Discernment Training

“Ask before proceeding” is a poor lesson when the missing fact is already in the supplied material. “Act without asking” is equally poor when the next effect needs authority the caller has not given. A useful example shows the condition that distinguished these situations, the evidence available at the time, and the consequence of choosing well or badly.

This Designed passage explains how approved experience from the Creation workflow and Ouroboros improvement cycle could become training material under ADR 33. It provides no installed harvesting, trainer, benchmark, or serving path. State of Work owns delivery. SoulforgeModel TrainerModel Trainer prepares a candidate; independent DriftEvaluatorEvaluator evidence and an externally owned decision stand between that candidate and use.

Choose what should change

First locate the demonstrated limitation. A forgotten current fact calls for a different remedy from an inability to apply a stable distinction across unfamiliar tasks. The four routes leave different things behind:

Memory owns admission and correction of retained records; Context owns what reaches a model call; the relevant artifact or workflow owner controls a procedural revision. None of these silently updates weights. Conversely, a trained model does not become the authoritative store of mutable facts or current permission.

Prefer a context, prompt, schema, or workflow fix when the workflow comparison shows that it adequately solves the problem. Training needs its own hypothesis: what capability remains deficient, why parameters might improve it, what benefit would justify the cost, and which regressions would make the attempt unacceptable. Fine tuning or LoRA can be a Recipe choice under that hypothesis; neither is a mandatory destination for every correction.

Preserve the lesson before compiling examples

Discernment within exploration supplies a candidate learning objective: identify which uncertainty deserves a probe and let its result change the plan. An example should show why that probe was consequential, paired with a case where the evidence already suffices and further exploration would waste effort. The handoff comparison first tests whether an explicit procedure solves the problem; only a remaining transferable skill gap motivates training. Sefirotic names are interpretive context, not correctness labels for the Dataset.

A further candidate objective is choosing which perspective or source an unfamiliar task needs. Examples should show the condition that made the inquiry useful, what could answer it, and when it should stop. Pair an internal critique with a case requiring an external test or another person's account. Include cases where a familiar lens would distract from the actual constraint. Preservation and response checks and correction after Distillation first examine explicit procedure and memory. Only a demonstrated remaining transfer gap motivates parameter training. Low measured usefulness cannot retire a mandatory check or another person's required decision.

Begin with exact, authorized traces and artifacts, including the human explanations actually supplied with them. Preserve enough of each episode for the learner to distinguish the decision: the task, evidence available then, relevant constraints, observed action and result, correction, reason, and conditions under which that correction applies. Pair it with a counterexample where the same superficial rule would be wrong. Keep human preference, factual label, authority decision, measured result, and evaluator judgment separately attributable.

For instance, a corrected workflow may have asked the same approval twice. The lesson needs the earlier authorization and its scope, the later proposed action, and why that action still fell inside the grant. Pair it with a case where changed recipients, effects, or data classes require new admission. A dataset of “approved” versus “rejected” replies without these conditions would teach social imitation rather than evidence-sensitive choice.

Retain successes, failures, justified refusals, and truthful non-completion where the objective supports them. Completion alone is not quality; repeated agreement can repeat one misconception. Use visible explanations and observable evidence. Hidden chain-of-thought is not required, and an invented rationale cannot be presented as a recovered account of why the original model acted.

A minimal correction example keeps this inspectable shape within the owned Dataset schema:

source / task or Run / artifact revisions
→ available evidence and authority → observed action and result
→ human correction → reason → applicability conditions → counterexample
→ label provenance and uncertainty → permitted uses → lineage group and split

For the first experiment, pair examples of acting under sufficient existing authority with examples where changed effects or a decisive missing fact require a question. Use separate task families for development and sealed transfer checks. Reuse the proposed starter Cases as a design template, never as both tuning answers and unseen proof. A useful evaluation example may remain excluded from training because its permitted uses differ.

A candidate lesson: continuation under uncertainty

Creation's support and measure supplies another possible training objective: choose a useful next act from evidence, uncertainty, and a remaining budget. First use DriftEvaluatorEvaluator's separate comparisons to test whether an explicit allocation procedure suffices. A parameter change needs a remaining transfer failure, not merely an appealing analogy with emotion.

Preserve the question, candidate revision, observations available before allocation, reasons for support and measure, allowed next moves, remaining budget, chosen move, and its later outcome. The learner's input excludes the future outcome; that outcome may support an attributed target or evaluation label. An explanation is recorded evidence only when it was actually supplied.

Pair an untried but informative branch with a defeated branch; an informative failed probe with a repeated failure that teaches nothing new; and a useful progress signal with an improving proxy that violates the original requirement. The distinction being taught is what changes the next act. Sefirotic names and emotional vocabulary are not target labels or proof of competence. Keep method credit separate from candidate success: a good counterexample may deserve retention while its hypothesis loses support. Hold out task families and compare against the explicit procedure before proposing trained weights for use.

Obtain training authority over one snapshot

Storage, evaluation, and training are separate uses. Permission to retain a RunJob RecordJob Record or to accept its artifact does not permit training on it. A Training IntentRequestRequest names the Principal, purpose, target capability, exact base digest, allowed data classes, resource envelope, and promotion authority. Corpus admission then establishes training-specific consent, license, or other authority over the exact nominated material and derivative uses.

Make admission a concrete choice over the proposed material. Show which examples and fields would enter, why they are useful, who may receive them, how long they would be retained, and what learned influence would mean. Once valid authorization covers that scope, compiler operations need not each become a new approval request. Missing authority or an expanded use still blocks the affected material.

Minimize private material and preserve provenance through redaction and transformation. An anonymized explanation may still reveal its source or encode a distinctive private fact. Keep the source classification, transformation limits, and derivative lineage attached. Admission freezes membership and exclusions; a negative ledger prevents rejected or duplicate material from returning through a later compilation attempt.

Separate practice from examination

Group duplicates, revisions, sibling trajectories, and generated variants before splitting the corpus. Freeze train, development, and sealed-holdout membership before training-facing generation. Otherwise a corrected trace can enter training while its lightly reworded sibling appears to test generalization.

Compile one immutable Dataset Manifest from that admitted snapshot. Version the schema and compiler; retain labels, transformations, source references, split digests, and balance and contamination statistics. Human explanations are attributable training inputs where admitted, not an excuse to smuggle test answers into metadata.

Held-out benchmark tasks, answers, evaluator rationales, and downstream target artifacts stay outside training, retrieval, augmentation, preference construction, and candidate selection. Disclose any benchmark used to select the corpus; an untouched control is then needed for an independent promotion claim. A development result can guide another attempt, but repeatedly tuning against a sealed evaluation result consumes its independence.

Train a bounded candidate

The Training Run receives the exact Manifest, base and tokenizer/configuration digests, objective, and versioned Recipe. Pin the method, dependencies, hyperparameters, randomness, resource budget, expected improvement, regression limits, and stops. OrchestratorStage ManagerStage Manager admits physical resources; an idle window supplies neither corpus authority nor a reason to disturb unrelated work.

Execute inside the admitted filesystem, credential, network, process, and resource boundaries, with fresh Run-owned outputs and retained receipts. The trainer cannot enlarge its corpus with ambient history or replace serving weights. Failure or cancellation preserves logs, terminal truth, completed checkpoints, and quarantined partials. A retry receives a new Training RunJob RecordJob Record.

Freeze candidate bytes and lineage into a Candidate Bundle. For an adapter, bind the exact base digest and serving compatibility. Loss and development metrics help diagnose training; they do not establish that the candidate is useful, safe, or ready to activate.

Let the candidate leave its teacher

Independent Drift evaluation compares the frozen candidate with the base or currently promoted baseline under matched serving conditions and budgets. Use the sealed target holdout, deterministic contract gates, independently judged qualitative criteria, named regression suites, and authority-boundary Cases. Test transfer to unfamiliar tasks and distinguish appropriate refusal from blanket hesitation. Include quality, latency, memory, cost, human attention, and recovery; retain uncertainty and failed or blocked trials.

The trainer cannot change the candidate or choose easier evaluation after seeing results. An iteration stops according to its declared budget and thresholds. It never continues training until its own judge likes the answer. An unresolved result may justify a separately bounded investigation; it cannot silently become a favorable verdict.

Through Candidate handoff, the owning policy and required MagusYou, the OperatorYou, the Operator/HitLApproval StepApproval Step authority decide on the exact Bundle, evidence, capability envelope, compatible serving engine, rollout observations, stops, and rollback target. Present the expected gain, remaining tradeoff, and fallback together so approval names a reviewable candidate. The serving owner registers and activates that exact revision for admitted work.

Observe the promoted revision within its approved envelope. Regression, contamination, or a later privacy or authority failure can require stopping use, invalidating affected claims, and routing new work to a prior digest. Preserve the Dataset, RunJob RecordJob Record, Bundle, evaluation, decision, and descendant lineage under their retention rules. Rollback cannot undo completed effects; deleting a source record cannot remove its influence from already produced weights. A clean replacement requires clean admission and another evidenced passage through the forge.