+Observe
Capture proposed actions and context through one engine-neutral protocol.
-
- Open-source compute governance for AI agents
-Fund only the next action worth taking.
-MARGINAL helps AI agents decide whether the next model call, tool call, search, retry, review or sub-agent is worth its token cost.
+Compute governance, measured
+The next action should add evidence, not just activity.
+MARGINAL evaluates whether another model call, tool call, retry, verification or reviewer is likely to add enough value to justify its cost — and accounts for the cost of making that decision.
-
-
- Local first
- Provider neutral
- Zero mandatory dependencies
- Apache-2.0 +
- Open source
- Local first
- Provider neutral
- Zero mandatory dependencies
-
Agent proposes action
↓
- Estimate value versus cost
↓
- ALLOWDENYSHADOW
- ↓
Settle actual usage + outcome
+
+
- Illustrative traceNot a benchmark
+ 01
+ Read README.mdnew information acquired
RUN02
+ Verify README.mdnew evidence acquired
RUN03
+ Verify README.mdsame state · no new evidence
DISCOUNT04
+ Verify README.mddiminishing return threshold reached
STOP?The question mark matters: enforcement is earned by evidence, not assumed by the demo.
-
-
-
The missing layer
-Budgets say what an agent can afford. MARGINAL asks what deserves funding.
-Agent runtimes often execute work because a model requested it or a hard limit has not been reached. MARGINAL compares expected improvement with tokens, cost, latency, risk, remaining budget and verification needs before compute is committed.
+
+
+
+ The actual thesis
+Not “models waste tokens.” Diminishing marginal value.
+
+
Coding agents can spend compute on actions whose incremental value becomes unclear. A repeated action is not automatically waste: another test may be exactly what a risky patch needs.
+MARGINAL looks for a stronger pattern: same semantic action + unchanged observable state + no new evidence. That is provider-neutral and remains meaningful even when models improve.
+How it works
Transactional governance, not token counting
- 01
- 02
- 03
- 04
- 05
+
+
-
Evaluate
Estimate marginal gain, uncertainty and full economic cost.
Reserve
Atomically protect budget and verification capacity before execution.
Settle
Record measured usage, failures, overruns and verified outcomes.
Learn
Replay versioned policies and improve calibration without overstating causality.
+
+ Proof before claims
+MARGINAL must earn its own compute.
+Matched OFF/ON evaluation is the product test. Saving agent tokens is insufficient if governance overhead, quality loss or false stops erase the benefit.
+
+ 01
+ 02
+ 03
+ 04
+ 05
+ 06
+
+
Verified quality
Same verifier, predefined non-inferiority margin, regressions and recoveries exposed.
Gross savings
Agent workload change before counting MARGINAL's own cost.
Net savings
Effective tokens, USD and latency after the governance tax. This is the claim surface.
Repeated calls
Shows whether behavior changed or the system simply moved cost somewhere else.
False stops
Explicitly reviewed deny recommendations that would have blocked helpful work.
Uncertainty
Repeated runs and bootstrap intervals separate a stable effect from one lucky trajectory.
-
-
Why it is different
Built for accountable autonomy
-
-
+
+
-
Shadow before enforcement
Observe recommendations without changing agent behavior, then enforce only after representative validation.
Real accounting lifecycle
Reserve, execute, settle, abort and report overruns across hierarchical treasuries.
Protected verification
Keep capacity available for tests and evidence instead of optimizing an agent into premature confidence.
Versioned evidence
Decision Ledger records correlate policy, estimator, cost, outcome and runtime identity for audit and replay.
Universal protocol
Thin adapters connect coding agents while the economic policy remains in one core.
Measured claims
Token reduction counts only when verified outcomes remain within a predefined quality constraint.
+
+
+ Graceful irrelevance
+What if GPT-5.7 — or any future model — is already efficient?
+
+
Then MARGINAL should get out of the way. The evaluator now distinguishes gross from net savings and can classify a configuration as pass_through when the governor does not demonstrate enough net value.
$ marginal public-eval baseline.jsonl marginal.jsonlintervention.status: pass_throughNo positive net intervention value demonstrated.
+ A negative or neutral benchmark is a valid result. The project should not manufacture a reason to intervene simply to justify its own existence.
The learning loop
From static policy to evidence-driven allocation
-MARGINAL's defensible advantage is not a single ROI formula. It is the closed loop between observed decisions, measured outcomes, versioned estimators, policy replay and calibration.
-Task success is kept separate from action-level causal attribution. Recorded correlation is not presented as causal proof.
+New control
State-aware diminishing returns, without a GPT-specific patch.
+
-
+
+ Same semantic keye.g. verify the same artifact for the same purpose
+ Same state hashthe underlying workspace has not changed
+ No new evidencethe previous pass did not change the evidence state
+ Repeat count risesexpected gain decays until a configured stop threshold
+
+
The detector is opt-in. Missing state fails open. Changed state or new evidence resets the pressure. That keeps the mechanism conservative while real agent telemetry is collected.
+Exact duplicate prevention still exists. Diminishing-return control adds a semantic layer for allowed retries and verification loops where fingerprints can legitimately differ.
+ Read the evidence model → +- Observe decisions
- Measure actual cost and outcomes
- Estimate marginal value with uncertainty
- Allocate budget
- Measure calibration and regret
- Improve the policy
-
-
Privacy by design
Operational evidence stays local. Shareable data is transformed deliberately.
-Prompt-free telemetry can still leak through identifiers, names and metadata. MARGINAL classifies fields and provides explicit export profiles.
-
- LOCAL_FULLComplete operational ledger for trusted local storage.
- SAFE_TELEMETRYFree text removed and identifiers pseudonymized with a local key.
- AGGREGATE_EXPORTGeneralized groups without identifiers; small groups suppressed by default.
+
-
-
+
+
Community pressure test
Feedback changes the product only when the reasoning survives review.
+ Accepted
+ Accepted
+ Partially accepted
+ Rejected
Show OFF vs ON benchmarks
Matched runs are required for performance claims. The 10-task canary remains engineering validation, not marketing evidence.
Future models may reduce the benefit
Converted into Graceful Irrelevance: MARGINAL must demonstrate positive net value for the workload in front of it.
“Less slop” on the website
The technical concepts stay, but the landing page now starts with a concrete failure mode and proof standard before architecture theory.
Providers intentionally preserve waste
There is no evidence needed or offered for that claim. MARGINAL is justified by user-controlled economics and observability, not provider motive speculation.
One product, multiple engines
Universal foundation first. Thin adapters next.
- Available
- Next
- Planned
- Planned
- Planned
+
+
+
-
Core runtime
Policy, treasury, ledger, privacy, replay and universal protocol.
Codex
Reference adapter and matched measured benchmark.
OpenCode
Open-source adapter and experimentation environment.
Claude Code
Hook-based integration using shared protocol contracts.
GitHub Copilot
Integration where official control surfaces permit enforcement.
Benchmark discipline
One benchmark is a surface, not ground truth.
+
+
-
The requested SWE-bench Pro comparison can be useful because it is recognizable and gives the community a common reference. It should still be versioned, audited and accompanied by task-quality notes.
MARGINAL-specific evaluation should also target the behavior the product claims to change: redundant same-state actions, repeated verification, overhead, false stops and verified outcomes.
Same modelSame promptSame toolsSame limitsSame task orderSame verifierRaw paired JSONLPreregistered gates
-
-
Evidence, not hype
Optimization is successful only when quality is preserved.
The bundled Killer Demo is deterministic and uses declared action costs. It explains the mechanism; it is not provider telemetry or a universal savings claim.
-Public evaluation requires matched model, prompt, tools, limits and verifier, then reports tokens per verified successful task, quality delta and uncertainty.
- Read the benchmark protocol → +
+
+
Next milestone
Codex Reference Integration
The core is being prepared now; the adapter and measured Codex runs remain the next implementation milestone.
+
+
v0.2Learning Loop Foundationuniversal protocol · ledger · privacy · replay
+ HardeningNet-value evidence layergovernance tax · false stops · diminishing returns
+ v0.3Codex integrationone-command target · telemetry · canary · public matched benchmark
-
Build economically disciplined agents
-Observe first. Measure honestly. Enforce what the evidence supports.
+Build the evidence first
+