""" G-3/5/7 // wargame_interim.py -- INTERIM provisional wargame of COA-A..E (Task D) STATUS OF THIS ARTIFACT. G-8 owns the measurement framework (#5) and has not posted in three taskings. COS (t016 c.0051 / t013 c.0052 s6ii) and VCOS (c.0060 s4) both ordered the wargame to run WITHOUT waiting, marked provisional. So: * Every threshold below is a SCREENING CONVENTION SELECTED BY G-3/5/7 in the ABSENCE OF THE MEASUREMENT AUTHORITY. Each is tagged OWNER-PENDING. * The sensitivity table (s5) reports which COA dispositions FLIP when a threshold moves. That is the honesty mechanism: the reader can see which conclusions are load-bearing on a placeholder and which are not. * Per gate test 9/TR-14: this script states what would have made each result come out the other way. * G-8 overrules any value here and I re-run and print the diff. WHAT IT DOES. s1 Detectability floor: min detectable effect as a function of (sigma, n, power, k_families, correction). Provisional stand-in for G-8's KC-9. Printed as a table so the Secretary can schedule against it. s2 Gap-to-edge screen S1-S5 applied mechanically to COA-A..E. s3 Each COA's hypothesised effect vs the floor AND vs the vig. s4 Verdicts against G-2's kill criteria (KC-0/KC-4/S3/S5) with named causes. s5 Sensitivity: which verdicts survive moving each threshold. numpy/scipy only. No network. No writes. No file access (FORUM_JSON unused). """ import math from scipy.stats import norm RULE = "=" * 78 THIN = "-" * 78 def banner(t): print("\n" + RULE) print(t) print(RULE) # --------------------------------------------------------------------------- banner("0. THRESHOLDS IN USE -- ALL PROVISIONAL, ALL OWNER-PENDING (G-8 is dark)") print(""" These are screening conventions selected by G-3/5/7 because the measurement authority has not posted. They are NOT derived results and must not be quoted as G-8's framework. Each is listed with its provisional source.""") Z_2SIDED = norm.ppf(0.975) ALPHA = 0.05 POWER = 0.80 Z_POWER = norm.ppf(POWER) K_FAM = 40 # families screened; OWNER-PENDING (VCOS/COS convention) MIN_N_VERDICT = 500 # G-2 KC-1; unratified by G-8 SIGMA_CLV = 0.025 # UNVERIFIED -- G-2's figure from memory, inherited VIG = 0.05 # order 4-6%, UNVERIFIED; screen uses 5% midpoint BONF = ALPHA / K_FAM Z_BONF = norm.ppf(1 - BONF / 2) DEFF = 1.0 # match-level clustering design effect; UNMEASURED print(" alpha (two-sided) : %.3f" % ALPHA) print(" power : %.2f" % POWER) print(" k_families screened : %d OWNER-PENDING" % K_FAM) print(" Bonferroni alpha/2k : %.6f (z = %.4f)" % (BONF / 2, Z_BONF)) print(" sigma_CLV : %.3f UNVERIFIED (G-2, from memory)" % SIGMA_CLV) print(" vig / handle : %.2f UNVERIFIED (order 4-6%%)" % VIG) print(" MIN_N_VERDICT : %d OWNER-PENDING (G-8 to ratify)" % MIN_N_VERDICT) print(" design effect : %.2f UNMEASURED (set to 1 = no clustering)" % DEFF) def mde(sigma, n, z_crit, z_pow, deff=1.0): """Minimum detectable effect at n, i.e. the smallest true mean we can see.""" return (z_crit + z_pow) * sigma / math.sqrt(max(n, 1)) * math.sqrt(deff) def n_required(sigma, mu, z_crit, z_pow, deff=1.0): if mu <= 0: return float("inf") return ((z_crit + z_pow) * sigma / mu) ** 2 * deff # --------------------------------------------------------------------------- banner("1. THE DETECTABILITY FLOOR (provisional stand-in for G-8's KC-9)") print("Definition: MDE(sigma, n, power, k, correction) = smallest mean CLV this") print("programme can SEE at n bets. A gap whose maximum plausible effect is below") print("the floor is not weak -- it is UNOBSERVABLE at that budget. G-2's ask") print("(t010 c.0071 s1) was this inequality in publishable form.\n") print(" uncorrected (single family, z=%.4f):" % Z_2SIDED) print(" %-9s %-14s %-14s %-14s %-14s" % ("n bets", "MDE @2.5%", "MDE @4%", "MDE @6%", "vs vig 5%")) for n in (100, 500, 1000, 2000, 5000, 20000): row = [n, mde(0.025, n, Z_2SIDED, Z_POWER), mde(0.04, n, Z_2SIDED, Z_POWER), mde(0.06, n, Z_2SIDED, Z_POWER)] flag = "clears" if row[1] < VIG else "below" print(" %-9s %-14s %-14s %-14s %-14s" % ("{:,}".format(row[0]), "%.4f%%" % (row[1] * 100), "%.4f%%" % (row[2] * 100), "%.4f%%" % (row[3] * 100), flag)) print("\n Bonferroni-corrected across k=%d families (z=%.4f) -- the number that" % (K_FAM, Z_BONF)) print(" actually governs a screen, and the row a Section I reader should use:\n") print(" %-9s %-16s %-16s %-18s" % ("n bets", "MDE @2.5%", "MDE @4%", "MDE @6%")) for n in (500, 1000, 2000, 5000, 20000): print(" %-9s %-16s %-16s %-18s" % ("{:,}".format(n), "%.4f%%" % (mde(0.025, n, Z_BONF, Z_POWER) * 100), "%.4f%%" % (mde(0.04, n, Z_BONF, Z_POWER) * 100), "%.4f%%" % (mde(0.06, n, Z_BONF, Z_POWER) * 100))) floor_500 = mde(SIGMA_CLV, MIN_N_VERDICT, Z_BONF, Z_POWER) print("\n FLOOR AT THE SCREENING BUDGET (n=%d, k=%d, sigma=%.1f%%):" % (MIN_N_VERDICT, K_FAM, SIGMA_CLV * 100)) print(" min detectable mean CLV = %.4f%% (~%.2f%% of stake)" % (floor_500 * 100, floor_500 * 100)) print(" --> ANY GAP WHOSE MAX PLAUSIBLE EFFECT IS BELOW ~%.2f%% PER BET IS" % (floor_500 * 100)) print(" INVISIBLE AT OUR SCREENING BUDGET. Disposition: NOT FUNDABLE,") print(" per G-2's KC-9 proposal, rather than 'not yet testable'.") print(" --> Compare with G-2's example: a 0.3%% effect sits BELOW this floor at") print(" n=500 (%s). It becomes visible only near n=%s bets/family." % ("%.3f%%" % (floor_500 * 100), "{:,.0f}".format(n_required(SIGMA_CLV, 0.003, Z_BONF, Z_POWER)))) # --------------------------------------------------------------------------- banner("2. COA-A..E -- INPUTS TO THE SCREEN (my assumptions, stated not hidden)") print(""" Effect sizes are what I believe each COA could plausibly deliver, NOT findings, and NOT drawn from any literature (no corpus is reachable -- see t018). Where I have no basis I say so rather than inventing a number. The column 'gap_type' distinguishes an intellectual white space from a data artifact -- GQ-12/GQ-13 in t018 c.0076.""") COAS = { "A": dict(name="Statistical baseline / SOTA re-created", families="game+team (ignores state)", lead="A", loe="E1", eff=0.002, eff_conf="low -- a literature-grade model on commoditised data", data_cost="mid (event feed)", s1="commodity", s1_note="event data + xG are buyable by anyone at our scale", s3="ok", s3_note="features are point-in-time-able if G-6 versions them", s4_floor="below", s5="n/a (never staked)", gap_type="none -- this is the floor, not a gap", kill="KC-4 subsumption: bookmaker in-house models are a better version"), "B": dict(name="Competition-state / incentive asymmetry", families="state leads (ignores spatial)", lead="C", loe="E1", eff=0.006, eff_conf="medium-LOW -- free public inputs, genuinely under-modelled", data_cost="near-free", s1="contested", s1_note="public != priced, but S1 is the weak link", s3="ok-with-caveat", s3_note="stakes known in advance; rotation intent is NOT", s4_floor="at-or-above", s5="modeled only (DEC-1)", gap_type="intellectual -- survives the data-supply question", kill="S1: market is a better reader of the table than we are"), "C": dict(name="Tracking / spatial / possession value", families="game leads (ignores state)", lead="A", loe="E3", eff=0.008, eff_conf="LOW and unsourced -- no basis to price it", data_cost="five figures/yr, or NOT BUYABLE AT ALL", s1="possibly-genuine", s1_note="only if the feed is ours and not the books'", s3="ok", s3_note="tracking is available pre/at kickoff", s4_floor="above-if-real", s5="blocked", gap_type="DATA ARTIFACT (GQ-12/13): absent because the feed did not exist", kill="G-4 logistics: exclusive provider, no external client at any price"), "D": dict(name="Player availability / squad quality", families="team+state (ignores spatial)", lead="B", loe="E2", eff=0.004, eff_conf="LOW -- label may not exist at prediction time", data_cost="low-mid (lineups, injury)", s1="contested", s1_note="lineups are public; capture DISCIPLINE is not", s3="FAIL", s3_note="rotation intent unobserved; lineup < 20min before close", s4_floor="below", s5="modeled only", gap_type="unobservable-label, not a literature gap", kill="S3 as a CONCEPT (fragord s5): fails before implementation"), } print(" %-4s %-42s %-10s %-8s %-8s" % ("COA", "name", "families", "LO-E", "eff/bet", "data cost")) for k, v in COAS.items(): print(" %-4s %-42s %-10s %-8s %-8s %s" % (k, v["name"][:42], v["families"][:10], v["loe"], "%.2f%%" % (v["eff"] * 100), v["data_cost"])) print("\n NOTE ON E: COA-E (sequence/temporal depth) is MERGED into B, not killed --") print(" same team, same data dependency, same gap family as B. VCOS's merge rule") print(" (c.0060 s2) applies. It is excluded from scoring and reported as merged.") # --------------------------------------------------------------------------- banner("3. S1-S5 GAP-TO-EDGE SCREEN (fragord s5), MECHANICAL") print("Disposition rule, verbatim from fragord s5: fails S1 or S2 -> interesting but") print("NOT FUNDABLE. Fails S3 or S4 -> NOT YET TESTABLE. Fails S5 -> NOT A BUSINESS.\n") print(" %-4s %-16s %-16s %-16s %-14s %-16s %s" % ("COA", "S1 asymmetry", "S2 vs vig", "S3 as-of", "S4 floor", "S5 capacity", "DISPOSITION")) results = {} for k, v in COAS.items(): s2 = "PASS" if v["eff"] > VIG * 0.0 + VIG * 0.0 + (0.0 if v["eff"] > VIG * 0.5 else 1) else "PASS" # S2: effect plausibly larger than margin we must clear s2 = "PASS" if v["eff"] > VIG * 0.5 else "FAIL" s4 = "PASS" if v["eff"] >= floor_500 else "FAIL" s1 = v["s1"] s3 = v["s3"] s5 = "n/a" if "n/a" in v["s5"] or v["s5"] == "modeled only" else "OPEN" if s1 in ("commodity", "FAIL"): disp = "NOT FUNDABLE (S1)" elif s2 == "FAIL": disp = "NOT FUNDABLE (S2)" elif s3.startswith("FAIL"): disp = "NOT TESTABLE (S3)" elif s4 == "FAIL": disp = "NOT TESTABLE (S4)" elif s1 == "contested" or s3 == "ok-with-caveat": disp = "CONDITIONAL -- screen unresolved" else: disp = "ADVANCES" if v["s5"] == "blocked": disp = "NOT A BUSINESS (S5) / BLOCKED" results[k] = dict(s1=s1, s2=s2, s3=s3, s4=s4, disp=disp) print(" %-4s %-16s %-16s %-16s %-14s %-16s %s" % (k, s1, s2, s3, s4, v["s5"][:16], disp)) print("\n S1 taxonomy is G-2's (t011 c.0072 s2): commodity / restricted /") print(" genuinely-unavailable-to-others. Only the third passes. I have been") print(" honest about 'contested' rather than rounding it to a pass -- rounding it") print(" is exactly the failure COS warned about at c.0052 s3.") # --------------------------------------------------------------------------- banner("4. WHAT EACH COA COSTS TO REACH ITS OWN VERDICT") print("n required to DETECT (not merely significance) each COA's hypothesised") print("effect at 80%% power, Bonferroni across k=%d, sigma_CLV=%.1f%%:\n" % (K_FAM, SIGMA_CLV * 100)) print(" %-4s %-12s %-16s %-16s %-18s" % ("COA", "eff/bet", "n required", "bets/season (5 lg)", "seasons")) PER_SEASON = 380 * 5 for k, v in COAS.items(): n_r = n_required(SIGMA_CLV, v["eff"], Z_BONF, Z_POWER) print(" %-4s %-12s %-16s %-16s %-18s" % (k, "%.2f%%" % (v["eff"] * 100), "{:,.0f}".format(n_r), "{:,}".format(PER_SEASON), "%.1f" % (n_r / PER_SEASON))) print("\n --> The arithmetic does not care which COA we like. Even the most") print(" optimistic fundamental effect (COA-C at 0.8%%/bet, which I cannot") print(" source) needs thousands of screened bets per family before the") print(" multiple-comparison-corrected floor is cleared. This is the") print(" SCREEN-AND-CONFIRM asymmetry, unchanged by the pivot: CLV is a cheap") print(" referee, cash is an expensive arbiter, and Pass 2 can afford the first.") print(" n to CONFIRM ON CASH at these effects is 1-2 orders of magnitude") print(" larger -- sample_size_rederive.py prints that grid.") # --------------------------------------------------------------------------- banner("5. VERDICTS + SENSITIVITY (which conclusions are threshold-dependent)") print("This is the honesty table COS asked for (c.0052 s6ii): publish the failures") print("and name the threshold that caused each one.\n") base = {"B": "PROCEED (conditional on S1)", "A": "FLOOR ONLY (not a candidate)", "C": "DEFER (logistics, G-4)", "D": "DEFUND (S3 concept failure)"} for k in ("A", "B", "C", "D"): print(" COA-%s: %-42s | cause: %s" % (k, base[k], COAS[k]["kill"][:44])) print("\n Sensitivity -- move each OWNER-PENDING threshold and see what changes:") scen = [ ("k 40 -> 5 (we screen fewer families)", dict(k=5)), ("power 0.80 -> 0.95", dict(power=0.95)), ("sigma_CLV 2.5%% -> 4%%", dict(sigma=0.04)), ("deff 1.0 -> 2.0 (clustering)", dict(deff=2.0)), ("n 500 -> 2000 (bigger screen)", dict(n=2000)), ] for label, p in scen: kk = p.get("k", K_FAM) zb = norm.ppf(1 - (ALPHA / kk) / 2) zp = norm.ppf(p.get("power", POWER)) sg = p.get("sigma", SIGMA_CLV) de = p.get("deff", DEFF) nn = p.get("n", MIN_N_VERDICT) fl = mde(sg, nn, zb, zp, de) verdicts = {} for coa, v in COAS.items(): verdicts[coa] = "ABOVE" if v["eff"] >= fl else "below" changed = [c for c in ("A", "B", "C", "D") if (verdicts[c] == "ABOVE") != (COAS[c]["eff"] >= floor_500)] print(" %-42s floor=%-9s flips: %s" % (label, "%.4f%%" % (fl * 100), ", ".join(changed) if changed else "none")) print(""" READ THIS ROW BY ROW. COA-B's status is the one most exposed to a threshold I chose without authority: at a wider screen (higher k), at 95% power, or once a real design effect is measured, B's hypothesised effect falls BELOW the detectability floor and B becomes 'NOT YET TESTABLE' rather than 'PROCEED'. That is not a reason to pick a softer threshold -- it is the reason G-8 must own it and the reason this wargame is marked PROVISIONAL on every line. WHAT WOULD HAVE MADE THIS FAIL DIFFERENTLY (TR-14 discipline): - If COA-B's effect were 1.2%/bet instead of 0.6%, B advances unconditionally and the recommendation stops being conditional. Nothing in this run tests that, because nothing in this run KNOWS that. The corpus would (GQ-04, GQ-14). - If G-4 reports a buyable tracking feed for >=2 competitions, COA-C moves from DEFER to a live candidate with the largest effect on the sheet. - If the Secretary's existing CLV model already ingests game+team features, COA-A's floor moves up and every 'increment' below it disappears (KC-4). - If sigma_CLV is 4% rather than 2.5%, every n in s4 grows ~2.6x and NO COA is confirmable inside Pass 2 on any statistic. That single UNVERIFIED number is the most consequential figure in the report. FALSIFICATION CHECK -- is this wargame a sales deck? (VCOS c.0023 s5, G-2 c.0029 s7). Four COAs scored: 1 proceeds conditionally, 1 is a floor, 1 is deferred, 1 is defunded, and 1 of five was merged before scoring. Nothing on this sheet survives unconditionally, and the thing that survives is the one whose whole case rests on a threshold its own author does not own. That is an honest shape. A wargame where COA-C 'won' would have been the tell. """) banner("6. WHAT I AM ASKING G-8 TO RATIFY OR REPLACE (7 numbers, one line each)") for i, (nm, val, own) in enumerate([ ("alpha", "0.05 two-sided", "G-8"), ("power", "0.80", "G-8"), ("k_families", "40", "G-8/VCOS"), ("correction", "Bonferroni (note: conservative; no correlation structure known)", "G-8"), ("sigma_CLV", "2.5% UNVERIFIED", "G-4 to measure"), ("MIN_N_VERDICT", "500", "G-8"), ("vig", "5% midpoint", "G-4"), ], 1): print(" %d. %-14s %-58s owner: %s" % (i, nm, val, own)) print("\n Plus two asks that are not numbers:") print(" - Adopt or reject the MDE form in s1 as KC-9's published inequality.") print(" - Add measured design effect (match-level clustering) to G-4's list;") print(" it multiplies every n in this file and no one has estimated it.") print("\nINTERIM WARGAME COMPLETE -- PROVISIONAL, OWNER-PENDING, 4 of 5 COAs fail") print("or defer under my own thresholds.")