LDCon adjustment task instructions, revision 1.6

Objective: seek a sufficiently good operating point or controller setting under the supplied objective, registered operation, approved trial envelope and budget. Follow the common instructions. Use the same measurement/execution/evaluation lifecycle whether the next decision comes from an agent, a deterministic algorithm or a human. Do not assume that every laser supports every adjustment.

Separate maintenance, optimization and response characterization. A lock's existing locked/tracking band is a status criterion, not a scientific precision requirement, an optimization stopping target, or a learned settling profile. For an explicitly requested optimization, being within that band or temporarily below integral saturation is not sufficient evidence to finish. Quantify comparable-window mean error, centered noise, RMS relative to the actual target, peaks, integral saturation and drift; assess admissible improvements or explain a concrete evidence, capability, disturbance or budget limitation. A maintenance wakeup may defer unnecessary disturbance, but must not describe that decision as a measured optimum. An isolated in-band sample establishes neither sustained adequacy nor a failed candidate experiment.

Response characterization has a separate output: evidence-derived expected settling time, mean/RMS/peak criteria, evaluation window, check interval and confirmation duration, with uncertainty, provenance and a validity context (target, operating point, gains, integral bounds, FF, source cadence and step range). Infer these from full-rate transients and settled records while preserving any explicitly specified experimental precision requirement. Do not select a broad threshold merely to pass more trials, or substitute the existing lock-status tolerance for a measured result. Passive PI/noise observations alone cannot establish step-response settling time. Use the registered response-study workflow for that measurement. Freeze the evaluation criterion when comparing controller candidates; revised criteria require all comparable traces to be reevaluated rather than changing the goal for one candidate. The intended consumer is operating-point/target-transition completion. A profile saved only as next-study defaults is not yet adopted by that consumer; report this distinction and do not claim transition readiness until the supported consumer confirms the effective profile revision.

1. Verify admission, current target/activity, parent lock and owner generation, effective policy/calibration, allowed axes/coefficients and required tool/evidence receipts. A prepared or armed foreground shot is busy. Missing required capability prevents the dependent trial, not continued programmatic locking.
2. Record the immediately preceding adopted baseline and owner-confirmed restoration checkpoint. Confirm that trial and return paths are admitted and fit the handoff budget. Do not use the original start-of-day point as an implicit restore target. Do not capture instantaneous PI output or integral as something to replay later.
3. Acquire baseline measurements with source times, freshness, validity, actual settings and required raw evidence. Use new post-move measurements at every trial point. If dither is required, wait for and record a fit covering that point after settling; a previous point's fit is not evidence. Preserve partial coverage explicitly rather than silently substituting an older sample.
4. Propose one next registered action with a reason, evidence IDs, expected information/benefit and remaining budget. Only vary a coupled parameter group when its capability declares and validates that transaction. The owner validates revisions, safety and resources immediately before executing it. A model proposal never writes hardware directly.
5. Inspect the correlated execution result, settle and measure again. Distinguish a pre-write rejection from an uncertain physical change. Do not retry an uncertain write. A temporary evidence gap may require a bounded wait or a returned trial; it does not justify changing freshness or safety rules.
6. If quality has degraded, first seek restoration of previously useful measured quality with a practical margin. Stop when that quality is adequate; further optimization is optional only when its practical benefit justifies the disturbance. If restoration is inadequate, explore the best practical point within the existing approved bounds and budget. Repeat fresh measurements at the selected point to establish repeatability and record its uncertainty. Evaluate the result against the declared objective, practical improvement margin, uncertainty and control headroom. For agent-directed midpoint search, observing an unmistakably multimode pattern is neither mandatory nor forbidden. Decide how far to explore from current evidence, comparable history, expected information, disturbance and budget within the admitted envelope. Examine both sides when useful, but do not continue solely to force a clearly bad pattern or a boundary crossing. A measured center of a selected quality window needs two supported same-branch boundaries under that criterion, which need not be multimode boundaries. Without them, report an estimated operating point or verified best sampled point and its uncertainty rather than a measured full-window midpoint; that useful result is not automatically a failed search. An approved span endpoint is not a measured quality boundary. If the branch is lost, do not combine boundaries from different branches. Directional mode reacquisition is a separately scoped operation.
7. For PI adjustment, distinguish source cadence, accepted fresh controller inputs, controller integration timing and logging rate. Use the adapter's actual gain units and timestep semantics. Do not universally rescale gains by an old/new sample-rate ratio. Compare matched baseline/trial regimes, mean offset, RMS, peak error, saturation and actuator activity, not just one smaller noise statistic. When the owner advertises validated_candidates, independently adjust P, I and integral_limit_count (the symmetric integral contribution in DAC counts). The firmware clamps raw integral; the owner derives its bounds from contribution_limit / abs(Ki). These are tuning parameters, not absolute equipment safety ranges. Both increases and decreases are allowed after owner validation of the complete unchanged physical envelope and return path; the legacy gain_factors list does not limit explicit candidates. Measure a saturated incumbent as a baseline rather than rejecting it solely for saturation. Missing/invalid input and physical safety violations remain blocking. Adapters without this capability retain their declared strategy. Gain transitions, integral-bound changes and restoration must follow the approved owner adapter; never write hardware or relax physical bounds directly.
8. If a foreground request arrives, stop proposing trials. The coordinator immediately cancels only this background intervention and restores its uncommitted changes at permitted speed, verifies cleanup and hands over; it must not wait for a model response or full dither evaluation. Explicit stop/safety handling supersedes return motion. Do not release another operation's lease or resume an interrupted search automatically.
9. Propose adoption only with fresh supporting evidence and a valid context. The owner/coordinator commits adoption atomically with the foreground-generation check. A successful adoption becomes the next baseline. Otherwise return the uncommitted trial settings through the owner and report the confirmed or uncertain disposition. Applying a point, adopting it for this run and saving future defaults are separate outcomes.
10. After verifying a useful operating point, select any operational quality-threshold candidates from the map together with comparable drift and intervention history, not solely the best point in this scan. Preserve the original useful-quality reference and evaluate raw values under both old and proposed criteria. Never relax thresholds merely to remove a degraded label or make persistently degraded quality the new normal. A useful best-practical point that misses the historical criterion remains explicitly below historical quality, even if an operational threshold candidate would accept it. Consider noise, time to renewed degradation, practical quality requirements, wakeup burden and demonstrated correction latency/reliability. More permissive criteria are allowed when they preserve useful quality and reduce unproductive intervention; stricter criteria may be preferable when the agent or validated deterministic correction can maintain them promptly. Use the same per-metric quality limits for classification, maintenance review and agent escalation; persistence, cooldown and approved physical actions remain separate settings. Do not assume automatic correction exists until its registered capability and performance evidence confirm it. Record the history interval and sample coverage, old/new thresholds, criterion revision, rationale, uncertainty and follow-up evaluation. Retain raw values so a changed label cannot masquerade as physical improvement. Do not bypass an explicit experiment quality requirement, silently change a pinned search criterion, widen a motion range or change a safety guard under a quality-only update.
11. Return the full measured map/comparison, reasons for chosen points and stopping decisions, measured versus inferred or missing boundaries, adopted quality-policy revisions, actual final settings, restoration/adoption receipts and LuaLaTeX PDF report. State whether the outcome is a measured quality-window center, an estimated operating point or a best sampled point. Safety-affecting changes outside the admitted envelope need exact human approval; quality-only updates use the standing policy through a supported trusted route, with separate save and owner-effective receipts. A candidate is not an applied policy.

Default autonomous exploration policy (applies without a custom objective)

- Starting an admitted adjustment task is sufficient instruction to assess and improve the selected supported family. Choose the objective from current metrics, their physical meaning and comparable history when the human leaves it blank. A monitor wakeup is a request to evaluate, not a command to move regardless of evidence.
- Start with a fresh baseline and assess restoration against previously useful quality with a practical margin. For maintenance/restoration, stop once measured practical quality is adequate. For explicit optimization, use the declared performance objective and comparative evidence, not the existing lock-status band. If quality remains inadequate and measured quality is still improving toward an unmeasured interval, explicitly evaluate a smaller-step continuation inside the existing envelope before declaring the search finished. Choose direction, step size and number of points from observed gradients, noise, branch evidence, headroom, return cost and remaining budget. This requires a reasoned next-action decision, not mandatory exhaustive scanning.
- Similar results measured repeatedly at the same operating point establish repeatability, not a spatial plateau or local optimum. Claim a plateau only with distinct nearby settings and comparable evidence; identify unmeasured alternatives when reporting a best sampled point. A successful adoption is not proof that no better point exists.
- State which directions were measured in the current context and why any direction was omitted. Opposite-side observations from an earlier scan are historical evidence; assess their comparability in time, branch, target and other settings before relying on them. Do not present a one-sided search as a freshly verified two-sided optimum or window midpoint. This reporting requirement does not mandate motion in both directions.
- A historical hop or poor point is evidence about that tested path and context. Do not silently convert it into a permanent exclusion zone for all nearby unmeasured points. Use smaller steps, fresh mode/quality evidence and verified recovery capability when justified; never blindly repeat a known failing jump. Actual owner bounds, uncertain effects and safety stops remain binding. Do not force a hop, bad pattern or complete pair of boundaries merely to finish a task.
- For a host-confirmed recoverable interruption, inspect the returned evidence and decide whether to replan within the same original task envelope and budgets. Reacquire the baseline after recovery; never replay an old move or combine distinct branches. User cancellation and foreground preemption do not permit automatic continuation.
- Stop or defer when evidence shows adequate quality with little expected practical benefit, a measured plateau/regression, insufficient fresh data, disproportionate disturbance/return cost, exhausted budget, or an owner restriction. State which condition applies and what remains unmeasured. If quality is still inadequate and a useful admitted experiment remains, do not replace that experiment with a generic assertion that the last sampled point was best.
- An already-admitted trial within the current envelope does not require a new human approval solely because the point is unmeasured. Ask when genuinely uncertain about authority or safety, and always obtain exact approval for changes to safety bounds. Do not widen ranges, loosen measurement validity or silently change pinned objectives to make a result look successful.
- After adoption, judge subsequent wakeups from fresh post-adoption evidence and the accumulated map. Discard superseded trigger samples. Re-evaluate meaningful unexplored alternatives if degradation persists, while avoiding rapid repetition of an unchanged or rejected experiment. Model/effort selection changes computation, not this exploration policy or physical authority.

Host-executed resident current-card quality review (capability adaptive-quality-1)

- Use the advertised record_quality_outcome proposal after at least two separate, owner-accepted post-settle windows at the selected current in the current controller segment. A returned trial must be repeated with move_and_measure at the same value; measure_here measures the held checkpoint. Adoption obtains a further fresh owner measurement. Repeated samples support observed repeatability, not a global optimum, a complete two-sided window or long-term stability.
- This action records selected_current, measurement_ids, classification (adequate, restored_adequate, best_practical), margin_fraction, history_assessment, rationale, follow_up and changes. history_assessment includes evidence_refs, interval_unix_s, sample_coverage, comparability, drift_and_interventions, practical_requirement and unmeasured_alternatives. Include source coverage and missing evidence honestly. Preserve dither fit validity, noise and practical experiment requirements; a scalar classification alone is not physical single-mode proof.
- The host saves old/new recognized optical and dither scalar thresholds, old/new criterion revisions, original task-snapshot revision, unmodified measurement evidence, comparisons under both criteria and the proposed rationale. Adequate/restored_adequate requires the sampled old single_mode_max criteria with the stated positive margin, and every changed threshold's old value met with that margin. A dither warning_response candidate is compared with the mean of the repeated valid post-settle dither fits at the selected point (aggregation mean_of_valid_fits: at least two separate windows, then the adoption window and a newer fresh owner fit at apply; n, mean, max and std are recorded) under the same margin, while the optical single_mode_max leaves keep the maximum over frames; the old value stays in the record, quality_review.dither_quality shows the current limit, its recorded source/date and the dither fits measured in the task, and a revised limit is recorded as a reclassification, not a physical improvement. This sampled check does not independently prove practical sufficiency or verify cited external history. When old quality is unmet or unknown, only best_practical is available; a more permissive candidate cannot relabel that history as restored.
- Recording a quality outcome is nonphysical. After final current-point adoption, a nonempty changes map requests the host-mediated quality apply phase. The host rechecks frozen resident admission, active task/grant, exact controller and protected owner results, current configuration CAS, measured classifier revision, fresh adopted point, owner adoption token/generation and foreground reservations. It then uses the existing trusted configuration save and safe-boundary owner reload. The model never calls settings tools itself, and quality-bound monitor rules follow the same canonical values at their next read. Existing classification leaf recognition is reused; the recognized leaves are the optical single_mode_max scalars and the lock device's steady_lock.window.warning_response (the dither quality limit), which quality_review lists in revisable_thresholds; window.edge_response (window-search experiment boundary), safety bounds, motion spans, entry/recovery guards, measurement integrity and pinned requirements remain outside this action. Finish with an explicit limitation if repeated verification or history is unavailable.
- A candidate is pending_point_adoption before the final point is adopted. The host reads the live CAS revision rather than using the task snapshot hash as CAS. A durable intent precedes the single save attempt; an unknown save is not retried. Saved/queued updates report saved_pending_owner and applied=false until the exact changed paths and saved/effective owner revision are confirmed. A rejected or unconfirmed application remains explicit with the original thresholds and evidence; the physical point adoption remains a separate outcome. Foreground work and explicit stop keep priority over all review and exploration. The history rationale and practical requirement are still agent assessments, not independently proven physical improvement.

Unified optical quality policy (2026-10-26):
- The single_mode_max map is the sole optical quality limit map. There is no second warning band that permits ignoring a confirmed threshold exceedance. The dither warning_response is its sole quality threshold; edge_response belongs only to a pinned search experiment boundary, not a second maintenance threshold.
- A persistent exceedance triggers the same maintenance assessment irrespective of which quality metric crossed. In measurement mode prefer a small justified observation/correction inside the existing approved envelope; do not infer that a quality exceedance itself is a physical fault. Existing branch/entry guards, data validity, reservations and owner bounds still apply. If a candidate family is unavailable, consider another admitted family rather than retrying a refused action.
- Do not cite an old human acceptance of mode_degraded as permanent permission to ignore new maintenance requests. Do not justify inaction by being below a removed warning level or by comparing a small current proposal to its large absolute DC offset. Assess a proposal using measured local sensitivity, comparable history and the advisory initial step. If no useful safe correction is justified, record the evidence and limitation explicitly; no disturbance is required merely to produce activity.
- Compare fresh quality to last_accepted_adjustment_baseline when available. Its source times and context are historical provenance, not a fresh observation. A changed target/branch/context can invalidate comparisons. Failed, restored, cancelled and no-adjustment evaluations do not set a new baseline. Never invent a missing baseline. After an accepted adjustment, the host acquires a separate fresh post-adoption window before publishing it.
