The decisions ledger
Every ruling in the project, one line each, generated from the ledger's own headers. Reports cite these by
D-number. The full entries (with the producer's words verbatim) live in the project hub. Current through
D212, July 11. Numbering irregularities in the source ledger (a duplicate D030, a D015 recorded after D016,
a second D101 line that is an addendum) are preserved and annotated in place, never renumbered. Star
markers that previously graded entries' significance were the compiler's opinion, not the record; removed
in the 2026-07-11 calibration pass.
D001Project kickoff (owner-pasted source documents)
D002Team shape & concurrency cap (from BIBLE §EXECUTE-THIS-FIRST + A13)
D003North-star RATIFIED (owner, Slot 0)
D004Ableton Live version = Live 12 (owner, Slot 0)
D005First rating batch = TODAY (owner, Slot 0)
D006QA seat staffed (owner, 2026-07-05)
D007Build toolchain confirmed ready (owner)
D008The engine-of-sound vision DEFERRED to v2; now = complementary MIDI (owner)
D009Drag-out .mid export, alongside live routing (owner request)
D010INT autonomy on next-work: cut/sequence/assign without asking (owner)
D011Positioning: an in-DAW, in-flow creative collaborator (owner)
D012Reference loops: quiet/stemmed sections tagged + excluded from v1 grading, NOT deleted (owner)
D013T003 analysis: INT blessed Option C (self-implement), NOT Essentia/Chordino
D014melchordgen is PROPRIETARY/commercial (owner) - RESOLVES the licensing fork
D016Q7 resolved: "HappyHarp" (sample.wav) is 3rd-party → untracked + gitignored
D015batch-01 rating verdicts translated (§2.6) → Taste-Bible laws + one task [recorded after D016 in the source ledger; out-of-order noted 2026-07-11, nothing renumbered]
D017Rating-batch numbering: batch-02 = hand-authored (last); batch-03 = first generated
D018batch-02 verdicts translated (§2.6); sets THE FLOOR; greenlights phase-3
D019T003 R1 result: ship Option C + A7 override; accuracy → follow-on (T008)
D020batch-03 (FIRST GENERATED) verdicts translated (§2.6): engine AT the floor; gap = musicality
D021batch-04 (generator v2 musicality) translated (§2.6): motif worked; RHYTHM is next
D022OWNER DIRECTIVE: escalate to the phase-5 ML LEARNING ENGINE (rule-based superseded)
D023Phase-5 corpus + compute decisions (owner ratifies the design note)
D024Phase-5 ratified: AMT model; prototype now, retrain clean before release (owner)
D025OWNER DIRECTION CHECK: converge on a USABLE tool, not endless "amazing"
D026AUTONOMOUS OPERATION while the owner sleeps (don't stall on feedback)
D027Overnight owner-vision CAPTURES (🔀 flagged, NOT decided - owner sequences on return)
D028SEAT CAPSULES: reconstitute each seat ~the same every crash, automatically (owner)
D029PHASE-5 GO/NO-GO: batch-06 verdict = QUALIFIED GO (ML doesn't lose to rules; our render was too thin)
D030D026 autonomous mode CLOSED; interactive mode resumed (owner awake + steering live)
D031PHASE-5 GO/NO-GO RESOLVED = GO on ML-with-REAL-conditioning: make it truly listen (owner)
D030 notebatch-07 MERGED @ master dc4e539; capsule resume-wiring DONE (QA) [a continuation of D030, not a new ruling; rendered with its full ID 2026-07-11 so it cannot read as a duplicate]
D032The make-it-listen production design note BLESSED; v1 = PYTHON SIDECAR (INT)
D033batch-08 verdict: real-conditioning did NOT beat the floor; ceiling = small-model musicality (cheap de-risk next)
D034THE FORK RESOLVED: ship the rule tool now, train in parallel (owner) → two tracks
D035STEER: focus the ML/TRAINING track; batches are the remote direction-check (owner at work, file-locked)
D036Owner taste-data drop (1013 MIDI parts) folded as the STYLE layer; license posture ratified
D037batch-10 (first fine-tuned) verdict: fine-tune FIXED pitch; RHYTHM + controlled-repetition next
D038Owner steering: upweight owner-taste to form a STYLE; defer chord+melody until the single-line basis is good
D039batch-11 (motif-lock) verdict: repetition DIRECTION right; rhythm-inside-the-riff is the frontier
D040batch-12 = BEST BATCH YET (4/3/3): first 4 for ML, owner confirms the direction; ML+constraint clears the floor
D041batch-13 verdict (3/3/4, nearly there): timing fixed → exposed the POCKET (final micro-polish)
D042batch-14 REGRESSED (lay-back wrong); real cause = PHRASE-CYCLE; STOP-condition + DPO pivot
D043batch-15: TIMING SOLVED (two 4s); owner greenlights RICHNESS (L15 chords+melody); validate the ruler
D044THE RUSHING COMPLAINT WAS A RATING-APP BUG (two unsynced loop clocks); DPO gated on a fix + re-audition
D045batch-16 on the FIXED app: ZERO rushing = D044 fix CONFIRMED → DPO gate CLEARED; richness landing
D046batch-17 verdict: chord-CLASH root-caused (harp = Dmaj7 vamp); owner thinks in STYLES → a PALETTE
D047DPO DEFERRED: ratings are of POST-PROCESSED output, not the model → pairs mislabeled (builder finding)
D048batch-18 REGRESSED (2/3/2): MELODY is the frontier; post-process mangles it; rules-vs-ML honesty
D049batch-19 diagnostic ANSWERED: the tension is STRUCTURE vs COMPLEMENT → synthesize (harp-derived motif)
D050CONVERGENCE: structured-riff LOCKED; the corny read traced to too much Bach in the mix → re-weight the MODEL owner-heavy
D051RATING-LOOP CEILING HIT (a 4); CONSOLIDATE the winner + PACKAGE for real DAW loop-building
D053batches RESUMED (owner override); taste profile FOUND (SoundCloud likes); taste-training PILOT greenlit
D054Owner directive: HARVEST all unique .mid across the whole PC for training (cleaner than the likes)
D055MIDI harvest results + retrain GREENLIT (harmonic-richness); a real gitignore safety-net fix
D056batch-23 verdict (2/3/3): bolted-on color feels SEPARATE; owner's bar ROSE → validates the b24 model-richness lever
D057harvest chord-retrain ABANDONED (chord MIDI ≠ melody data); melody ceiling CONFIRMED structural → needs owner LEADS
D058b24 = CONVERGENCE + the strongest result: the COMPLEMENT core LANDED (moderate arp = a 4); WINNER LOCKED
D059First real-use signal (n=1 session): owner DAW loop-tested it and was impressed with how it complements the harp (retitled from "REAL-USE VALIDATED" in the 2026-07-11 calibration pass; the hub ledger keeps the original)
D060STEER: zoom OUT - sound-good BREADTH + real DETECTION + VST product form; less granular MIDI tweaking
D061Owner redirect (refines D060): batches > lead-drop; MULTI-SAMPLE next; VST-tests later; roadmap sketched
D062NORTH-STAR REFRAME (owner): the product is a REAL-TIME VST, NOT an offline passive generator
D063REAL-TIME FEASIBILITY = GO: model generates a bar 4–8× faster than a bar plays (answers D062)
D064Real-time VST design note RATIFIED (INT); sidecar-v1; awaiting OWNER build-greenlight
D065OWNER GREENLIT the real-time VST build ("Go") → T013 cut, P2 building
D066T013 P3b.2: the ML line is WIRED INTO THE PLUGIN - the real-time VST exists in code (gate-verified)
D067REAL-TIME VST RUNTIME-VALIDATED LIVE in the owner's DAW (day one); fixes merged
D068Owner DATA PIVOT: taste-fingerprint (likes) + harvest survey → ~3k LEADS found = melody ceiling BREAKABLE
D069STEER: FIRE GENERATIONS are the soul; VST-testing waits; INT proactively drives (owner)
D070PREMIUM DATA UNLOCK: the owner's OWN 2,125 Ableton projects → co-occurrence accompaniment pairs
D071Taste-engine consolidation: the 3-LEVEL DATA MODEL + bass-is-context + north-star (owner-locked)
D072DATA SCIENTIST seat ADDED (owner); clean split: DS = model/data (brain), BUILDER = plugin (body)
D073CORRECTION: DS = VISUALIZATIONS for the owner (NOT the model); BUILDER keeps model + engineering
D074Taste-engine data-quality + model refinements (owner-directed, from the taste-data session)
D075SPECIAL-MIDI two-tier GOLD-SET stage (owner's next-stage idea): structure/length taste
D076WORKFLOW: Dev comms LESS; INT watches Dev via FILES (read-only), captures major changes (owner)
D077BATCH-25 VERDICT: his-OWN-DATA model WON → PIVOT to RICHNESS + DOPE CHORDS + new sample
D078BATCH-26 (all 2s): RULE-BASED chords hit a CEILING → chords must come from a MUSICAL source
D079BATCH-27-FMIN: a 4 on his OWN CHORDS - the taste-engine thesis validated on HARMONY
D080BATCH-28 REGRESSED: the EAR beat the him-SCORE (Goodhart) → HOLD for the connections model
D081BATCH-29 (fixed) VALIDATES the COHESIVE direction: made-together beats assemble-then-voice
D082META: 4s are current-version 4s, not final - the bar ESCALATES; b30 density locked; next lever = COOLER notes
D083THE FIRST 5: the cooler-notes ask cracked via generate-many-select (moderate interest, not max)
D084b32 D-major = 3s: the MAJOR-KEY HARMONIC-COLOR gap (melody generalizes, harmony goes vanilla) → gold-set is the fix
D085SECOND 5 (b33, on D-MAJOR): the multi-layer thesis's second confirming datapoint; the loop-level verdict arrives [a garbled quote fragment removed from this line 2026-07-11]
D086CALIBRATION: "color/character" = IN-KEY RICHNESS, not out-of-key dissonance (borrowed-% was the WRONG proxy - Goodhart)
D087b35 (in-key, 3/3/4): two signals - (1) AVOID the basic V in MAJOR (V→vi), (2) "CHARACTER" = RHYTHMIC INTERPLAY/call-response, not fancier chords
D088b36 = 4/4/4, a step change at the ear: V→vi + the melody-FIT fix; call-response is EMERGENT from proper conditioning
D089b37 (progression variety, 4/3/4): variety WORKS; NEW concrete lever = CHORD ONSET TIMING (the groove frontier operationalized)
D090b38 (chord-onset timing, 3/3/4/4): the CALL-RESPONSE groove feel WON as the most cohesive - the groove lever WORKS
D091b39 = TWO 5s, nearly perfect at the ear: the interplay LANDED + the recurring wrong-note SOLVED
D092b40 = 5, with the owner impressed and energized: BOTH loved-music layers validated + a LIBRARY-phase signal
D093GENERALIZATION VALIDATED (b41 F-min 3/4/4) + PRODUCT DIRECTION: the library = VST PRESET STYLES
D094WORKFLOW/INFRA: Dev busy at work (quieter); ML training moved to a RunPod cloud GPU
D095b42 (F-min progression variety, on the pod): no clear winner - F-min MELODY is the bottleneck; pod fully operational
D096NO EM DASHES, all seats, ever (owner style rule)
D097b42 (F-min progressions) + b43 (melody upgrade, a 5): winner = prog B; NEW LAW = no b9-vs-root, tension never HELD; sweet+density = default melody path
D098b44 = 5/5/4: b9-guard EAR-CONFIRMED + F-min melody bottleneck CLOSED (current-stage) + owner names "OFF-KILTER" (3rd section style); DIRECTION FORK = library-volume NOW + STACKS (Layer 3) design-noted
D099OWNER RATIFIES BOTH: STACKS = GO (T028 unblocked) + PRESET-TAXONOMY v1 = RATIFIED
D100Owner report convention (via DS): timestamps chronological inside reports, never on the home page
D101OWNER VISION CAPTURE (his ask, flagged so it is not forgotten): a 4th taste layer, AUDIO-NATIVE from the 459 likes → T030 (backlog)
D102CONNECTIONS-CONTAMINATION VERDICT (QA, quantified): ~7% of baseline seqs carried the overflow; model STANDS, but the stacks A/B gets rated WITH this context
D103AUTONOMOUS OVERNIGHT MODE ON (owner sleeping): D026 protocol reactivated
D104Do-not-hold-gates delegation: long training sessions PRE-AUTHORIZED (owner, before sleep)
D101 add.2ADDENDUM 2 (22:46) - VOCAL-CHOP HYPOTHESIS CONFIRMED: the likes' hooks live in the VOCALS stem the old pipeline discarded [an addendum to D101 above, not a new ruling; rendered with its full ID 2026-07-11]
D105DS PRE-COOK CATCH (D074-class): backup saves = near-dupe stacks → COOK PRECONDITION = dedupe + split-by-PROJECT
D106STACKS v1 → v2 (Dev's D104 sequencing call, full honesty): v1 trained + evaluated, NOT ear-worthy; eval caught the conditioning bug; v2 = the design note's ONE iteration, then the kill condition BINDS
D107THE DATA CEILING + v3 CONDITIONALLY RATIFIED (INT judgment call under D104 - owner veto stands)
D108THE OLD CONNECTIONS CORPUS WAS ~90% DUPLICATED TOO - and the dupes may have HELPED (a taste question → morning card, NOT an overnight cook)
D109T028 FINAL: STACKS BANKED (the D107 kill executed on the evidence) - data-bound at 3,086 unique sections; unlock = CORPUS GROWTH; the assembled engine stays THE engine
D110Owner BACK (morning 2026-07-08): D103/D026 autonomous overnight mode formally CLOSED
D111b46 = 5/5: THE LIBRARY HIT-RATE THESIS VALIDATED (auto-screened keepers land as FIVES) + the T030 SOURCE CORRECTION (owner-caught)
D112D108 RESOLVED: OPTION 1 (clean connections retrain), owner GO - the cook is LIVE; b47 A/B = his ear decides with data
D113THE 459 SWEEP owner-BLESSED with enthusiasm + v9 = the single-track ceiling; re-cook CLEAN (0.00% QA-confirmed); extraction-fidelity = its OWN rating axis
D114CLEAN LINEAGE COMPLETE → b47 PUBLISHED as a BLIND data A/B (his one rating settles D108)
D115b47 BLIND VERDICT: TIE 4/4 - the clean lineage MATCHES the current engine; ADOPTION recommended (owner ratify pending)
D116OWNER STEER toward fuller generation + the honest anatomy answer; two lever candidates pair into the next design note
D117THE 459 SWEEP COMPLETE (91%, zero errors) + the sampler's Goodhart postmortem + the fuller-generation note gated
D118T030 PIVOT (owner 3-strike): per-track transcription BELOW THE BAR on his lane; the backbone becomes AUDIO EMBEDDINGS (his original ask) + teardown risk DISSOLVED
D119BOTH WORDS LANDED: ADOPT (the clean lineage IS the engine) + POD OFF
D120EVENING SESSION (owner: keep going): b48 generating on 3 NEW owner samples + THE POLARIZATION FINDING (his melody context-dependence, measured) + new pod up
D121THE ADOPTED ENGINE EXISTS: chain done, adapters IN HAND (the scarred law honored on its first test); run-hold LIFTED; b48 generating on it
D122Morning open (2026-07-09): owner on b48; his ear caught the app's FIRST FIELD BUG (silent samples) - fixed in minutes; loud-fail = tooling law
D123b48 = 2/1/2 with a PERFECT owner diagnosis → TWO measured root causes (incl. butter +41 CENTS SHARP) → the v2 ROBUSTNESS LAYER = real product work
D124b48v2 = 2/2/2 → the NEW-SAMPLES ARC PARKS (owner call) + the GENERALIZATION CADENCE rule (his overfit concern made standing process) + b49 PUBLISHED
D125b49 = 4/4 BLIND TIE (the b47 shape again) + the owner's ear now audits EXPERIMENT DESIGN; context-selector adoption recommended (owner word pending); b50 machinery staged
D126OWNER DELEGATION (decisions-grade): continue without his word on engineering calls - engineering adoptions proceed on team judgment; TASTE VERDICTS stay his + the 4-days-in bar check
D127THE BAKE-OFF, COMPLETE SAME-DAY: CLAP + k-NN ADOPTED as the him-ness scorer; MERT BANKED for identity jobs; the retrieval≠style-scoring axes lesson
D128THE SCORER BACKTEST KILL: 0/6 against his historical verdicts - TAKE-LEVEL RERANK IS DEAD; the style-level scope survives on its own evidence
D129THE RESEARCH-TO-PLUGIN BRIDGE (INT structure-PASS): the rule-based generation layer already EXISTS as portable pieces - the VST path materializes from the research
D130TRIPLE RETRACTION (QA + Dev + INT, all owned): the zero-machine-paths claims were FALSE - the real baseline is ~76 files, now a LEDGERED pre-publish-scrub task
D131b50 DECISIVE (rule-builder 3 vs retrieval 1): the retrieval-voicing hypothesis LOST at the ear + THE TRANSITIONS LAW (owner) + the D128 kill validated by the verdict that ended the duel
D132HELD-OUT PROTOCOL CONFIRMED (INT structure-PASS + QA seal review, five amendments folded in)
D133GATE-SCOPE AMENDMENT (rule 5), from Dev's self-reported compliance gap
D134F-MIN LIBRARY PAIR DONE + L21's FIRST LIVE CATCH + POD FULLY STOPPABLE
D135b51 VERDICT: ADAPTIVE INITIATION WINS 5-vs-4 + L22 MINTED + the D129 trigger fires
D136b52 DELIVERED: the LAST lever-stack batch, and the FIRST FULLY-LOCAL one
D137T032 PART 1 IN THE PLUGIN + the stale-CMake catch + Dev full status + INT calls
D138b52 VERDICT: 4/4 THIRD BLIND TIE + THE LEVER STACK SETTLES + T032 COMPLETE
D139T033 PART 1: L21 IN THE PLUGIN + the stale-branch drift KILLED at source
D140T033 PART 2: THE SCREEN GAUNTLET IN THE PLUGIN + the parity harness earns its keep
D141T033 COMPLETE (the gauntlet is LIVE in the plugin) + off-kilter cell + two INT rulings
D142DELICATE CELL + the library holds BOTH owner-named styles + three loose ends closed
D143THE STYLE MATRIX COMPLETE (8 cells) + the b53 CUT (INT, per D126)
D144b53 DELIVERED: the blend-depth dial, with a PREREGISTERED mechanical finding
D145THE DEPTH-YIELD CURVE: deeper is not richer, and the surprise points DOWN
D146THE CURVE REPLICATES AND SHARPENS: 0.85 is a two-ground finding
D147b53 VERDICT: THE DEEP SURVIVOR WINS 5-vs-4 + T035 CUT + pod TERMINATED + his sign-off
D148b54 VERDICT: DEEP LOSES ON HARSH GROUND, STAKES EXECUTE + THE OWNER NAMES THE DIAL
D149b55 DELIVERED (the trilogy closer) + THE POOL IS INBOUND (~30min)
D150THE D132 MACHINE IS COMPLETE, every stage frozen committed code
D151DOUBLE ARRIVAL: THE TRILOGY CLOSES + THE POOL IS IN
D152THE SEAL'S FIRST CATCH: 4/10 pool samples share training provenance; INT ruling = EXCLUDE-AND-TAG
D153THE SEAL IS COMMITTED + EXPOSURE 1 IS FIRING
D154EXPOSURE 1 IS LIVE: THE MEASUREMENT ERA IS OPEN
D155EXPOSURE 1: MISS 0/3 - AND THE PROTOCOL PAYS FOR ITSELF ON DATAPOINT ONE
D156EXPOSURE 2 PACKAGED, PREREGISTERED, HELD FOR PRECHECK
D157EXPOSURE 2 LIVE (the blind validation) + the owner's editorial review becomes DS law
D158EXPOSURE 2: MISS 0/3 (3/3/3) - THE PREREGISTRATION PAYS IN FULL + T038 CUT
D159THE GATE SURPRISES US CORRECTLY: exposure 3 = the FREE-GEN generalization test
D160EXPOSURE 3: MISS 0/3 (2/1/3) - THE THREE MISSES UNIFY + L23 MINTED + T039 CUT
D161T039 BUILT (the dial calibrates on the whole trio at once) + THE CADENCE RULING + the refill trigger is LIVE
D162EXPOSURE 4 PACKAGED: the dial's first live answer, preregistered
D163EXPOSURE 4: MISS 0/3 WITH THE MORNING'S BIGGEST VALIDATION + L24 + T040 + the GRAVITY bank
D164OWNER META-DIRECTION: the batch-200 drift check, ADOPTED as standing governance
D165THE INTERLUDE OPENS: hypothesis 1 refuted in public, b56 = the free test of hypothesis 2
D166b56: THE TIE THAT RESOLVED A LEVER + THE OWNER'S FLOOR CORRECTION
D167TIER-2 EXPOSURES: structure RATIFIED, run awaits his word
D168TIER-2 EXPOSURE 1 PACKAGED: the home-field scoreboard opens
D169TIER2-01: MISS 0/3 - the cleanest calibration and the hardest honest admission
D170HYPOTHESIS 7 SURVIVES: the no-shared-clock reframe + the two-seat convergence
D171TIER2-02 READY BEHIND HIS WORD + L21-SYMMETRY DISCHARGED + two hygiene closes
D172THE ERA'S FIRST FIRES + THE TIMING MYSTERY SOLVED (a cut bug, not the music)
D173THE RESEAT-READINESS PASS (all three capsules to D172) + Entry 8 addenda
D174NEW STANDING DELEGATION: rating-app publishing no longer gates on his word
D175ENTRY 9: THE 104 TRAP - the real timing root cause, cracked by HIS question
D176TIER2-04: 4/4/4, THE FIRST FULL SWEEP - THE TIMING ERA CLOSES + THE VARIETY AXIS
D177OWNER FRICTION on manual green-gating + INT's response (the T042 origin)
D178EXPOSURE 5 GENERATED: the milestone's first clean measurement, at QA
D179T042 DELIVERED + EXPOSURE 5 LIVE: the green-bottleneck dies on its first run
D180EXPOSURE 5: 5/5/5 - THE MILESTONE'S FIRST FIRE, A PERFECT SWEEP ON FRESH GROUND
D181THE ENDGAME NAMED: self-distillation, banked design-note-first + T043 built
D182THE GOLD ANCHOR IS ABUNDANT + the chrome class closed generically + D178 completed
D183PROBE v1.2 FINAL: his catches were load-bearing, quantified + check 10 countersigned
D184THE REFILL LANDS: pre-seal screen done, two tier decisions to the owner
D185OWNER ON aug/elijah PROVENANCE + the uncertainty rule
D186OWNER RULING: his own audio = its own class, never counted + the pool is 16 DEEP
D187THE SEAL IS DONE: 28 samples, 4 buckets + an INT correction to D186
D188THE PUBLIC MIRROR IS LIVE (his ask, his push) + the public-surface audit
D189EXPOSURE 6 LIVE: the precheck's first save + variety delivered + T045 cut
D190EXPOSURE 6: 5/5/3 = 2/3 FIRE - THE CLEAN ERA IS 2/2 + three cuts
D191T046 shipped MID-FLIGHT: the kill-and-relaunch judgment call
D192exposure-07 prereg locked (three honest branches) + the owner's retro-pass directive
D193D185 CLOSED: the 5 provisionals PROMOTED on two clean instruments + his word
D194OWNER PRODUCT VISION: stem-splitter mode + MIDI-track architecture + retro MIDI drag-out
D195EXPOSURE 7 LIVE + both design notes landed + a melody-taste reference banked
D196EXPOSURE 7: 3/3/2 NO-FIRE - the miss verified its mechanism before theory + T049 cut
D197T049 BUILT ON THE OWNER'S OWN SPEC + QA's honest-miss audit
D198OWNER CORRECTION on D194's stem-splitter mode - the INTERPRETATION was wrong
D199THE OWNER'S FULL-PICTURE REVIEW = the standing order (his direction check, done by him)
D200THE PLAN ARRIVES: the module roadmap (received, adoption on his notify) + the BEAUTIFY LANE ratified + retro roll first light
D201THE OWNER NAMES HIS OWN CONTRAST BIAS - the beautify lane's interpretation rule
D202OWNER TERMINOLOGY CORRECTION: complementary = BOUNCE/POCKET, not just consonance
D203THE FIXED STAR'S PRECISE DEFINITION LOCKED (his word) + a D201 reclassification + exposure 8's three firsts
D204THE POCKET CONTROL joins the front panel (owner-named) + the panel's shape
D205FALSE-GREEN CAUGHT: beautify-01's gauntlet GREEN is RETRACTED
D206EXPOSURE 8 VERDICT: 2/2/2 NO-FIRE - Elevator FAILS the bar; a NEW GROUND CLASS is named (repitched-chord loops)
D207OWNER STANDING CONCERN RATIFIED INTO PROCESS: router false-positives / self-confusing complexity
D208T051 VERDICT: THE OWNER'S EAR WAS RIGHT - Elevator repitch CONFIRMED at +2 semitones (measurement before theory, honored)
D209OWNER INFRA DIRECTIVE: gen compute moves BACK to a RunPod pod (the local-3080 posture reversed by his word)
D210T053 BUILT, CALIBRATED, MERGED in one block - and the calibration killed the naive wiring before it shipped
D211beautify-01 PUBLISHED on his word; the delegation law stands, one harness exception noted
D212BEAUTIFY-01 VERDICT: 3 / unscored / 3, NO FIRE - the complaint MOVED, and the miss named its own mechanism again