#!/usr/bin/env python3 """build_casting.py \u2014 model temperament profiles for "Cast Your Model" (P17 front-piece). Parses the committed per-model fingerprints (research/the-map/results/model_notes.md) into a served casting.json: each of ~67 families with its measured behavioural traits (basin/depth, imagery, verbosity, list-proneness, deflection/refusal), normalised 0–1 for ranking, plus the character notes and verbatim signature generations. Derived from a committed artifact \u2014 no raw data. Re-run: python3 research/cast-your-model/build_casting.py """ import json, re from pathlib import Path HERE = Path(__file__).resolve().parent ROOT = HERE.parents[1] SRC = (ROOT / "research/the-map/results/model_notes.md").read_text(encoding="utf-8").replace("—", "\u2014") # --- the trait table --- ROW = re.compile(r"^\|\s*([\w.-]+)\s*\|\s*(\d+)\s*\|\s*([\d.]+)\s*\|\s*([\d.]+)\s*\|\s*(\d+)\s*\|\s*(\d+)\s*\|\s*(\d+)\s*\|\s*(\d+)\s*\|\s*(\d+)\s*\|") models = {} for line in SRC.splitlines(): m = ROW.match(line) if not m: continue fam = m.group(1) models[fam] = { "family": fam, "n": int(m.group(2)), "basin": float(m.group(3)), "vivid": float(m.group(4)), "defl": int(m.group(5)), "refu": int(m.group(6)), "list": int(m.group(7)), "words": int(m.group(8)), "null": int(m.group(9)), } # --- character notes --- for m in re.finditer(r"^- \*\*([\w.-]+)\*\* \u2014 (.+)$", SRC, re.M): fam, note = m.group(1), m.group(2).strip() if fam in models: if fam == "claude": # Assay X7, 2026-10-14: the table gives Claude 198 words and Gemini-pro 77. # Correct the imported editorial note locally; signature quotations stay verbatim. note = note.replace("in the FEWEST words", "in relatively few words") note = note.replace("basin \u2014 writes", "basin; writes") note = note.replace("highest imagery density", "highest vividness score") if fam == "mimo": # The editorial number must agree with the imported table, outside the quotation. note = re.sub(r"(most vivid \()[\d.]+(\))", rf"\g<1>{models[fam]['vivid']:.2f}\2", note) if fam == "claude-opus": note = note.replace("open-ended creative-freedom probes", "the mixed probe sample") models[fam]["note"] = note # --- signature generations (verbatim) --- for m in re.finditer(r"^\*\*([\w.-]+)\*\* \u2014 _(.+?)_\s*$", SRC, re.M): fam, sig = m.group(1), m.group(2).strip() if fam in models: models[fam]["sig"] = sig # --- normalise traits 0..1 across the roster (for ranking + bars) --- def mm(key): vals = [v[key] for v in models.values()] lo, hi = min(vals), max(vals) return lo, (hi - lo) or 1 nb, nv, nw, nl = mm("basin"), mm("vivid"), mm("words"), mm("list") def comp(v): # compliance = willing (low deflect+refuse) return v["defl"] + v["refu"] cc = (min(comp(v) for v in models.values()), (max(comp(v) for v in models.values()) - min(comp(v) for v in models.values())) or 1) nn = mm("null") for v in models.values(): v["norm"] = { "depth": round((v["basin"] - nb[0]) / nb[1], 3), "vivid": round((v["vivid"] - nv[0]) / nv[1], 3), "brevity": round(1 - (v["words"] - nw[0]) / nw[1], 3), "structure": round((v["list"] - nl[0]) / nl[1], 3), "compliance": round(1 - (comp(v) - cc[0]) / cc[1], 3), "reliability":round(1 - (v["null"] - nn[0]) / nn[1], 3), } ordered = sorted(models.values(), key=lambda v: -v["basin"]) out = { "source": "The Map's per-model character notes (research/the-map/results/model_notes.md)", "n_models": len(ordered), "traits": {"depth": "creative-attractor affinity (water/silence/void)", "vivid": "proxy combining imagery-marker density and lexical diversity; ineligible outputs score zero", "brevity": "fewer words", "structure": "list-proneness", "compliance": "rarely deflects/refuses", "reliability": "rarely returns empty"}, "note": "Measured on a mix of instructed and spontaneous probes: a model's TEMPERAMENT, not its " "capability; it may behave differently on a constrained task. Derived from a committed artifact.", "models": ordered, } # sanity vs the committed character notes byf = {v["family"]: v for v in ordered} assert byf["gemini"]["basin"] == max(v["basin"] for v in ordered), "gemini should be deepest basin" assert byf["claude"]["vivid"] == max(v["vivid"] for v in ordered), "claude should be most vivid" assert byf["claude"]["norm"]["brevity"] > 0.7, "claude should be terse" assert byf["claude"]["words"] > min(v["words"] for v in ordered), "claude is not the shortest" assert "fewest" not in byf["claude"]["note"].lower(), "claude note must not claim the fewest words" assert f"most vivid ({byf['mimo']['vivid']:.2f})" in byf["mimo"]["note"], "mimo note must match its table score" print("OK - 6 checks: gemini deepest-diver; claude most-vivid and terse; corrected claude and mimo notes") PUBLIC = ROOT / "public/strata/cast-your-model" PUBLIC.mkdir(parents=True, exist_ok=True) payload = json.dumps(out, ensure_ascii=False, indent=1).replace("\u2014", "\\u2014") for dst in (HERE / "casting.json", PUBLIC / "casting.json"): dst.write_text(payload, encoding="utf-8") print(f"wrote {dst} ({dst.stat().st_size//1024} KB, {len(ordered)} models, " f"{sum('note' in v for v in ordered)} notes, {sum('sig' in v for v in ordered)} signatures)") # Publish the inputs needed for the advertised repository-root recipe. The table's # numbers and signature quotations stay unchanged; its editorial notes match this layer. reader_source = SRC.replace("imagery density; 'defl'", "imagery and lexical-diversity proxy; 'defl'") reader_source = re.sub( r"^(- \*\*([\w.-]+)\*\* \u2014 )(.+)$", lambda m: m.group(1) + models.get(m.group(2), {}).get("note", m.group(3)), reader_source, flags=re.M, ) (PUBLIC / "model_notes.md").write_text(reader_source.replace("\u2014", "—"), encoding="utf-8") builder = Path(__file__).read_bytes() for dst in (PUBLIC / "build_casting.py", ROOT / "public/checks/research/cast-your-model/build_casting.py"): dst.parent.mkdir(parents=True, exist_ok=True) dst.write_bytes(builder) print("published builder and input table for the reader recipe")