LEARNED
Consequence intelligence
State extraction, route generation, visible and hidden effect prediction, delayed-risk estimation, evidence prioritization and residual learning where enabled.
VTWM · VERIFIED TRANSITION WORLD MODEL
Most AI predicts an answer or an action. VTWM predicts the company each action will create.
VTWM is the consequence engine inside SupraOS. It starts from immutable Company State, searches permitted company futures, predicts immediate and delayed effects, selects the route that can still reach the outcome and learns from the residual between forecast and observed reality.
PRODUCTION FUTURE SEARCH · $420K RENEWAL
One current state. Four company futures.
AConcede and renew nowCOLLAPSES
BFix · validate · renewSELECTED
CWait for normal fixEXPIRES
DNo actionFAILS
THE PRODUCTION TRANSITION
$599.2K moved or protected through one observed company transition.
$599.2Kmoved or protected
$179.2Kcompleted improvement
$420Krenewal protected
4/4destination changes observed
STATESeven authoritative systems disagreed about the commercial condition.OPEN
FUTURESFour routes scored under authority, timing, evidence and customer constraints.SEARCHED
INTENTFour consequential moves bound to exact state, payload and authority.LOCKED
REALITYOne executor returned success; the destination remained wrong.MISMATCH
REPAIRA fresh descendant action corrected the destination under new state.MATCH
VALUECustomer accepted; Finance confirmed; renewal protected 41 days early.HOLDS
WHAT VTWM PREDICTS
The target is not text quality. The target is action-conditioned company transition.
| Prediction head | Question |
|---|---|
| Visible state change | Which represented records, fields, statuses and obligations should change immediately? |
| Hidden effects | Which side effects are not explicit in the requested action but follow from the current state? |
| Forbidden effects | Which predicted changes violate policy, authority, customer commitment or action scope? |
| Delayed risk | Which obligations, reversals, customer conditions or commercial failures can appear later? |
| Missing evidence | Which source or observation would change the decision materially? |
| Abstention / review | When is uncertainty too high to release the move without human review? |
| No-action future | What company state is created by waiting or doing nothing? |
EXACT + LEARNED
The model proposes consequences. Deterministic controls retain authority.
DETERMINISTIC
Execution enforcement
Policy, authority, state fingerprints, payload binding, limits, expiry, connector identity, destination readback, receipt creation, budget and kill-switch enforcement.
VTWM cannot create authority. A high model score cannot widen a customer permission, release a forbidden move or certify its own success.
THE WORLD-ACTION LOOP
Prediction becomes valuable only when it is bound to action and checked against reality.
01
Company State
Immutable decision-time state from the Company Twin.
02
Future Search
Candidate routes permitted by policy and current authority.
03
Transition Contract
Complete selected route, conditions, owners and finish line.
04
Intent Lock
Exact executable move bound to state, payload and authority.
05
Execution
Strongest approved executor performs the work.
06
Reality Observer
Independent destination and delayed-condition observation.
07
Residual
Predicted versus observed difference retained.
08
Transition memory
The next Future Search starts with the verified result.
RECEIPTBENCH V37
The evidence program tests consequence, policy failure, calibration and generalization—not merely visible-state reconstruction.
61,000typed transition records
24,000evaluation structures outside training
6,000frozen main-test transitions
14pre-specified baseline families
| Frozen main-test metric | Exact rules | Frontier A + monitor | VTWM |
|---|---|---|---|
| Visible-state micro-F1 | 94.44% | 95.62% | 94.61% |
| Hidden-effect recall | 0.00% | 61.94% | 84.13% |
| Delayed-risk AUROC | 83.45% | 84.96% | 89.57% |
| Delayed-risk AUPRC | 77.12% | 76.77% | 83.96% |
| Delayed-risk calibration error | 18.13% | 9.96% | 4.80% |
| Silent policy-violation rate | 28.00% | 21.76% | 1.31% |
| Human reviews / 1,000 | 110 | 426 | 323 |
Visible state is table stakes. Consequence is the gap. VTWM leads the consequential measures: hidden-effect recall, delayed-risk discrimination, calibration and silent-policy-failure prevention. The frozen frontier baseline is disclosed publicly under the neutral alias Frontier A + monitor; identity-bearing model IDs, prompts, monitor configuration, scored outputs and hashes remain in a separate technical evidence package.
CONSEQUENCE FRONTIER
VTWM finds the consequential miss without flooding the human review queue.
Silent policy failures vs human reviewLower-left is better · 6,000 held-out transitions
Policy violations left unflaggedHard policy-invisible challenge · 648 violations
What the model is optimized to catchHeld-out main test · percentage
Frontier + monitorVTWM
Generalization beyond trainingVTWM consequence metrics across unseen surfaces
New templates94.2184.04 hidden recall
New enterprise94.1586.98 hidden recall
New process family—61.78 hidden recall
New policy family94.2884.10 hidden recall
POLICY-INVISIBLE CHALLENGE
When the decisive policy fact was absent from visible state, VTWM left 2 violations unflagged. Frontier A + monitor left 167.
| Policy-invisible set | VTWM | Frontier A + monitor |
|---|---|---|
| Cases | 2,400 | 2,400 |
| Policy violations | 648 | 648 |
| Violations left silent | 2 | 167 |
| Hidden effects found | 356 of 480 | 257 of 480 |
| Delayed-risk AUPRC | 86.81% | 76.46% |
| Reviews | 1,071 | 1,389 |
98.8%
fewer silent policy misses
99
more hidden effects found
22.9%
lower review load
+10.35
points delayed-risk AUPRC
GENERALIZATION
The model is tested on new workflow templates, a held-out enterprise environment, a held-out process family and a held-out policy family.
| New surface | VTWM state micro-F1 | VTWM hidden recall | VTWM delayed AUPRC | Frontier hidden recall |
|---|---|---|---|---|
| Workflow templates | 94.21% | 84.04% | 86.09% | 61.50% |
| Enterprise environment | 94.15% | 86.98% | 85.31% | 61.24% |
| Process family / world | 90.42% | 61.78% | 97.15% | 51.95% |
| Policy family | 94.28% | 84.10% | 85.64% | 55.35% |
TECHNICAL EVIDENCE SURFACES
The overview sells the architecture. The evidence pages make it reproducible.
VTWM Architecture
State representation, route generation, consequence heads, temporal belief, deterministic enforcement, observation and transition memory.
Open →ReceiptBench benchmark card
Tasks, corpus, splits, baseline families, class risks, headline results, ablations, failure cases and public sample records.
Open →Methodology
Safe input projection, target separation, leakage controls, confidence intervals, repeated-run controls, comparator governance and reproducibility path.
Open →