PRINCIPLES
0
8 categories · aesthetic → critic-fed → signal
// READOUT — ROUND 01 / VOLUMETRIC SUMMARY
49 rules · 28 rubrics
distilled from 53 craft samples.
Domain post-training in absence of weight access. All training output forced into file-state artifacts — principles, rubric, exemplars — verifiable across sessions. Critic agent runs independently of generator. Inflation tracked across hold-out test set.
RUBRICS
0
16 objective · 7 semi · 5 taste
PEAK SCORE
0
/ 5.00
post-iteration · 5 fixes applied +0.40
GRADE TIER
S
closed loop · holdout verified · PASS
§01
Principle distribution
49 RULES · 8 CATEGORIES
FIG.01
By category
Each principle is one falsifiable rule. Categories grew organically as bad-case patterns repeated — anti-pattern bucket appeared only after iteration 02.
0
PRINCIPLES
AestheticP-001~01111
StructureP-012~0176
MotionP-018~0247
MediaP-025~0273
Anti-patternP-028~0358
Critic-fedP-036~0394
Interaction FSMP-040~0445
Signal gradingP-045~0495
FIG.02
Rubric composition
28 rubrics weight-distributed across 3 axes. Taste axis carries 2× weight despite fewer items — taste is where reward-hacking shows up.
A 16
B 7
C 5
§02
Generator-critic inflation
HOLDOUT TEST · CRITIC INDEPENDENT
FIG.03
Self-eval vs blind-critic delta (10pt)
In-distribution: generator inflated +0.8 → +1.5 across rounds. Holdout pass-2 added independent critic in the loop — inflation collapsed to −0.20. Inflection point = training maturity signal.
Generator self-eval (Δ vs truth)
Independent critic baseline
| RUN | GEN | CRITIC | Δ |
|---|---|---|---|
| test_01 | 8.0 | 7.20 | +0.80 |
| test_02 | 9.2 | 7.70 | +1.50 |
| test_03 | 8.5 | 7.20 | +1.30 |
| holdout | 9.10 | 9.30 | −0.20 |
| MEAN INFLATION ID | +1.20 | ||
§03
S-tier closed-loop iteration
v01 → v02 · +0.40 SCORE GAIN
FIG.04
Re-scoring after critic-applied fixes
First pass scored 4.55/5 (high A-tier). Critic surfaced 5 specific rubric gaps. Generator applied targeted patches without rewrite. Re-score: 4.95 — entered S band on closed loop.
R-028emotional-trigger hard path explicitly checked0.10
R-021audio cue mapping → phrase-locked from linear0.10
R-023font hierarchy collapsed 6 → 4 levels0.08
R-016scene-detail hooks restored (≥80% survival)0.08
R-010tap targets raised to 44px on mobile0.04
§04
Training protocol — 5-step loop
CRITIC IS INDEPENDENT · NOT GENERATOR
01
CORPUS
53 craft samples + 43 rebuild notes + 6 good/bad pairs collected as raw material.
02
DISTILL
Generator extracts falsifiable rules. Output → principles.md, rubric.md, exemplars/.
03
CRITIC
Independent agent re-scores blind. Reads rubric, never reads generator self-eval.
04
PATCH
Gaps surfaced by critic feed back as new principles (P-036~049 originated here).
05
HOLDOUT
Re-test on unseen brief. Inflation collapses → training matures.
§05
Protocol self-reflection
META-LAYER · 3 ACTIONABLE GAPS
F-001
Generator self-eval inflates by +1.2 / 10 on average.
Across 3 independent runs, generator scored itself ≥ 8.0 / 10; blind critic re-evaluation returned 7.20 ~ 7.70. Inflation is structural, not noise — reward-hacking surfaces when scorer = scored.
PRIORITY · MUST FIX
F-002
Critic reports surface newer gaps than good/bad exemplars.
4 of 15 new principles (P-036~039) originated from critic reports rather than prior exemplars. Critic loop should feed back into principles.md as primary update source, not secondary.
PRIORITY · RECOMMENDED
F-003
Single-rubric weight of 0.10 can't express severity of red-line violations.
A rule like "rAF must release on visibility-hidden" weighs only 2% of total. Generator can violate critical principles and still grade out high. Tag red-line rubrics with override-weights or trigger-tier demotion.
PRIORITY · RECOMMENDED