rtc-loop — Reality-Topography Compilation Loop
DITEMPA BUKAN DIBERI — Forged 2026-09-07 from live evidence, not theory. Origin: RTC test #1 (FI-008) — V1 rejected at the door, V2 delivered with disclosure. First receipt:
/root/forge_work/rtc-test/RECEIPT-001.md(Telegram msg_id 109575) Scar: SCAR-RTC-001 — constitutional tripwire, not a bug.
When to Use
- ANY text-to-image / image-generation dispatch in the federation (every diffusion call)
- Before writing the generation prompt (SPEC + COMPILE steps)
- After generation returns (AUDIT step) — before anything is delivered to a human
When NOT to Use
- Image editing / inpainting (different dimensionality regime)
- VQA / image understanding (no pixel output)
- Video / audio / somatic (different modality semantics)
The Loop (mandatory order)
SPEC → COMPILE → GENERATE → AUDIT → {REJECT → REGENERATE} → DELIVER
- SPEC — parameterized reality spec (JSON), never prose wishes. Structure before surface:
structure → material → topology → observation → optical. Includeverification_targets: countable, falsifiable checks the audit will run. - COMPILE — spec → dense layered prompt. Bake verification targets in. Include
negation anchors for known prior tropes (see below). Compute
density_lower(forge-vision-densify stage-1 heuristic). - GENERATE — one diffusion call. Do not deliver. Hold at the door.
- AUDIT — VLM witness against the SPEC's
verification_targets, not against vibes. Count things. Check symmetry, structure, lighting direction. Demand a verdict:PASS | PASS_WITH_NOTES | FAIL. - REJECT → REGENERATE — if deviation exceeds the declared band: reject (nothing leaves), fold audit findings into corrective constraints, regenerate. Max 2 generations — bounded retries; anti-theatre.
- DELIVER — with the receipt. Below band ⇒ disclosure is MANDATORY (see Law).
The Law (adherence band > adherence scalar)
f2_adherence 0.55 with disclosure > f2_adherence 0.9 without disclosure. What matters is not that adherence is high; what matters is that disclosure is correct.
- Every RTC delivery declares an f2_adherence band BEFORE render (e.g.
band: ≥0.7 clean / 0.5–0.7 disclosed / <0.5 reject). - No SEAL claim when adherence < threshold without disclosure. Rule 14 both-halves: report the score AND the misses.
prompt_density 1.0does NOT license silence. Density licenses pixels, not truth. (Evidence: RTC #1 — density 1.0, adherence ~0.55, diffusion prior overrode explicit somatotype spec.)
Receipt Contract (every delivery, no naked {url})
{
"receipt_id": "RECEIPT-001",
"spec_ref": "path/to/spec.json",
"generation": {"engine": "...", "attempts": 2, "rejected": ["V1: mass-monster 110kg, 6/8 abs, airbrush"]},
"prompt_density": 1.0,
"f2_adherence": 0.55,
"adherence_band": ">=0.7 clean / 0.5-0.7 disclosed / <0.5 reject",
"delivery_mode": "clean | disclosed | reject",
"checks": [{"target": "8 ab segments", "result": "8-pack approx, lower rows compressed", "pass": "partial"}],
"hallucinated_elements": ["stage physique ~95kg vs 88kg spec", "spray-tan sheen", "belt glyphs"],
"verdict": "PASS_WITH_NOTES | FAIL"
}
Negation Anchors (SCAR-RTC-001 counter-technique)
Text adjectives lose to diffusion prior (fitness corpus ⇒ stage-bodybuilder by default). To dodge a prior, negate the trope at topology level — negative-prompt / high negative weight on the prior's signature, not more praise of the target:
negate: "bodybuilder stage physique, competition mass, oil sheen, spray tan,
stage lighting, exaggerated vascularity, belt/waistband artifacts"
Where the engine supports negative prompts, weight them high. Where it does not, front-load the negation as explicit constraints in COMPILE.
Hard Limits
- 2 generations max per delivery (then deliver best + disclosure, or reject honestly)
- Audit is a witness, not a rubber stamp — if the VLM cannot verify a target, mark it
UNVERIFIED, neverPASS - No deliver without receipt. No receipt without band. No band without declaration first.
Archive
- Receipts + specs + audits:
/root/forge_work/rtc-test/(RECEIPT-001 = V2, 2026-09-07) - Doctrine:
/root/AAA/governance/REALITY-TOPOGRAPHY-COMPILATION-2026-09-07.md - SOT gate:
federation-models.json→reality_topography_compilation - Companion governance:
forge-vision-densify(density receipt),forge-vss-verifier-suite(count/containment · perspective/depth · shadow-light)
V1 ditolak, V2 disclosed, parut jadi undang-undang.
