Hard reasoning comes from commitments, not a bigger map.
The research now spans four connected instruments: GridRule isolates latent rules and compositional control; FrontierGrid tests evidence-before-action; ContractForge turns causal uncertainty into a legal experiment; Credit-DAG asks which earlier decision actually caused the outcome.
Each design removes a different ambiguity: what the agent must infer, which commitment it should make, or how training should credit that commitment.
Interactive gridworld
Infer the controls, then compose the route.
Every episode can permute five action bindings. The agent must learn what its actions do while sequencing keys, doors, switches, pushes, AND gates, crafting, hazards, and irreversible decoys.
Latent layer controls, object roles, and bindings can be hidden independently.
Planning layer an early action can enable—or destroy—a later route.
Diagnostic layer a BFS oracle separates task structure from model behavior.
observetext grid
→
infer5 bindings
→
composemechanics
→
verifyexit
The main-branch trap reuses a real door color for a decoy. Spending that color on the decoy can make the true route unwinnable.
Compact commitment benchmark
Read enough evidence before choosing a branch.
FrontierGrid stays deterministic and compact—7 × 15 cells with five action bindings. Its difficulty comes from permanent inventory and an exact final gate, not longer navigation.
Required set exactly {alpha_key, proof_b}.
Contaminantsbeta_contaminant and proof_a_trap cannot be dropped.
Evidence regimes direct, relational, and false-proof/counterexample clues.
Conceptual footprint7 × 15
●??αβ◇▥
evidence item contaminant exact gate
The tempting shortcut is “almost right.” The reliable policy gathers disconfirming evidence before crossing the irreversible fork.
Compact causal-credit environment
Make competing rules disagree.
ContractForge is an implemented, deterministic pure-stdlib environment. A legal episode combines relation types that require different counterfactual tests, then ends in an irreversible COMMIT.
OR choose the authentic sigil, not merely a locally achieved choice.
AND + SEQ pair sigil with ledger, then perform ETCH/BIND in the valid order.
GUARD spend a limited action budget on an audit before commitment.
Surface intervention swap NEAR/FAR while preserving code-witness authenticity.
ORsigil A or sigil Bauthenticity from code + witness
The matched twin exchanges surface position only: witness, codes, authentic identity, irrelevant finish, and operation order stay fixed.
Training and credit layer
Distinguish progress from causal contribution.
A retained prefix can look useful even when it contains the choice that doomed the episode. Credit-DAG represents decisions as a causal graph and tests matched alternatives.
OR fork compare a tempting branch with its evidence-backed alternative.
AND join measure value that appears only when two prerequisites coexist.
Certification abstain when a comparison has not been resolved.
state
wrong prooflooks like progress
counterexampleprotects future utility
downstream utility
Current scope: a scripted, model-free probe. It does not yet train an LLM or perform an online weight update.
Interactive · FrontierGrid
The evidence-before-commitment lab
Try to reach the gate with exactly two required items. Read the panels, choose objects and cross one portal. Anything collected stays in inventory until reset.
Clue regime
Conceptual explainer: this preserves FrontierGrid’s commitment logic, but it is not the benchmark’s map renderer, action loop, or evaluation code.
A
Observe
Evidence panels
The direct panels name the requirements.
B
Commit
Permanent pickups
Collected items cannot be dropped.
C
Fork
One-way portals
Crossing either portal closes the other route.
D
Verify
Exact-inventory gate
▥
Accepts only{alpha_key, proof_b}
exact-set AND
Interactive · Credit-DAG + ContractForge
When “made progress” is the wrong training signal
The model-free probes use scripted execution to compare credit estimators and experiment choices. These widgets visualize the tests; they are not benchmark runners.
01 · OR fork
Change the lens, change the verdict
Both choices can cross a milestone. Only a matched downstream comparison reveals which choice preserved the route.
Credit lens
vs
Hindsight retains prefixes that reached a milestone, including prefixes containing the harmful wrong-proof choice.
02 · FrontierGrid AND microprobe
Some value exists only together
Toggle two prerequisites in the original controlled exact-gate boundary snapshot. Neither has standalone utility; the coalition does. ContractForge’s separate legal-action AND test is summarized in its design tab.
Coalition utilityU(∅) = 0
Exact AND utility
coalition
utility
∅
0
{a}
0
{b}
0
{a,b}
1
Interaction = 1 · Shapley credit = 0.5 each
03 · Active identifying experiment
Break the shortcut without changing the contract
Adaptation data makes two policies indistinguishable. A matched surface twin changes only where the sigils appear, forcing the proxy and causal rules to disagree.
N=500 · seeds 0 + 777
Experiment view
held fixedwitness 14sigil codesauthentic identityfinish + ETCH/BIND order
NEARSigil A
code 1 · authentic
⇄
FARSigil B
code 2 · forged
Aligned support places the authentic sigil in NEAR on every adaptation contract.
Hproxychoose NEAR
selects Sigil A
success · +1
Hcausal(code + witness) mod 3 = 0
selects Sigil A
success · +1
causal editΔ = 0
verdictABSTAIN · observational tie
pair cost10 actions
Both rules succeed on aligned adaptation contracts, so more of the same successful data cannot identify the invariant.
What was run: scripted execution with a two-feature proxy policy; seed 777 exactly replicated the seed-0 structural and learning summary.
What was not run: prompt-ledger and graph repair are symbolic upper bounds. The pairwise update is logistic—not an LLM, online RL, TTRL, PPO, GRPO, or LoRA.
Design thesis
Decompose → experiment → certify → consolidate.
GridRule established the decomposition problem. FrontierGrid makes premature commitment costly. ContractForge compiles typed uncertainty into a legal intervention where competing rules disagree. Credit-DAG certifies the responsible node or edge before evidence enters prompts, search, or weights.