local and transition-cost explanations.
+- k-best alternatives are deterministic and measurably distinct under the
+  documented diversity rule.
 
 ## Open questions
 
 - Minimum corpus size before the graph beats rule-based v0.
 - Exact cost-term weights (calibrated on the corpus; later tuned by S9).
 - State-size vs exactness trade-off before beam search is needed.
+- Which multi-bar client produces enough value to justify the first reusable
+  path contract.
 
 ## See also
 
+- [`../audit/2026-07-symbolic-harmony-and-evolution-research.md`](../audit/2026-07-symbolic-harmony-and-evolution-research.md)
+- [`S15-tonal-context-and-harmonic-control.md`](S15-tonal-context-and-harmonic-control.md)
 - [`../glossary.md`](../glossary.md) §9
 - [`../adr/0013-dp-viterbi-traversal.md`](../adr/0013-dp-viterbi-traversal.md)
 - [`../adr/0018-rich-note-model-fretboard-and-techniques.md`](../adr/0018-rich-note-model-fretboard-and-techniques.md)
diff --git a/docs/stages/S8-preview-app.md b/docs/stages/S8-preview-app.md
index 2a748a8e..767f431d 100644
--- a/docs/stages/S8-preview-app.md
+++ b/docs/stages/S8-preview-app.md
@@ -20,6 +20,12 @@ ADRs: —
 >
 > The `griff-preview` binary launches the TUI, or prints one headless frame with
 > `--snapshot=WxH`.
+>
+> Research update (2026-07): the future cockpit/playground should borrow the
+> **editable text + immediate visual/audio feedback** shape from symbolic music
+> editors, without adopting their notation format as Griff's domain model. S8
+> owns the surface for S7 path alternatives, S9 feedback/evolution lineage, and
+> S15 tonal/harmonic provenance; those stages retain their own semantics.
 
 ## UI design reference
 
@@ -50,6 +56,16 @@ front-ends and audio build on them:
 - [ ] `eframe`/`egui` native window — the canonical desktop target (piano-roll
       canvas, pan/zoom), reusing the same `PianoRollView`.
 - [ ] MIDI playback via `midir`, with a playhead overlay.
+- [ ] **Griff textual playground** — editable request/constraint text with live
+      parse diagnostics and immediate candidate refresh. Future S15 harmonic
+      fixture scripts may be edited here, but typed core structures remain the
+      source of truth.
+- [ ] **Candidate/provenance inspector** — candidate scores, novelty, register,
+      playability, tonal hypotheses, S7 path/transition explanations, and stable
+      ids sufficient for S9 feedback.
+- [ ] **Feedback/evolution surface** — like/dislike/favorite controls first;
+      parent/child lineage and generation history when the S9 Evolution Lab is
+      active. S8 displays and edits; S9 owns preference/evolution semantics.
 - [ ] Curation actions feeding the S5 corpus schema — **first slice landed
       2026-06-11**: approve/reject intents in the interaction core
       (`Viewport::decision`, ADR-0016 — repeat to undo), 'a'/'x' keys and a
@@ -95,24 +111,36 @@ the CLAP plugin, to debug transport/slicing/graph without DAW quirks.
 ## Inputs / Outputs
 
 - In: `.mid` / corpus chunks / candidates.
-- Out: piano-roll view, MIDI playback, boundary overlays, history,
-  approve/reject/split/merge/rename/tag actions feeding the corpus.
+- Out: piano-roll/tab view, MIDI playback, boundary overlays, candidate history,
+  score/provenance inspectors, and approve/reject/split/merge/rename/tag actions
+  feeding the corpus.
 
 ## Approach
 
 - New workspace member `preview/` using `eframe`/`egui` (immediate-mode,
   native; not Tauri — IPC/HTML overhead for an offline MIDI tool).
 - Playback via `midir`. Headless fallback: a `ratatui` TUI.
+- Text is an editable request/fixture surface, not a replacement for the typed
+  canonical model.
 
 ## Acceptance criteria
 
 - Loads a `.mid`, shows a piano-roll, plays it back.
 - Curation actions persist into the S5 corpus schema.
+- Candidate and provenance views use stable ids and headless-testable view
+  models.
+- S7/S9/S15 data is displayed without reimplementing their inference or policy
+  in the UI.
 
 ## Open questions
 
 - Playback engine details on each OS.
+- Minimal textual request syntax before the S15 fixture DSL exists.
 
 ## See also
 
+- [`../audit/2026-07-symbolic-harmony-and-evolution-research.md`](../audit/2026-07-symbolic-harmony-and-evolution-research.md)
+- [`S7-graph-layer.md`](S7-graph-layer.md)
+- [`S9-feedback-layer.md`](S9-feedback-layer.md)
+- [`S15-tonal-context-and-harmonic-control.md`](S15-tonal-context-and-harmonic-control.md)
 - [`../glossary.md`](../glossary.md) §11
diff --git a/docs/stages/S9-feedback-layer.md b/docs/stages/S9-feedback-layer.md
index 68a2e776..0317b5c4 100644
--- a/docs/stages/S9-feedback-layer.md
+++ b/docs/stages/S9-feedback-layer.md
@@ -7,14 +7,36 @@ ADRs: —
 ## Goal
 
 Let like/dislike/favorite steer ranking and sampling of candidates — not train
-a model in a vacuum.
+a model in a vacuum. Once basic preference learning is accepted, experiment with
+structural parent selection, crossover, and mutation while retaining full
+lineage and existing validators.
 
 ## Inputs / Outputs
 
 - In: candidates + their feature vectors, user ratings.
-- Out: a `PreferenceProfile`; reranked/ resampled candidate sets.
+- Out: a `PreferenceProfile`; reranked/resampled candidate sets.
+- Later experiment: parent selections + structural evolution operators → a new
+  candidate population carrying explicit lineage.
 
-## Approach
+## Planned phases
+
+### Phase 0 — feedback capture
+
+Persist inspectable events with stable candidate/session identity:
+
+```rust
+pub struct HumanFeedback {
+    pub candidate_id: CandidateId,
+    pub verdict: FeedbackVerdict,
+    pub generation: u32,
+    pub session_id: SessionId,
+}
+```
+
+The final names and storage format remain a design decision. `Skip`/no-op must be
+representable so absence of a like is not silently treated as a dislike.
+
+### Phase 1 — preference reranking
 
 - Like/dislike/favorite → update feature weights.
 - Baseline EMA update: `w_i ← (1-α)·w_i + α·sign(approve)·feature_i_norm`,
@@ -22,16 +44,75 @@ a model in a vacuum.
 - Explainable rerank by similarity / features / tags.
 - No gradient descent / RL before S10.
 
+### Phase 2 — Evolution Lab (experiment)
+
+Adapt only the human-in-the-loop idea from `perfect-shuffle-music`; do not port
+its note-array genome or crossover implementation.
+
+```text
+candidate set
+→ user selects parents
+→ bar/motif/parameter crossover
+→ mutation
+→ meter/register/playability/novelty validators
+→ existing rerank as a safety/quality guard
+→ next generation
+```
+
+Candidate operator vocabulary:
+
+```rust
+pub enum EvolutionOperator {
+    AlternateBars,
+    AlternateMotifs,
+    RhythmFromAContourFromB,
+    PrefixSuffixCrossover,
+    ParameterBlend,
+}
+```
+
+Lineage records generation, parents, operator, mutations, and session. Musical
+units are bars, motifs, rhythm grids, contours, gesture plans, register windows,
+endings, and strategy parameters — never a blind alternating array of MIDI note
+indices.
+
+### Phase 3 — diversity and collapse controls
+
+Measure and expose:
+
+- population and strategy diversity;
+- rhythm/contour diversity;
+- repeated-parent dominance;
+- operator survival rates;
+- population collapse across generations.
+
+Evolution Lab remains an S9 experiment. A new stage is justified only if it
+becomes a standalone persistent workflow with branching histories, undo/fork,
+operator analytics, and its own acceptance contract.
+
 ## Acceptance criteria
 
 - Ratings measurably shift subsequent ranking toward liked features
   (deterministic test with synthetic ratings).
 - The profile is inspectable and resettable.
+- Feedback events retain stable candidate/session provenance.
+- Evolution operators preserve structural validity and pass the normal generator
+  validators.
+- Fixed inputs, feedback sequence, and seed reproduce the same population and
+  lineage.
+- Diversity/collapse metrics are reported before Evolution Lab is considered for
+  product use.
 
 ## Open questions
 
 - Per-tag Beta-prior alternative vs single EMA.
+- Minimal parent-selection UX and population size.
+- Whether lineage belongs in the corpus schema or an experiment/session store.
 
 ## See also
 
+- [`../audit/2026-07-symbolic-harmony-and-evolution-research.md`](../audit/2026-07-symbolic-harmony-and-evolution-research.md)
+- [`S7-graph-layer.md`](S7-graph-layer.md) — deterministic k-best global alternatives
+- [`S8-preview-app.md`](S8-preview-app.md) — feedback and lineage UI
+- [`S15-tonal-context-and-harmonic-control.md`](S15-tonal-context-and-harmonic-control.md)
 - [`../glossary.md`](../glossary.md) §10