GAAD Technical Diary - 2026-07-25

Bundle 0_143 in, 0_144 out. Addon and generator untouched (v1_031 + v1_002).
The day ran on three fronts at once: the website, the AFAL bench, and - from
the moment the bench's HTML mockup started rendering real forms - the genome
programme that now sits under everything.

THE WEBSITE - THREE PUBLISHES (1_045, 1_046, 1_047)

• About had DRIFTED from build_about.py: nine hand-edits never carried back

(canonical, og/twitter, JSON-LD, two nav links, hero padding, font-display,
the whole team block). Regenerating would have silently deleted the lot.
All folded into the generator; output diffed line by line against the
published page before anything moved. 1_045.

• The Team is its own page (build_team.py v1_001); "Who it's for" left the

nav, the #audiences section stays on the homepage. Five navs updated. 1_046.

• The thesis photograph (Film 001 scene 1) joined About after the paragraph

that names the thesis, and its 1200x630 crop is now About's share card.
New images/ folder. 1_047.

• Generator map measured before touching anything: About/Team generate whole;

index/architecture are patched in place by gaad_site_stats.py (the html IS
the source); diary lifts its chrome from episodes; episodes' two generators
are ~300 lines stale - relics, ticket-worthy.

THE BENCH - v3_028 ACCEPTED, v3_029 ROLLED (F051, A0024)

• Miles accepted the Shell window AT DEFAULTS ("shell is fine") - the first

per-species emergence window to stand without tuning.

• v3_029 rolled the mechanism to Phyllotaxis: blooms spilled plots threefold,

traced to r = Radius x sqrt(i) drawing from the shared Wild envelope; the
phyllo window is a verified 2.89x shrink. Both species now ignore
Calm/Wild/Extreme - the window IS the envelope.

• Honest status: Miles saw "some issues" in v3_029, not yet itemised (A0027).

THE CLEAN UX (D208)

• Four-section IA: Species -> Generate -> Presets -> Parameters -> RNG.
• Miles's taxonomy: six families, eighteen species; planned species visible

but locked with the reason stated.

• Kill list ruled item by item; breeding OUT ("I have not got this to work

so for now its out"). Mode reborn as RNG Strength; Module fixed to Cube.

THE GLASSHOUSE (P070, P071, F052, D203)

• The clean-panel HTML mockup became a live instrument: the real growth laws

in rotating 3D. Verified: no external references, D197 correct, Canvas 2D.

• Named the GAAD Glasshouse, GLAS - a SURFACE, not a place. Hosts: the web,

Blender, one day the product, one day Revit. AFAL is the testing ground;
if it works it works for all languages (Q045).

• F053 parked years ahead: Export Champion, the selected winner leaving GAAD

for Revit carrying its ancestry.

THE GENOME FORMAT (D202, D204-D207; spec v0.001 -> v0.009)

• D202: one genome, one format, host-free and language-free; it buys a TEST.

Claude writes the format, the Code Technician builds it. Format -> export
-> import, in that order.

• D204 (Miles's ruling, Q046): lineage REQUIRED. Biology's genome carries

inherited information, not the parent; GAAD's carries parent, lineage id
and generation. Ancestry derives by walking the chain.

• D205 (after bench review): PARAMETERS are authoritative; the seed is

provenance and the breeding input. Two tests: conformance proves the LAWS
agree, fidelity proves the HOST renders what the genome says.

• D206 (after bench review): compare UNROUNDED float64 within sized

tolerances (1e-9 m; one f64 ULP at 30 m is 3.6e-15, the millimetre is a
million times up); the module is a language-defined GENE; only ACTIVE
genes are written.

• D207 (after bench measurement): float32 fails the laws at scale. Real

Blender: Phyllo 321.6 mm at 5000; the helix species 1.33 mm - the proxy's
12-micron figure was wrong, the graph is RADIAN-NATIVE and the storage
error rides up with index. The law moves to float64 Python; nodes place
modules. No second approximate path per host. Both fidelity runs complete;
the vertical axis is benign (z has no wraparound - its magnitude is its
content).

• The mapping (bench first task) answered F1-F7 in v0.009 and caught THE

FIRST LIVE DRIFT: the web scales Hyperboloid modules, the engine holds
them at 1.0. Two copies, one author, one week. T032 carries it; the
engine's rule stands.

RECORDS FILED TODAY

D202-D208, F052, F053, P070, P071, T032, Q043-Q048; F051 and A0019/A0024
updated; A0026 (exporter, unblocked) and A0027 (v3_029 issues) opened;
releases 1_045-1_047 and 0_144. The 0_143 release record's date corrected
(26 -> 25 Jul) with 43 other places.

LESSONS BANKED

• Fix at the source: the About drift and the spec-HTML date lapse were the

same disease at two scales, one of them mine.

• A proxy is a hypothesis, not a result; identity-compare nothing - measure.
• Drift needs neither two teams nor two years. One author and one week.

OPEN ACTIONS

AFAL

• A0019 high / Claude - the language file is behind the engine (now gates the exporter, F1)
• A0024 high / Claude Code - per-species ranges: PART DONE (Shell validated, Phyllo rolled; three species remain)
• A0026 high / Code Technician - the genome exporter, then import, then the conformance check (the check is the deliverable)
• A0020 medium / Claude - T1 research for the bench species
• A0021 medium / Claude - surface topology: the tier that makes a panel system possible
• A0027 medium / Miles + Code Technician - itemise the v3_029 issues Miles saw in Blender
• A0025 low / Claude Code - geometric NATURE row on the Singularity panel

GALF

• A0011 high / Miles - rule the fourth element: does a relationship carry a rationale
• A0012 high / Miles - TWAL s6: map the three retired types
• A0013 medium / Miles - ARAL s6 declares types but states no triples

MHAL

• A0001 high / Miles - Midtown East recount under the D181 tax block rule
• A0015 medium / Miles - the prototype carries MHAL verbatim at v0.003

Platform

• A0022 high / Claude - AFAL has no assert contract and no Validate steps
• A0008 low / Miles - Prototype/ carries stray ARAL versions

Website

• A0023 medium / Claude - Episode 6: AFAL and the singularity (now also debuts the four-seat team)
• A0006 low / Miles - latest episode first: name the surface

STILL OPEN AT CLOSE

• First task next session: A0027 (itemise the v3_029 issues) before the range

rolls on; A0026 (the exporter) runs at the bench in parallel.

• Open rulings for Miles: does a genome pin the language version strictly;

is the AFAL Studio panel the Blender instance of GLAS; Cone's home in the
taxonomy; the Growth family-vs-gene name.
zip (D159), never the live bundle.