Spec17 Curriculum Blueprint

spec17 is planned as the bounded intent bridge after the frozen spec16 shared-bundle winner. It is not another narrow raw repair rung. The goal is to move from explicit layout/style prompts to bounded [topic] + [goal] + [audience] planning while keeping the output contract the same: one clean shared [bundle], deterministic lowerers, deterministic compiler, and external content.json.

Why this exists: the run history from spec02 through spec16 shows a stable pattern. Better contracts and better full curricula create the real progress. Post-hoc raw repair rungs, especially warning-language fixes around wrappers and stop markers, regress more often than they help once a strong winner already exists.

Hard Boundary

Prompt Contract

Primary prompt:

[task:svg] [topic:<topic_id>] [goal:<goal_id>] [audience:<audience_id>] [OUT]

Controlled hints may add [emphasis], [constraint], or [content_pack].

Do not give the model explicit layout, form, theme, tone, density, background, or topology counts.

Output Contract

Emit exactly one shared [bundle] only.

Do not emit raw SVG, free-written prose, topic ids, asset ids, or family-specific lowerer syntax.

Visible facts, copy, and numbers remain in external content.json.

Repair Boundary

Allowed deterministic repair is narrow:

  • truncate at the first valid [/bundle]
  • strip copied prompt/control tags outside the bundle
  • dedupe singleton tags when family and form already match
  • fill missing prompt-controlled fields only when family and form already match

Repair must never invent a new family, form, or style bundle.

Founding Families

Family Forms Topology fields Intent profiles
memory_map typed_regions, arena_sections, layer_stack segments, brackets, cards memory_layout_map, allocator_regions, arena_guard_flow
timeline milestone_chain, stage_sequence stages, arrows, footer ir_evolution_timeline, why_pipeline_timeline, operator_changeover_timeline
system_diagram linear_pipeline, build_path, selection_path stages, links, terminal, footer pipeline_overview_flow, build_pipeline_flow, registry_selection_flow

Stage Mix

The curriculum is designed to teach the predictable failure frontiers from the beginning rather than patching them later. All rows must stay compiler-valid and bundle-canonical. No warning-language repair rows are allowed.

Stage A

Surface Weight Teaching purpose
explicit_bundle_anchor 35% Preserve the frozen spec16 bundle grammar and anchor non-regression.
clean_stop_anchor 10% Teach one clean bundle and stop without introducing wrapper-warning prose.
intent_family_bridge 20% Teach topic + goal + audience to family/form routing.
form_disambiguation_contrast 10% Separate sibling forms before the line defaults to one favorite form.
style_bundle_inference 10% Teach theme/tone/density/background inference from audience and emphasis.
topology_budget_bridge 10% Teach topology and density changes from compact/balanced/dense constraints.
family_confusion_contrast 5% Keep the three-family boundary visible from the first stage.

Stage B

Surface Weight Teaching purpose
explicit_bundle_anchor 20% Keep the shared bundle grammar frozen while the prompt contract weakens.
clean_stop_anchor 5% Preserve decode hygiene with positive clean-stop coverage.
intent_family_bridge 20% Make bounded intent prompts a first-class path, not a rare ablation.
topic_goal_recombination 15% Test whether the model recombines seen topics and seen goals instead of memorizing pairs.
form_disambiguation_contrast 10% Continue pressure on sibling-form choice under weaker prompts.
style_bundle_inference 10% Keep style inference from collapsing to the dominant bundle.
topology_budget_bridge 10% Keep count and density planning inside the bounded contract.
paraphrase_bridge 5% Teach wording robustness without changing the underlying task.
family_confusion_contrast 5% Block hidden family imbalance before promotion.

Predicted Failure Frontiers

Failure frontier Why it is predictable How the curriculum teaches it up front
default family attractor Shared-family lines often collapse to the most familiar family when the prompt weakens. intent_family_bridge, topic_goal_recombination, family_confusion_contrast
default form attractor Sibling forms inside one family are easy to blur together under CE pressure. intent_family_bridge, form_disambiguation_contrast
default style attractor Theme/tone/background often drift to the training-dominant bundle. style_bundle_inference
topology/count drift Counts often default to familiar values when the prompt only gives high-level intent. topology_budget_bridge
topic leakage into output Once topic and goal enter the prompt, the model may try to serialize them into the bundle. explicit_bundle_anchor, clean_stop_anchor
paraphrase brittleness Minor wording shifts can change the chosen bundle if the prompt surface is under-trained. paraphrase_bridge
hidden family imbalance One family can improve on visible probes while another silently collapses on hidden probes. family_confusion_contrast, balanced hidden splits

Holdouts And Promotion

Before Training Starts

Source of truth in the repo: version/v7/reports/spec17_curriculum_blueprint.json, version/v7/scripts/audit_curriculum_blueprint_v7.py, and the generated audit report under version/v7/reports/.

Related Pages

Image
100% | |
Scroll to zoom | Drag to pan | W/H to fit | 0 to reset | ESC to close