---
vault_clearance: EUCLID
halo:
  classification: FORM ΓÇö Functional Orthodox Reality Model
  confidence: HIGH
  front: "34_Project_DeadShape"
  custodian: "Claude Code seat (Resolve line)"
  created: 2026-04-24
  updated: 2026-04-24
  wing: CONDITIONAL
  containment: "Orthodox counterparts this project is expected to beat or cross-validate against, with explicit free-parameter counts."
---

# FORM ΓÇö The Dead Shape vs the Field

Per [FORM_PROTOCOL](../FORM_PROTOCOL.md), every project names the orthodox counterpart it competes against or cross-validates with. For Dead Shape, the orthodox counterparts dominate the last 20 years of protein/fold modelling, and every one of them is heavily parametrised.

## §1 The parameter-budget lens

The Dead Shape project adopts a strict discipline: **name the free-parameter count of every tool it touches, including its own.** The vault's methodological gold standard is [LENG](../05_Project_LENG/README.md), which derives 87+ Standard Model observables from **zero free parameters**. Any protein-topology claim should be read against that benchmark, not against the field's implicit standard of "enough parameters to fit the training set."

| Tool / Method | Free parameters (approx) | Trained on | Truth-quadrant |
|---------------|-------------------------:|------------|----------------|
| **LENG (S^5/Z_3 spectral action)** | **0** | geometry only | Rogue + (Righteous/Daemonic) both ΓÇö zero-param |
| **Dead Shape `topology.py` LRF/RCO** | 3 (cutoff=8.0Å, min_seq_sep=3, LR_thresh=12) | n/a — direct measurement on given structure | Rogue + Daemonic (geometric, not trained) |
| **Dead Shape `topology.py` circuit topology (P/S/X)** | 1 (contact cutoff inherited from LRF stage) | n/a | Rogue + Daemonic (true invariant, one upstream param) |
| **Dead Shape `topology_multimodel.py` ensemble reporting** | 0 additional (aggregates over provided ensemble) | n/a | Rogue + Daemonic when fed experimental ensemble |
| **AMBER ff19SB force field** | ~1500ΓÇô2500 (bonds, angles, dihedrals, LJ, partial charges) | QM of amino-acid dipeptides + crystal/liquid data | Orthodox + Righteous |
| **CHARMM36m** | ~1500ΓÇô2500 | similar, different training | Orthodox + Righteous |
| **Water models** (TIP3P / TIP4P-Ew / OPC / SPC/E) | 3ΓÇô9 per model | liquid-phase observables | Orthodox + Righteous |
| **Rosetta REF2015 scoring** | ~20 weights | decoy discrimination datasets | Orthodox + Righteous |
| **AutoDock Vina scoring** | ~6 weights | PDBbind + similar | Orthodox + Righteous |
| **Glide empirical** | ~10 weights | proprietary training sets | Orthodox + Righteous |
| **BLOSUM62 substitution matrix** | 400 (symmetric) | BLOCKS database of aligned protein families | Orthodox + Righteous |
| **PSIPRED secondary-structure** | ~10k neural-net weights | DSSP assignments on non-redundant PDB | Orthodox + Righteous |
| **AlphaFold 2** | **~93 million** learned parameters | PDB (~170k structures) | Orthodox + Righteous (the FORM Dao Monarch) |
| **AlphaFold 3** | **~500 million+** | PDB + protein-ligand + nucleic-acid complex data | Orthodox + Righteous (the 2024 FORM Dao Monarch) |
| **ESMFold** | **~15 billion** (language model) + million-scale folding head | UniRef + PDB | Orthodox + Righteous |
| **RoseTTAFold2** | ~hundred million | PDB | Orthodox + Righteous |
| **Generalized Born implicit solvent** | ~10ΓÇô20 (Born radii per atom type) | PB numerical solutions on test set | Orthodox + Righteous |
| **Poisson-Boltzmann** | dielectric ╬╡ (typically chosen 2, 4, 8, or 20) + atomic radii + probe radius | none; chosen by author | Orthodox + Righteous (but often under-documented) |
| **PROPKA pKa prediction** | ~15 per titratable residue type | experimental pKa benchmarks | Orthodox + Righteous |
| **Martini 3 coarse-grain** | ~200 (bead-type interaction matrix) | higher-resolution MD | Orthodox + Righteous |
| **Gō-model native contact sim** | 2 (native cutoff, interaction strength) | the native structure itself (circular) | Orthodox + Righteous |

**The total parameter budget of a typical 2025 protein-modelling pipeline:**

> Take an AlphaFold 3 prediction (~500M params), relax it with AMBER ff19SB + TIP4P-Ew (~2500 + 4 params), score binding with Vina (~6 weights), compute pKa with PROPKA (~15×20 sites = 300 params), and compare to a benchmark with the BLOSUM62 matrix (400 params). Total: ~500 million free parameters, nearly all trained from the PDB lineage.
>
> Each one was chosen to reproduce the training distribution. When the claim is "AlphaFold predicted this novel amyloid fibril," the honest reading is "a 500-million-parameter function interpolated on the PDB returned a value, and the PDB mostly doesn't contain amyloid fibrils."

## §2 What Dead Shape must beat (or pair with)

For each orthodox counterpart in §1, Dead Shape's stance is:

| Counterpart | Dead Shape's posture |
|-------------|---------------------|
| AlphaFold 2/3, ESMFold | **Acknowledge as FORM Dao Monarch.** Use for exploratory sequence→fold guesses; **never as the sole witness** for a deployment-state claim (HALO_TOPOLOGY_AUDIT §4d + §11). When both AlphaFold and an experimental ensemble exist, report both and flag disagreement. |
| MD force fields (AMBER, CHARMM, etc.) | Accept as valid for *relaxation* of experimental models, not as ground truth for novel conformations. Always report which force field + water model + parameters were used. |
| Rosetta / Vina / Glide scoring | Do not use for primary claims. These are fitted to binding or decoy sets; out-of-domain generalisation is known poor. |
| BLOSUM / PAM / substitution matrices | Do not use for topology claims. These are sequence-level, not structure-level. |
| Gō-models | Circular (fit to native state, then used to "predict" the native state). Do not use as independent witness. |
| Persistent homology / knot invariants | **Allies, not counterparts.** These drop toward zero parameters. Adopt them as core metrics (O-TOPO-9, O-TOPO-10). |
| Experimental ensembles (ssNMR, cryo-EM, SAXS) | **The daemonic witnesses.** Pull, ingest, compute topology directly, report distributions. The `osiris/experimental_pull/` directory is the proof-of-concept; expand systematically. |

## §3 What counts as "real" vs "modeled" in the upstream data

Every structure file Dead Shape consumes is itself a *model* fit to an *observation*. This matters because my metrics (even parameter-parsimonious ones) inherit the fitting regime of whatever they are run on.

| File format | Truly-measured observable underneath | Model layer applied to get atomic coordinates | Free params in that model layer |
|-------------|-------------------------------------|-----------------------------------------------|-------------------------------:|
| Crystal PDB | X-ray diffraction pattern (electron density map) | Rigid-body + flexible refinement minimising R-factor against density | ~4 per atom (x, y, z, B-factor) × thousands of atoms + scale factors + TLS groups + occupancy parameters. **Tens of thousands of fitted params per structure.** |
| Cryo-EM PDB | Electron-counting detector images (motion-corrected, CTF-estimated, class-averaged to a density map) | Real-space refinement against the map | Similar to crystal ΓÇö atomic coordinates + B-factors are fitted. Additionally, the map itself has CTF + motion parameters. |
| ssNMR PDB (e.g., 2RNM) | Chemical shifts + NOE distance restraints + dihedral restraints + paramagnetic relaxation enhancements | Simulated annealing with a force field, constrained to satisfy observations | Force-field parameters (~1500+) + restraint weights (~5ΓÇô20 weight parameters) + 20-model ensemble generation protocol |
| AlphaFold PDB | ΓÇö (no direct experimental observation of the target) | The entire 500M-parameter model | **~500 million** |
| MD trajectory (.dcd, .xtc) | Initial structure (PDB, itself a fit) | Newtonian integration of force-field energy | Force-field parameters (~2500) + all MD sampling parameters (timestep, thermostat, barostat, cutoffs) |

**What IS real** (i.e., the raw measurement you could in principle re-process with a different model):

- `.hkl` / `.mtz` crystallographic reflection data
- Raw cryo-EM movies (motion-uncorrected)
- NMR peak lists, coupling constants, NOE tables, chemical shift tables
- SAXS scattering curves I(q)
- CD spectra ╬╕(╬╗)
- Fluorescence / FRET intensities
- Mass spectra (m/z, intensity)
- Native-PAGE / size-exclusion chromatograms

**What IS NOT real** (i.e., already a model):

- The atomic coordinates in a PDB file (even a crystal structure ΓÇö it's fit to density)
- "Binding affinity" from a scoring function
- "Fold" or "structure" as a singular answer
- Any AlphaFold / ESMFold / RoseTTAFold output
- Any MD-relaxed structure
- Any "native state" inferred from a Gō-model

The Dead Shape posture: when we can, work one layer deeper (ingest the raw observable). When we can't, name the model layer and its parameter count explicitly.

## §4 Orthodox counterparts Dead Shape explicitly declines

| Counterpart | Why declined |
|-------------|--------------|
| **Empirical docking scoring for binding claims** | Known poor out-of-domain generalisation. If a binding-affinity number is needed, it must come from ITC, SPR, or other direct biophysical measurement ΓÇö not from a 6-weight scoring function fitted to PDBbind. |
| **Secondary-structure-only arguments** | DSSP thresholds are arbitrary. Helix/sheet/coil is a lossy projection of full topology; when possible, use circuit topology or persistent homology instead. |
| **Ramachandran plot alone as structure validation** | Only checks backbone dihedrals. Says nothing about tertiary topology. |
| **RMSD alone as structure similarity** | A single number. Insensitive to topologically distinct rearrangements (e.g., a ╬▓-solenoid vs a ╬▓-sandwich can have similar RMSD after alignment; they are topologically different). Use LRF + PSX + RMSD together. |

## §5 Meta-discipline: the parameter budget ledger

Every result Dead Shape publishes should carry a **parameter budget** line, like a nutritional label:

> Example: *"HET-s LRF = 0.875 ┬▒ 0.008 (n=20 models, 2RNM ssNMR ensemble). Free parameters in the measurement chain: 1500 (force field used in NMR restrained annealing) + 10 (restraint weights) + 3 (topology.py cutoffs). Upstream raw observation: 114 NOE distance restraints + 189 dihedral restraints + 72 hydrogen-bond restraints (PDB 2RNM header)."*

This is the audit trail LENG's zero-parameter discipline forces. Dead Shape is not zero-parameter (it cannot be ΓÇö it is a measurement tool on structures). But it *can* be honest about how many parameters sit between the raw observation and the number it prints.

## §6 TRUTH quadrant placement

| Tool | Orthodox ←→ Rogue | Daemonic ←→ Righteous |
|------|-------------------|----------------------|
| AlphaFold prediction | **Orthodox** (trained on PDB) | **Righteous** (uses sequence as biology-guided input) |
| Rosetta scoring | **Orthodox** (trained weights) | **Righteous** (uses physics terms) |
| AMBER MD relaxation | **Orthodox** (trained force field) | **Righteous** (biology-informed atom types) |
| `topology.py` LRF on a single PDB | **Rogue** (no training) | **Daemonic-leaning** (pure contact geometry, no biology input) |
| `topology.py` circuit topology | **Rogue** | **Daemonic** (a true invariant) |
| `topology_multimodel.py` on ssNMR ensemble | **Rogue** | **Daemonic** (the structure came from experiment, the metric came from geometry) |
| Persistent homology on C╬▒ (proposed, O-TOPO-9) | **Rogue** | **Daemonic** (no cutoff, no training, pure topology) |
| Knot invariants on backbone (proposed, O-TOPO-10) | **Rogue** | **Daemonic** (mathematical invariant) |

**Convergence rule (TRUTH_PROTOCOL):** when Dead Shape's Rogue+Daemonic result on an experimental ensemble agrees with a Rogue+Righteous literature measurement (e.g., the coupling tensor) and an Orthodox+Righteous prediction (e.g., AlphaFold) is the outlier, the convergence wins. HET-s was the first instance (LADDER 0.852 Γëê ssNMR 0.848ΓÇô0.875 ; AlphaFold 0.467 is outlier).

---

*This FORM document is append-only for the tool-inventory rows. When a new orthodox counterpart is published or a new Dead Shape metric is added, append a row; do not edit prior rows except to correct a factual error.*