Research programme · Porsche 993 Turbo · M64
Experimental model · cylinder-head preparation
An engineering copilot.
One part at a time.
Teach a small local model to write geometry and scene code. Check what it produces. Build a documented digital twin, starting with the cylinder head.
Evidence snapshot ·
01 / The model
Better code needs a measurable test
The completed local pilot uses Qwen2.5-Coder-1.5B-Instruct, quantized to 4 bits, with MLX LoRA on an Apple M1 Max. PicoGK C# runs in the native geometry kernel; OpenUSD Python is checked with Pixar's tools.
Latest training: scene composition and variant selection
Two local trials add 1,000 training steps. Coding-007 learns from 288 new three-object USD examples and reaches 27/30 USD and 39/40 PicoGK validation passes. It is rejected because two earlier USD cases select the wrong variant. Coding-008 adds 96 matched training examples: the scene stays the same, while the requested small or large variant changes.
| Domain | Previous coding-003 | Coding-008 | Previous passes lost |
|---|---|---|---|
| OpenUSD | 22/30 | 30/30 | 0 |
| PicoGK | 9/40 | 31/40 | 0 |
Coding-008 passes the fixed validation rule and is selected as an experimental coding candidate. Every previous validation pass is retained. Final tests were opened only after selection; default inference remains unchanged.
| Partition | Previous adapter | Candidate | Previous passes lost |
|---|---|---|---|
| OpenUSD · regression cases | 28/36 | 36/36 | 0 |
| OpenUSD · 32 new instances | 11/32 | 31/32 | 0 |
| PicoGK · regression cases | 10/32 | 24/32 | 0 |
| PicoGK · 24 new graphs | 1/24 | 18/24 | 0 |
On the final tests, coding-008 satisfies 67/68 USD requests and 42/56 PicoGK requests. All 68 USD answers pass the available native checks. Of 56 C# answers, 54 pass the code boundary, compile and generate accepted native geometry; two are rejected before compilation. No previously complete task is lost, but compilation falls from 56 to 54. These remain experimental coding results.
These are adaptive trials using shared synthetic API recipes, not independent engineering benchmarks. All new reference programs passed the bounded native checks before training. The final model tests remain separate from training and validation. Download both composition trials and their receipts · JSON.
Earlier training: recovering OpenUSD retention
Two further training trials test retention. Coding-005 completes 600 steps but still loses seven previous OpenUSD passes and is rejected. Coding-006 starts from the retained checkpoint and trains only four LoRA layers for 800 steps at a lower learning rate. Both use more OpenUSD training replay. Validation and test prompts are unchanged; they never enter training.
| Domain | Retained coding-003 | Focused coding-004 | New coding-006 |
|---|---|---|---|
| OpenUSD | 22/30 | 17/30 | 19/30 |
| PicoGK | 9/40 | 40/40 | 7/40 |
The candidate failed the fixed retention rule and was rejected for the combined assistant. Coding-003 was retained, and the 24 fresh PicoGK model tests were left unopened at this stage.
This was a validation-driven continuation of the earlier experiment. Only the 24 new PicoGK cases were untouched by model inference before selection; all older tests were regression checks. The synthetic tasks share grammar with training. Download the earlier retention results · JSON.
| Test partition | Previous adapter | Selected adapter | Previous passes lost |
|---|---|---|---|
| OpenUSD · original regression cases | 3/12 | 12/12 | 0 |
| OpenUSD · fresh instances in this run | 1/24 | 16/24 | 0 |
| PicoGK · original regression cases | 8/16 | 9/16 | 0 |
| PicoGK · fresh graphs in this run | 1/16 | 1/16 | 0 |
Across these tests, OpenUSD improves from 4/36 to 28/36, and PicoGK from 9/32 to 10/32. All 32 final C# answers compile and 31 produce nonempty geometry. Only 10 satisfy the complete request. Compiling successfully is one check in a longer chain.
Read the earlier coding-003 protocol
The curriculum contains 528 unique training records, 54 validation cases and 68 test cases. Selection uses validation results; final tests are excluded from training. Original cases are development-informed regressions. That experiment selected step 600 of its retention continuation; the adapter remains experimental.
The report records per-case outcomes, dataset hashes, the selected adapter hash and native verification receipts. Model files stay local; the training code and reports are available in the engineering repository.
Earlier coding-004 trial: why retention training was needed
The 800-step continuation added 512 training graphs and expanded LoRA from four to sixteen layers. PicoGK validation rose from 9/40 to 40/40. OpenUSD fell from 22/30 to 17/30, losing seven previous passes while gaining two.
That candidate failed the fixed rule requiring improvement without losing previous passes in either domain. Coding-003 was retained, and the 24 fresh PicoGK tests were left unopened for the retention continuation. A disk-full interruption was recovered using saved answers and unchanged graders. Read the earlier continuation and recovery record · JSON.
02 / Part 01
The cylinder head comes first
The active record is 993-ENG-CYLINDER-HEAD-4V-F0-0001: preparation for the existing finned four-valve M64 head concept, with two intake and two exhaust valves per cylinder.
What exists
A catalogue identity, an interface contract, a creation brief and prior native body, valve, seat and guide studies. The existing 935 scan is a research proxy; it does not establish the dimensions of an exact 993 Turbo head.
This milestone prepares continued work on that geometry. It does not introduce a new finished head or a printable replacement.
What still needs evidence
Access to an exact donor, measured datums, stud and sealing interfaces, hidden oil galleries, valve-train packaging and thermal loads. Twin ignition remains an unresolved packaging requirement.
Alloy, process, fitment and physical validation remain open. Manufacture and engine operation are not authorised.
Catalogue research starts with the PorscheFanatics engine specification and Swindon head-kit record, then checks the manufacturer's primary information. Swindon's kit is a comparison reference. The site's proposed 4.0-litre build is separate from this project's 3.6-litre historical baseline.
Parts queued after the cylinder head
- Intake valve
- Exhaust valve
- Upper valve cover
- Three-runner intake
- Exhaust manifold
- Intercooler end tank
- Fan housing
- Cooling impeller
This is the preparation sequence. Each part needs its own interface review before the sequence advances.
03 / The architecture
Local work first. Evidence at every stage.
Active
Define
Identify the catalogue part, sources, measured interfaces and unresolved requirements.
Research
Create
Review PicoGK code and native CAD against the part contract. Retain each candidate and its checks.
Pending
Simulate
Check mesh quality, then run flow, heat-transfer and structural studies with documented conditions.
Pending
Qualify
Select material and process; inspect coupons and prototypes and correlate the physical tests.
Deferred
Review
Prepare OpenUSD locally. Rent an Omniverse GPU only for the final approved visual review.
Mac · training and USD
The Mac runs the current Qwen pilot and native checks. Pixar's CPU tools prepare and validate USD before a cloud GPU is needed.
Kali1 + Kali2 · connected workers
SSH access to both machines is verified. Kali2 passes the main suite (3,270 tests, including 156 optional skips) and all 15 Docker voxel-audit tests. Full repository validation stops at an existing catalogue-count mismatch; the remaining checks pass separately. Kali1 remains available for user-space work; its current account still needs Docker access. These software checks do not qualify a physical part.
Vast · final Omniverse review
GPU rental is deferred until the local assets and checks are ready. Visual inspection in Omniverse complements the engineering evidence.
Traceable architecture
The repository uses a tailored TOGAF, ArchiMate and UML approach with Mermaid diagrams: requirements connect to catalogue records, code, native checks and review decisions.
Current milestone: measured local coding pilot and cylinder-head preparation. OpenFOAM training, calibrated simulation, material qualification and physical print tests remain future work.
Engineering source and documentation
The result records above are public. The engineering repository contains the cylinder-head creation brief, architecture and Mermaid diagrams, and training code and reports.