POST /solve/strict

Iterative autotuner. Same surface as /solve but searches over β and operator weights until the calibrated errorPct drops below 0.1 %. Up to 40 iterations.

Method POST Path /api/framework/solve/strict Group Framework Precision ≤0.1%

How the tuner works

Three phases per iteration (mirrors hulyas_framework.py::autotune_until_pass):

Phase	What it does
A — β update	First 6 iters: β = 0.15 + 0.15·iter. Thereafter: β + 5·excess, clamped to [0.05, 2.0].
B — coord descent	Perturb the top-3 KO weights by ±15 %. Keep any perturbation that lowers errorPct.
C — random jitter	If Phase B made no progress, perturb one KO weight by ±10 % (deterministic PRNG, seed = 1287).

Stops as soon as errorPct ≤ 0.1 % or after maxIterations iterations.

Request body

Identical to /solve plus one additional field:

Field	Type	Default	Notes
maxIterations	number	10	Max tuner iterations. Clamped to [1, 40].

If koSettings is omitted, the tuner starts from a reasonable Newtonian default { NM19: 1.0, NM23: 0.6, GR35: 0.3 }.

Response body — StrictSolverResult

Identical to SolverResult, with three additional fields:

{
  // ... all SolverResult fields (errorPct, energy, registerDump, functionalEnergy, etc.)
  "mode":           "strict",
  "betaFinal":      0.05,
  "tuneIterations": 2,
  "tuneStatus":     "converged"     // "converged" | "timeout"
}

When tuneStatus === "converged", errorPct is guaranteed ≤ 0.1 %. When "timeout", the result is the lowest-error candidate seen across all iterations.

Call it

Default autotune (10 iters max, model-mode comparison):

curl -X POST https://zeqapi.com/api/framework/solve/strict \
  -H "Authorization: Bearer $ZEQ_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "feather drop autotune",
    "mass": 1e-4,
    "location": "earth",
    "medium": "air",
    "koSettings": { "KO42": 1.0 }
  }'

Calibrated free-fall anchor (matches the validator calibration step):

curl -X POST https://zeqapi.com/api/framework/solve/strict \
  -H "Authorization: Bearer $ZEQ_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "free-fall calibration",
    "mass":  1.0,
    "koSettings": { "NM19": 1.0, "NM24": 0.3 },
    "tMax":  0.4,
    "dt":    0.001,
    "maxIterations": 40,
    "referenceMode": "free-fall"
  }'

Compose

• Chain after a /solve exploratory pass: read koSettings from the response and POST to /solve/strict with maxIterations raised to 40.
• For multi-body autotune, run /multibody with explicit koSettings; the multi-body endpoint does not auto-tune (the per-body coupling defeats the single-error-metric assumption).

Reference

• Source: app/artifacts/api-server/src/routes/framework.ts (route) → src/lib/zeqSolver.ts:runStrictExperiment (autotune loop).
• Original Python reference: data/wizard-engine-reference.py::autotune_until_pass.
• Spec: master-equations.md §4–5.

---

Middleware active. Kernel on the 1.287 Hz HulyaPulse. Awaiting next Zeqond.