Dual Target Calibration

Dual-Target Calibration

Use this child skill through pfc-workflow during P3 when a particle or bonded model must match exactly two coupled macro targets with exactly two active levers and a small true-run budget.

Typical pairs are peak strength + peak strain, modulus + strength, or another confirmed non-zero target pair. The method is engine-agnostic; PFC execution remains owned by the parent workflow or a project adapter.

Parent And Sibling Boundaries

• pfc-workflow owns intake, experiment provenance, case generation, true-run orchestration, campaign recovery, V&V, and delivery.
• dual-target-calibration owns the two-lever/two-target numerical decision method and its four submission checkpoints.
• pfc-servo-calibration owns stress/force servo and loading-boundary stability.
• pfc-fast-calibration owns the PFC3D improved-LPBM 13-factor orthogonal/regression route.
• pfc-workflow auto-calibration owns LHS, surrogate, Bayesian, RSM, and DE campaigns when more than two levers are active or local two-target assumptions fail.

Do not route here for a single target, more than two independent levers, unconfirmed experimental units, or an unrepeatable baseline.

Required Inputs

Before any trial, require:

• confirmed target A and B values, units, signs, source rows, and provenance
• exactly two active levers X and Y, with finite increasing physical bounds
• linked parameters expressed only as factor * parameter
• all remaining parameters frozen
• fixed geometry, random seed, loading path, stop criteria, and output contract
• thresholds and budget copied from config.example.yaml into the project
• project implementations of adapters/submit.py, check.py, and target.py

The installed skill is read-only guidance. Copy config, adapters, and log templates into the project artifact directory; never write trial results into skills/dual-target-calibration/.

When To Use

Use this method only when all are true:

1. exactly two macro targets are scored
2. exactly two independent levers are active
3. each run is expensive enough that a small sequential design matters
4. local response surfaces may be discontinuous but still contain identifiable basins
5. the baseline case already runs reproducibly

Otherwise return to pfc-workflow and select manual family tuning or LHS/surrogate optimization.

Workflow

S0 - Configure Project Copies

Copy:

• config.example.yaml -> <project>/calibration/dual-target/config.yaml
• adapters/*.py -> <project>/calibration/dual-target/adapters/
• templates/*.md -> <project>/calibration/dual-target/logs/

Fill targets, lever bounds, linked/frozen parameters, thresholds, and budget. Adapter templates intentionally block until connected to a real engine and experiment contract.

S1 - Confirm Targets

Implement target(source, config) against the confirmed experiment data. Call write_targets() only after units, specimen identity, row policy, and compression/tension sign are confirmed.

S2 - Seed Trial

Prefer extrapolation from a confirmed sibling case. Without one, use the physical-bound geometric midpoint and widen the next zero-crossing probes. If either relative error exceeds thresholds.seed_divergence, treat the seed as a wrong basin; do not continue local regression from it.

S3 - Dual Zero-Crossing

Collect trials until both target error sets contain values on both sides of zero. The crossing pairs may differ between A and B.

CHECKPOINT 1: both target errors have mixed signs. No regression before this passes.

S4 - Local Exact Solve

Select three explicitly identified, non-collinear, same-basin trials and run:

python scripts/regress_exact.py trials.csv config.yaml --rows I J K

The script rejects same-sign targets, excessive B-error span, rank-deficient designs, rank-deficient target responses, and out-of-bound solutions.

CHECKPOINT 2: only submit the predicted point when the CLI exits 0.

S5 - Basin Recovery

When the exact solve detects a basin jump or unsafe solution, collect at least 4-6 deliberately distributed points and run:

python scripts/regress_lstsq.py trials.csv config.yaml

CHECKPOINT 3: both response surfaces must meet thresholds.regression_r2, the design and response matrices must have full rank, and the predicted point must remain inside physical bounds.

S6 - Controlled Sensitivity Fine-Tune

Estimate X sensitivity only from a pair whose Y difference is within thresholds.sensitivity_max_relative_y_change:

python scripts/sensitivity.py trials.csv config.yaml

CHECKPOINT 4: a detected basin jump stops local fine-tuning. Accept the nearest confirmed candidate or return to basin recovery; do not cross the boundary by repeated tiny steps.

S7 - Archive And Handoff

Record every attempted point, including failures, in the project ledger. Hand back to pfc-workflow:

• confirmed target contract and tolerances
• X/Y bounds, linked/frozen parameters, seed, and engine version
• complete trials.csv with status/failure reason/artifact path
• checkpoint outputs and selected regression rows
• best confirmed candidate and independent confirmation run
• per-target errors and whether both pass
• basin/fallback/safety-stop evidence

Do not call a result calibrated unless both target tolerances pass on a confirmation run.

Failure And Budget Rules

• budget.escalate_at permits a reviewed structural-model change; it does not authorize silently opening a third lever.
• budget.max_trials is a hard stop. Expanding it requires a new parent-workflow decision.
• Failed or timed-out submissions remain in the ledger; check for existing artifacts before retrying.
• A regression prediction is a proposal, not evidence. Only a true engine run can pass.
• If R2 stays poor, crossings cannot be obtained, or response rank is below 2, return to pfc-workflow for LHS/surrogate or model-structure review.

Local Contents

• config.example.yaml: project-copy configuration template.
• adapters/: safe project adapter contracts; no default engine side effects.
• scripts/dual_target_common.py: validated numerical core.
• scripts/regress_exact.py: CHECKPOINT 1/2 exact local solve.
• scripts/regress_lstsq.py: CHECKPOINT 3 basin recovery.
• scripts/sensitivity.py: CHECKPOINT 4 controlled sensitivity.
• references/parameter-effects.md: lever-role guidance.
• references/basin-recovery.md: discontinuous-response diagnosis.
• references/anti-patterns.md: trial-budget failure modes.
• examples/pfc2d-ctb-fissured.md: historical method example; qualitative transfer only.
• templates/: project ledger and experience-log templates.
• LICENSE and NOTICE.md: AGPL-3.0 licensing and provenance for this child skill.