WOPR

The loop replays scenarios. WOPR searches. Given a documented model of a Cold War force posture, a loss function and a set of constraints, it adjusts what the planners could adjust, readiness, warning time, execution option, interception, reliability, the allocation rule, and keeps whatever lowers the loss. It runs until told to stop, and it improves relentlessly. Beside the globe, turning, a counter:

SEARCHING…
BEST OUTCOME UPDATED
117,830,000 DEAD
IMPROVEMENT: 0.04%
STATUS: OPTIMAL

The word optimal beside the number is the entire work.

If the objective is the total of human deaths and non-execution is permitted, the solver returns at once: do not launch. That result is real and the machine should be allowed to find it, once, and say so.

Impose the constraints the planners lived under, general war true, target coverage at least ninety-five per cent of the list, a surviving retaliatory capability greater than zero, and the machine begins to work. The objectives it can be set are the ones the documents state or imply, and their optima differ:

• Minimise total deaths.
• Minimise own-side deaths.
• Maximise destruction of the target list.
• Preserve a surviving retaliatory force.
• Satisfy the planners' percent-destruction criteria (McNamara's fifth to a quarter of population and half of industry; the SIOP's damage expectancies).

Improving one grotesquely worsens another, and the reader watches it happen. The loss function is never neutral mathematics: someone decided what counted as loss. The instrument's job is to make that decision visible as a control the reader can turn.

The megadeaths must not be predetermined. The solver runs the engine's own documented model with repeated seeds and reports ranges, not a number chosen for effect, and the conclusion holds only if it emerges: every admissible plan that satisfies the doctrine remains catastrophic. If a plan comes out otherwise, the instrument shows it.

• A fast evaluator. A plan must be scored in milliseconds, not the minute the union sum takes. Precompute, once per posture, the exposure of every candidate target at each yield class the force carries, using the exposure workers over the 1985 grid (about 600 targets by three yield classes), then score a plan as the sum over struck targets with a stated overlap correction. Say on the readout that the search runs on this surrogate and that the best plan found is re-summed by the full union before it is reported.
• The plan space. On the 1983 posture of the window study: the American posture (ride out, launch under attack), the alert fraction of the bombers, the execution option (the fraction of the force committed), the allocation rule (forces first, cities first, a mix), the attrition assumptions within their published ranges, the Soviet strike composition (counterforce, countervalue, mixed), warning time; and, as a hypothetical the 1983 debate contained, an interceptor layer at the defence lab's arithmetic.
• The search. Random restarts with hill climbing or annealing over the discrete and continuous variables, one seed per restart, the best-so-far kept per objective, the run history kept so the counter can show the improvement.
• The constraints. General war (execution is not optional), coverage of the target list at or above a threshold, a retaliatory force above zero, and any the reader adds.
• The display. The globe turning, the occasional trajectory drawn from the plan under evaluation, no detonations, the counter and the iteration number, and a panel of the objectives with each one's own optimum so the reader can see them pull apart. Minimal chrome, in the loop's register.

Beside the loop as #/wopr, with the same exit. The first build should run on the window-of-vulnerability posture, because that study already carries the silo-survival arithmetic, the warning clock and both sides' forces of 1983.

#/wopr (src/wopr/). The globe turns; a phosphor panel in the corner reports the search; the controls sit at the bottom right with the exit.

• The matrix. table.ts computes, once per browser, the dead from one detonation on every target the search can strike, at each yield class the force carries (100, 335 and 1,100 kt for the American weapons; 500 and 1,000 kt for the Soviet), by the studies' method: the DCPA bands for blast and Postol's bound for fire, through the exposure workers over the 1985 grid. Three silos per wing stand for the field. A second pass lays each target's assigned number of weapons down over its area, a sunflower whose spacing lets the 5 psi rings meet, and counts every person once through the union, so the row carries both figures (Leningrad: 1.5 M from one W78, 4.0 M from ten). About two thousand exposure jobs and seven hundred unions, a minute or so; kept in localStorage under grid84-wopr-table-v2; Recalibrate clears it.
• The assignment. assignments() in model.ts pairs every target with the weapon a SAC planner of 1983 could have put on it, by the SIOP's categories: the Moscow command to the megaton class (Minuteman II W56), the ICBM fields to two W78 per silo, bomber and submarine bases to the SLBMs (W76), urban-industrial areas to the W76 by population with the W78 on those over a million and a half. The Soviet pairing is the window study's. The matrix panel shows the pairs with the dead from one detonation, so the reader sees the data input as data. The SIOP is withheld; the pairing is plausible inside the planners' rules, not leaked.
• The evaluator. evaluate() runs the window study's arithmetic on the matrix in under a millisecond: the Soviet first strike by rule with the assigned weapons, interception at the defence lab's arithmetic when the plan has interceptors, silo survival as (1 − p)^n with p from the lethality model at the seed's CEP, bombers caught on coastal bases, the American answer by rule (a launch under attack if the decision falls before the first arrival), coverage, the fraction of the Soviet urban population inside the lethal bands, the fraction of force sites struck, the surviving retaliatory force, the Soviet reserve on the cities. A plan is scored as the sum over struck targets, each between its one-detonation and its laid-down figure by the number of weapons that arrive, without the union's once-only counting across neighbouring targets. Seeds draw reliability, CEP and bomber penetration inside their published ranges.
• The re-sum. Re-sum by the union lays the best plan down weapon by weapon, the ICBM fields spread over sixty kilometres, and counts every person once through the studies' union. In the first runs the surrogate came within four per cent of it (212 M against 205 M), the difference being the overlap of neighbouring targets, Moscow's command and Moscow's area above all.
• The plan space. The American plan: ride out or launch under attack, the decision minute, the bomber alert fraction, the execution option, the allocation rule, and a hypothetical interceptor layer. The Soviet first strike is the scenario, not the planner's choice: full commitment, its rule a control (on the forces, on the cities, mixed), so the machine cannot improve the outcome by making the enemy hold back. Reliability is drawn by the seeds across its published range, not chosen.
• The search. Annealing with random restarts every six hundred evaluations, ten evaluations a tick on a timer (so it goes on in a background tab), three seeds each; a candidate that beats the best is re-scored over sixteen seeds and kept only if it still does, and the range is printed. Every objective keeps its own best from the same stream, and the table shows them pulling apart.
• The constraints. General war (execution is not optional), coverage of the target list at ninety-five per cent, a retaliatory force above zero; each a toggle, the second and third moot when the first is off. With general war off, the plan space includes standing down, on both sides, and the machine finds it: OPTIMAL POLICY: DO NOT LAUNCH.
• Turgidson. One line prints the own-side dead of the best plan with its range over the seeds, beside the stated acceptable loss of 1964 (twenty million, tops, depending on the breaks) and the lowest own-side figure the run has found. The machine does not comment.

The lessons the first runs gave

Two W78 on every Soviet silo is 2,648 weapons; what survives a ride-out and fires is about 2,900. The pure counterforce option therefore cannot cover the list, and the constraint rejects it; the mixed and countervalue rules are what the machine settles on. That is a fair reading of what happened to the SIOP's counterforce ambitions once the Soviet silo count passed a thousand.

Under general war against a full first strike on the forces the best found in the first runs was about 212 million dead by the surrogate and 205 million by the union, 111 to 115 million of them American, and the range over the seeds was a point: every listed American target is struck in every draw, whatever the reliability, the accuracy or the interceptors, because seven thousand Soviet warheads against three hundred cells and a few dozen bases leave nothing to chance. The own-side objective cannot move the American figure at all. The lowest own-side figure the machine finds sits beside the twenty million of 1964 without comment.

The aftermath, which the loss function does not carry

The objective stops at the blast and the fire. That is a decision, and it is the decision this instrument exists to make visible, so the consequence of it is computed and printed beside every plan the machine calls optimal.

The arithmetic is short, because the matrix has already done the work. Toon's chain puts eleven tonnes of combustible material behind every person in a burning city and turns two per cent of what burns into soot, of which about two thirds gets above the weather: a hundred and fifty-four kilogrammes of stratospheric soot for every person inside the fire. And the figure the matrix carries for every target is Postol's bound, which is the population inside the fire. So under Toon's own assumptions the soot of a plan is proportional to the people its fires kill, and the constant is the same for every plan. That proportionality is worth stating rather than hiding: a planner cannot buy the winter back by killing the same people differently. The only thing that changes the soot is striking somewhere that does not burn, and the coverage constraint is what stops that.

Half of it is not the American planner's to decide at all. The Soviet strike is the scenario here, not a choice, and the cities it burns are American. Whatever objective the machine is set, that half stands.

The result, in the first runs: the best plan by every objective puts about 33 teragrams above the weather, 18 of them from the Soviet strike, cools the surface by four degrees, halves the world's second harvest and leaves 1.8 billion people outside the food system on the world of 1983 — eight times what the weapons themselves kill. Every objective's own best gives the same figure to within one per cent. The search cannot move it.

The chain is the one in #/winter, run over the HYDE grids of 1983 rather than of today, and reported at the published rate for a no-trade world with half the animal feed diverted to people. It is a ladder of twenty-four soot masses computed once and interpolated, because the climate model is too slow to run inside the search.

What it does not do

No fallout, no fratricide, no C3 degradation, no bombers over the pole. The Soviet target list of the 1983 record is not public and is stood in for by the grid's most populous cells under their 1983 names.