Scaffold pattern

Appendix F — The Scaffold Withdrawal Pattern

F.1 Why this is constitutional

The architecture repeatedly faces problems whose unrestricted form is initially too noisy or sparse to teach anything reliable:

• Tolaria cannot attribute branch differences while execution noise is unknown;
• Narset cannot learn intervention timing when every host trajectory diverges for unrelated reasons;
• Momir cannot learn useful design when almost every unconstrained graph proposal is invalid or useless.

The response is not to pretend the unrestricted problem is easy. It is to establish a controlled classroom in which causal signal exceeds nuisance variation, calibrate the instruments there, and then remove the classroom constraints one at a time.

This is a uniform theory of generalisation rather than three unrelated training tricks.

F.2 The three primary scaffolds

Scaffold	Protects against	Acquisition regime	Relaxation	Withdrawal gate	Retained role
Tolaria Academy exactness	Attribution error	Bitwise-exact paired worlds	Repeated stochastic worlds and surrogate calibration	Ranking, decision, uncertainty and tail criteria	Causal oracle, CI, regression and disputed-case retest
Repeated host trajectories	Host variance	Fixed acquisition seeds and identical trajectories	Held-out in-family initialisations and one-axis variation	Stable tactical timing and lifecycle outcomes	Policy-language and regression fixtures
Sarpadian reference ancestry	Generator collapse	Reconstruction, imitation and bounded mutation	Ancestry dropout and partial de novo design	Structural validity and positive coverage with null ancestry	Blinded controls and historical precedent

F.3 Tolaria's training wheels

Bitwise replay is deliberately strict because it establishes the reference equality:

\[ \text{branch difference} = \text{intervention effect}. \]

Field execution instead operates under:

\[ \text{branch difference} = \text{intervention effect} + \epsilon, \]

where \(\epsilon\) must be measured, modelled and priced. The progression is not exactness followed by sloppiness. It is:

text unknown noise ↓ eliminated noise ↓ measured noise ↓ modelled and tolerated noise

Academy exactness therefore remains available after withdrawal. It defines the unit of causal measurement, calibrates Field QA, diagnoses divergence and adjudicates cases too close to call under stochastic execution.

F.4 Independent gate ownership

The primary gate owners are:

• Urabrask certifies Tolaria execution and Field-surrogate evidence against Academy results;
• Augustin authorises use of that evidence for a declared assurance class and applies uncertainty margins;
• Narset curriculum evaluation certifies host-distribution generalisation;
• Momir curriculum evaluation certifies null-ancestry design competence;
• Leyline validates that the run's declared ScaffoldState matches the actual configuration.

No owner may certify another dimension merely because its own dimension is ready.

F.5 One-axis withdrawal and interactions

The default confirmatory sequence is:

text Academy exact + repeated hosts + ancestry ↓ relax execution only Field-calibrated + repeated hosts + ancestry ↓ relax host distribution only Field-calibrated + held-out hosts + ancestry ↓ withdraw ancestry only Field-calibrated + held-out/open hosts + null ancestry

The order may change, but only one scaffold changes per confirmatory transition. After the main effects are known, declared interaction experiments test whether scaffold effects are correlated.

A multi-axis transition without those controls is not necessarily unsafe, but it is scientifically uninterpretable and cannot support an attribution claim.

F.6 Adding a fourth scaffold

Any future proposal for restricted grammar profiles, fixed blend schedules, synthetic tasks, single-slot hosts, known-rank repairs or other training wheels must answer:

text What failure mode does this scaffold protect against? What is the Learn-the-Land regime? What is the controlled relaxation path? What measurable gate permits withdrawal? Who owns that gate? What remains as a reference after withdrawal? What other scaffolds might it correlate with? How will ambiguous Field cases escalate?

A scaffold proposal without those answers is an undocumented permanent assumption.

F.7 Review shorthand

The review question is:

Where is the classroom, where is the graduation test, and where is the retained laboratory?

If the classroom can never be left, the system has not generalised. If the laboratory is dismantled after graduation, the system can no longer calibrate or explain itself.