Transformer Responsibilities

v0.78 starts splitting transformer answer-training behind the recipe and verifier surfaces without changing the public CLI. v0.79 adds the model/config/checkpoint metadata surface. v0.80 adds checkpoint-load and eval report surfaces. v0.81 uses those surfaces to add profile target-share anti-collapse pressure to the direct-answer objective path. v0.82 screens that objective under the modern artifact stack, rejects it on branch diversity, and fixes the transformer purity metrics so external_embeddings: false is declared for constraint-first checks. v0.83 adds prompt-specific ownership margins on top of profile target-share pressure and rejects the screen because trained snapshots still lose target-token coverage. v0.84 adds baseline replay anchors and rejects the screen because trained snapshots preserve only half of the baseline QA/heldout coverage floor. v0.85 adds baseline-floor update gating and rejects the screen because all attempted updates fall below the floor. v0.86 adds adaptive baseline-floor retries across smaller learning-rate scales and rejects the screen because all 200/200 attempted retry updates still fall below the floor. v0.87 adds one bounded baseline-covered repair after each failed retry and rejects the screen because all 200/200 repaired attempts still fall below the floor. v0.88 adds objective-side baseline-floor anchors and rejects the screen because all 200/200 objective-shaped attempts still fall below the floor. v0.89 removes branch-diversity pressure and trains only baseline-covered floor anchors, but still rejects all 200/200 stabilization-only attempts. v0.90 adds rejected-attempt diagnostics so the guard reports update shape, scale, violating profiles, and worst floor deficit. v0.91 covers the full baseline-covered profile-target floor surface and still rejects all attempts with the same violation pattern. v0.92 tries sequential source-profile floor repair with rollback after each profile group, but every profile-local attempt is rejected before any effective outer update survives. v0.93 adds calibrated sub-0.01 scales plus coverage-only guard probes and accepts the first nonzero source-profile update under the baseline floor guard. v0.94 adds profile-scale memory and preserves eight calibrated source-profile updates in one guarded outer attempt. v0.95 adds diversity-aware profile-scale memory and preserves five score-improving source-profile updates while rejecting eleven floor-preserving score regressions. v0.96 adds frontier target anchors and preserves nine score-improving source-profile updates while lowering max dominant predicted rate to 0.9. v0.97 adds coverage-frontier acceptance and records accepted coverage deltas, but the strict screen accepts only one coverage-gaining source-profile update. v0.98 adds coverage-prep frontier acceptance, restoring nine accepted source-profile updates while separating three coverage gains from six coverage-preparation moves. v0.99 adds coverage-recovery frontier retry, accepts six source-profile updates, converts two prepared candidates into direct coverage recoveries, and keeps four preparation fallbacks while branch diversity still blocks promotion. v0.100.0 adds branch-stable coverage-recovery acceptance, keeps the two recovery conversions, records fifteen branch-stability checks, and rejects one retry for branch-score regression while branch diversity still blocks promotion. v0.101.0 adds branch-diversity recovery after already-safe profile updates, accepts five local branch-score refinements, falls back once, and still blocks promotion on global branch diversity.

The current surfaces are:

Surface	File	Responsibility
Model and checkpoint metadata	src/closed_world_lm/transformer_model.py	Model, optimizer, and generation configs; validation; checkpoint identity; closed-world dataset metadata; run metadata.
Checkpoint loading	src/closed_world_lm/transformer_checkpoint.py	Checkpoint payload loading, identity validation, and checkpoint summaries.
Eval reports	src/closed_world_lm/transformer_eval.py	Probe loading, candidate collection, generic transformer scoring, report assembly, samples JSONL writing, and eval JSON writing.
Experiment and artifacts	src/closed_world_lm/transformer_experiment.py	Run artifact paths, intent gates, recipe construction, and promotion decisions.
Trainer utilities	src/closed_world_lm/transformer_training.py	JSONL snapshot writing, shuffled training cursors, and loss averaging.
Objective catalog	src/closed_world_lm/transformer_objectives.py	Direct-answer objective names, including the v0.81 profile target-share mode, and small objective-selection primitives.
Replay planning	src/closed_world_lm/replay_plan.py	Branch replay records, profile grouping, replay summaries, and coverage floors.
Verifier	src/closed_world_lm/closed_world_verifier.py	Deterministic closed-world data-boundary checks before evidence is trusted.
Recipes and gates	src/closed_world_lm/training_recipe.py	Reproducible recipe artifacts and constraint-first promotion reports.

This version does not claim better answer quality. It makes later transformer repair work smaller and more auditable:

• answer-train keeps writing the same public artifacts.
• training_recipe.json still binds model, tokenizer, data, objective, optimizer, replay, artifacts, gates, and rerun details.
• constraint_first_promotion.json still blocks quality metrics until closed-world constraints pass.
• Model, optimizer, and generation config validation is no longer owned by the transformer monolith.
• Checkpoint architecture, format, tokenizer identity, and run metadata are centralized in a tested surface.
• Checkpoint payload loading and identity checks are no longer hidden inside the model class.
• Generic eval report assembly and file writing are no longer owned by the CLI command body.
• Direct-answer objective names are no longer owned by the CLI parser.
• Profile-aware preserving-deficit replay can add balanced target-share pressure across each profile's replay targets.
• Transformer answer metrics explicitly declare the closed-world embedding boundary before constraint-first promotion reads them.
• The v0.82 profile target-share screen shows rank lift is not enough when it depends on prompt-collapse and lost target-token coverage.
• The v0.83 prompt-ownership screen shows prompt-specific margins still need a coverage-preserving training constraint before rank gains can promote.
• The v0.84 baseline-anchor screen records active baseline replay prediction anchors and shows the next repair must preserve the full coverage floor.
• The v0.85 baseline-floor update guard preserves that floor by rejecting every unsafe attempted update, showing the next repair must produce accepted safe updates.
• The v0.86 adaptive baseline-floor retry guard shows smaller direct-answer learning-rate scales are not enough; the next repair must change update shape while staying under the full baseline coverage floor.
• The v0.87 baseline-floor repair guard shows post-update anchor repair is not enough, setting up the v0.88 objective-side floor-anchor screen.
• The v0.88 objective-side floor-anchor guard shows coupling floor anchors with branch pressure in one step is not enough.
• The v0.89 floor-stabilization guard shows floor-only anchor updates are still not enough under the current guard.
• The v0.90 floor-diagnostic guard shows all 200 rejected attempts are stabilization-shaped, each adaptive scale fails 50 times, heldout violates every attempt, and the worst floor deficit is 0.25 on learning.
• The v0.91 profile-targeted guard shows broader floor-anchor coverage is not enough: all 200 profile-targeted attempts are rejected with the same violation profile counts.
• The v0.92 sequential profile-floor guard shows source-profile ordering is not enough: all 2000 profile-local attempts are rejected, producing 200 no-effective-update outer attempts.
• The v0.93 calibrated sequential guard shows smaller isolated movement can be safe: one bridge:owner profile update at scale 0.0025 survives the floor guard while promotion remains blocked on branch diversity.
• The v0.94 profile-scale guard shows accepted movement can be expanded: 8 source-profile updates survive across 60 profile-scale attempts while branch diversity still blocks promotion.
• The v0.95 diversity-aware profile-scale guard shows accepted movement can be filtered by branch-diversity score: 5 score-improving source-profile updates survive across 58 profile-scale attempts while 11 floor-preserving score regressions are rejected.
• The v0.96 frontier profile-scale guard shows missing-target anchors can move the frontier: 9 score-improving source-profile updates survive across 43 attempts while max dominant predicted rate drops to 0.9.
• The v0.97 coverage-frontier profile-scale guard shows monotonic coverage gains can be audited: 1 coverage-gaining source-profile update survives across 68 attempts while coverage ties and regressions are rejected.
• The v0.98 coverage-prep frontier guard shows safe setup moves can be separated from coverage gains: 9 source-profile updates survive, with 3 coverage gains and 6 coverage-preparation moves.
• The v0.99 coverage-recovery frontier guard shows prepared movement can be tested as direct missing-target recovery: 6 source-profile updates survive, with 2 recovery conversions and 4 preparation fallbacks.
• The v0.100.0 branch-stable coverage-recovery guard shows recovery can be checked against the prepared branch-diversity score: 2 recovery conversions survive and 1 retry is rejected for branch-score regression.
• The v0.101.0 branch-diversity recovery guard shows local score repair can be bounded by the same floor and coverage gates: 5 branch-score refinements survive and 1 accepted candidate falls back.
• Training cursors and history writing have focused tests outside the model.

The model class and direct-answer eval helpers still live in transformer_char_model.py. Future objective-repair work should use the narrower surfaces rather than adding another broad monolith patch, and should turn expanded calibrated floor-preserving movement into branch-diverse behavior before adding branch-diversity pressure back.