# Composites / MatSci — Sketching Checklist + Evidence-Ontology Enhancements

*Companion to `BIOTWIN_ACCELERATOR_CHECKLIST.md`. Same FORZA governance spine;
composites-specific rows. Refines the generic sketching checklist and closes the
five gaps a general evidence ontology leaves open when chemistry **and** geometry
are both first-class.*

**The one-line thesis:** a composite is a *statistical* material that fails at
its weakest link. So the deliverable is never "the strength" — it is a
**B-basis allowable with declared confidence**, computed from the lower tail.
Every gate below serves that truth.

---

## Part A — Sketching checklist (refined)

### 1. Purpose — what are you sketching?
Fiber-orientation distribution · layup schedule · reaction network / energy
landscape · property tensor across conditions · defect population · NN weights.

### 2. Choice of sketch — **and declare its geometry + order-statistic**
The sketch family is not enough; the manifold and the statistic it preserves are
the load-bearing declarations.

| What | Sketch | **Geometry** | **Order-statistic** |
|---|---|---|---|
| Strength / life | **Weibull / EVT (REQ, DDSketch)** | log or linear | **minimum / lower tail** |
| Fiber orientation (ODF) | spherical-harmonic / moment (Advani–Tucker tensor) | **SO(3) / sphere** | moments |
| Layup schedule | **HRR / VSA — position-tagged, order-preserving** | sequence (non-commutative) | structure |
| Microstructure field | two-point S2 / variogram (Torquato; PyMKS) | spatial | correlation length |
| Connectivity / percolation | union-find + giant component / HyperANF | graph | threshold |
| Composition | CLR/ILR then quantile | **simplex (Aitchison)** | moments/quantiles |
| Distributions (general) | t-digest / KLL | Euclidean | median-band |

Red rules: a **permutation-invariant** layup sketch is a launder (stacking
sequence drives the CLT D-matrix). A **p50** sketch is the wrong tool for
strength (weakest-link → tail). Averaging **compositions** or **orientations** in
Euclidean space leaves the manifold.

### 3. Parameters & error budget — **tail-aware**
Sketch size / buckets · seed (reproducibility) · acceptable error **stated per
order-statistic**: 4% at the median may be fine; 4% at the p1 tail where the
coupon fails is disqualifying. The **>4% built-in red line** stands, but it is a
tail red line, not a median one.

### 4. Provenance & metadata (FORZA)
family · mergeability · error model · **geometry** · **order-statistic** ·
grain-at-sketch-time · error bound + confidence · algorithm + version · source
fingerprint · freshness · invalid-use envelope.

### 5. Validation — **contract, not vibe**
Reconstruction vs held-out ground truth at small **and** large N (the low-data
weakness you already caught). Two property tests, always:
- **Mergeability:** `sketch(A∪B) ≈ merge(sketch(A), sketch(B))` within declared error.
- **Manifold round-trip:** transform→sketch→inverse lands back on the manifold
  (compositions sum to 1, orientations stay unit, strength stays ≥ 0).
Where the claim feeds design: **conformal-calibrate** the interval (measured
coverage), don't just label it.

### 6. Authority boundary
"Approximate similarity" · "cardinality" · "not for precise quantiles under
N=10k." A sketch **nominates, never certifies**: `sketch_derived` evidence is
capped at `candidate` posture until a count-last exact pass confirms it.

### 7. Integration point
Register `sketch_derived` as a first-class evidence type in FORZA · link to the
`composite_architecture` / `material_system` it summarizes · store the sketch
bytes in HCD versioning (content-addressed).

---

## Part B — Evidence-ontology enhancements (the five "don't-gots")

Your seven classes (Material System · Composite Architecture · Processing History
· Test Specimen · Evidence Type · Authority Boundary · Grain Scale) are sound.
Add these:

### B1. Interphase as its own class (pillar, not bullet)
Where the borosilicate↔nylon↔ceramic gradient buffer lives. Own grain (nm-scale
gradient), own tests (single-fiber fragmentation, microbond, pull-out), own
failure competition (debond vs matrix crack vs fiber break — a multi-channel
committor; report the branching, not one saddle).

### B2. A/B-basis statistical gate (the real output)
Coupon prediction emits a **distribution**, and design-facing claims emit an
**A-basis (99/95) or B-basis (90/95) allowable** per CMH-17 methodology — a
lower-tolerance bound, i.e. the Weibull/EVT tail from Part A. No design allowable
without this gate. (This is the composites twin of BioTwin's conformal gate.)

### B3. Homogenization operators, named and ELA-classed
Make scale-bridging explicit and bounded:
`micromechanics (rule-of-mixtures / Halpin–Tsai / Mori–Tanaka)` →
`CLT / ABD (lamina → laminate)` →
`progressive damage (Hashin / Puck / LaRC)` → `RVE-FE homogenization`.
Each is a **lax functor**: declare its REV and its homogenization error; posture
degrades across the step, never silently upgrades.

### B4. Process → defect → tail chain (close the loop)
Processing is not a metadata list; it *induces the defect population that governs
the tail*. Cure kinetics (DiBenedetto Tg–conversion), consolidation, and thermal
history emit voids / wrinkles / residual stress → that population is the **input
to the weakest-link EVT** → which produces the B-basis allowable in B2. Model the
chain, not the boxes.

### B5. Environmental / aging axis
Composites degrade in ways metals don't: **hot/wet knockdown** (moisture
plasticizes the matrix, lowers wet Tg), thermal-cycling microcracking, moisture
diffusion (Fickian/dual-Fickian), oxidation/UV. Own evidence axis, own standards,
own knockdown factors — the Space-Shuttle-thread failure regime.

---

## Overlap map (Composites ↔ BioTwin — the paired exercise)

| Layer | Shared spine | Composites-specific | BioTwin-specific |
|---|---|---|---|
| Governance | evidence classes, refusal, provenance, FORZA | interphase class | strictest-tier default, SaMD/CDS |
| Validity | committor, spectral gap, κ-ledger | debond/crack/break branching | ensemble/allostery |
| Scale bridge | grain / REV / MEO | micromechanics → CLT → damage | **PK/PD compartment** |
| Causality | — | process→defect→tail (mechanistic) | **target-trial / counterfactual** |
| **Calibrated tail** | **the shared hard gate** | **Weibull + A/B-basis (CMH-17)** | **conformal under covariate shift (n=1)** |
| Sketch | order-statistic + geometry | HRR-ordered layup, ODF, EVT | log-transcript, torus-dihedral, EVT |
| Privacy | — | (IP/export sensitivity) | PHI custody + DP |

**Read-out:** the spine is mature and identical. The single muscle missing in
*both* accelerators is **honest tail statistics** — build it once (weakest-link /
lower-tolerance-bound / conformal), skin it twice (CMH-17 for composites,
conformal-under-shift for the patient).

---

**Status:** v0.1 — companion to BioTwin checklist.
**Next:** review B1 (interphase) and B4 (process→defect→tail) with the
Hexcel/DuPont thread SME; those two are where the borosilicate/nylon gradient
work becomes first-class instead of a footnote.