# Layer 1 — Vision and thesis of Objective Guard

## Problem

Programming agents can function with great effectiveness as long as the relevant context for a task remains small and observable. As a repository grows organically, implicit dependencies, dispersed constraints, and insufficient feedback emerge. The agent can then:

- mislocate the problem;
- build on an incorrect hypothesis;
- progressively expand the change;
- weaken tests to make a solution pass;
- replace components whose historical behavior it does not understand;
- declare success without sufficient evidence.

The most avoidable cost is not merely introducing an error. It is discovering it several steps later, when new decisions already depend on it.

The observation of a "small-good, medium-bad, large-structured-better" curve remains a confounded hypothesis. Objective Guard does not presume it is universal. The product will gather evidence to measure when it appears and what factors explain it.

## Current thesis

The product thesis is more limited and falsifiable than the initial PRD:

> An agent can work with greater reliability and lower cost if it receives programmatic feedback at the first moment where reproducible evidence of a violation exists, and if statistical expansions of scope are observed and calibrated before becoming blocking decisions.

The associated scientific thesis proposes that performance depends on the relationship between effective complexity, necessary context, locality, contracts, and observability; not merely on raw size.

A useful mnemonic, not a directly measurable quantity, is:

\[
\rho = \frac{K}{C_{\mathrm{eff}}}
\]

Where \(K\) is the minimal founded context necessary for a task and \(C_{\mathrm{eff}}\) the agent's effective capacity to use context. Both are latent variables and will only be used through validated proxies.

Objective Guard does not need to prove this theory to deliver initial value: it can detect local facts such as broken syntax, disabled tests, or new dependencies from the first release.

## Principles

### P-01 — Reflex before control plane

The first level must react on edit, with zero configuration and delta-only output. More sophisticated components are added when evidence exists that they justify their cost.

### P-02 — Facts, inferences, and decisions are distinct objects

- Fact: an assertion was removed.
- Inference: the surface of the change is at the 99th percentile.
- Decision: a policy requires pausing and justifying.

The system must never present a statistical anomaly as a proof of incorrectness.

### P-03 — Blocking proportional to evidence

A reproducible violation can block. An uncalibrated signal is only recorded. A calibrated statistical signal can warn or require justification according to explicit policy.

### P-04 — Cost proportional to impact

The cost of validation must grow with the radius of impact and uncertainty, not with the total size of the repository.

### P-05 — Push and delta-only

New findings are delivered automatically in the agent's cycle. Known findings do not consume tokens repeatedly.

### P-06 — `unknown` is a valid result

Absence of evidence does not equate to correctness. Each criterion can be `proven`, `violated`, `unknown`, or `not_applicable`.

### P-07 — Fail without breaking the loop

An internal error in the guard must not block an edit. It must be logged and be observable without confusion with a `pass`.

### P-08 — Measure before governing statistically

Drift metrics begin in passive mode. Only after out-of-sample calibration and an intervention trial can they generate automatic alerts.

## Epistemic limits

Objective Guard cannot:

- formally prove the general correctness of a program;
- recover an intention that was never expressed nor observable;
- guarantee that a human contract correctly represents the objective;
- infer complete historical requirements from code;
- use model consensus as a substitute for independent evidence;
- turn incomplete tests into a perfect oracle;
- assert a probability of incorrectness from a percentile of anomaly.

The product can reduce the error space, make uncertainties visible, require evidence by criterion, and detect deviations before they accumulate cost.