Adaptive Survival Protocol Documentation

A complete reference for the Adaptive Survival Protocol: hypothesis, variables, causal pipeline, measurement, experimental design, statistical models, and falsification criteria. This page documents what the protocol intends to test, what it does not claim, and how its own failure can be made visible.

Overview

The Adaptive Survival Protocol is a measurement protocol for testing whether revision and action mechanisms explain system survival under model–reality divergence, independently of prediction accuracy.

The protocol separates, both conceptually and operationally, the survival outcome from the mechanisms that might explain it. Survival is not defined by revision or action; revision and action are candidate mechanisms whose contribution to survival is to be tested.

We do not ask the instrument to prove the theory; we ask it to make its own failure visible.

Central Hypothesis

The central hypothesis is stated in deliberately neutral terms:

Survival under model–reality divergence is not explained by the absence of divergence, but by the capacity to revise and act under it.

This formulation does not claim that what is measured is intelligence. It claims only that a specific class of mechanisms may explain survival better than prediction accuracy alone.

Formal Statement

Let:

• S: operational survival of the system under divergence
• E: model–reality divergence event
• PAcc: prediction accuracy
• O: opportunity structure (signal, authority, action)
• R^K: model revision capacity
• R^Ω: policy revision capacity
• A^cap: action capacity

H0:
After accounting for prediction accuracy (PAcc) and relevant
environmental factors (including O), R^K, R^Ω, and A^cap
add no independent explanatory power for survival under E.

H1:
R^K, R^Ω, and A^cap add independent explanatory power for
survival under E, beyond PAcc and relevant environmental factors.

What the Hypothesis Does Not Claim

• That R^K, R^Ω, A^cap are necessary for survival.
• That R^K, R^Ω, A^cap are sufficient for survival.
• That PAcc is irrelevant to survival.
• That the mechanisms are the same across system types.

Variables

The protocol defines seven primary variables. Each is defined independently of the others. No variable is defined in terms of the mechanisms it is supposed to be explained by.

Two further variables are used in the models:

Key Distinctions

Survival as an Operational Outcome

S_E(τ) = Pr( S^op(t+τ) = 1 | E_t )

where:
  E_t   : the divergence event at time t
  τ    : continuous elapsed time since E_t
  S^op  : the system is operational over [t+τ, t+τ+δ]

S^op is operationalized through four binary indicators:
  M1 : basic function continues
  M2 : irreversible collapse is avoided
  M3 : decision capacity persists
  M4 : capacity to receive new information persists

S^op(t) = 1   iff   Σ_j w_j · M_j(t)  ≥  θ_S

Causal Pipeline

The protocol specifies the following causal ordering:

E_t
  → Detection_t
  → C_t
  → ( R^K_t , R^Ω_t )
  → A^cap_t
  → Outcome_t
  → S_E(τ)

Each arrow represents a conditional transition, not a deterministic implication.

Probabilistic Form

P( Detection_t = 1 | E_t , O^sig_t )
P( C_t = c       | Detection_t , E_t )
P( R^K_t = 1     | C_t , O^sig_t , Detection_t )
P( R^Ω_t = 1 | C_t , O^auth_t , Detection_t )
P( A^cap_t = 1   | R^K_t , R^Ω_t , O^act_t )
P( Outcome_t = 1 | A^cap_t , environment )
P( S_{t:t+τ}  | Outcome_t , history )

Conditioning on Opportunity

Each stage is conditioned on a specific component of O:

O^sig_t   conditions  Detection_t  and  R^K_t
O^auth_t  conditions  R^Ω_t
O^act_t   conditions  A^cap_t

O^sig_t  = 0  ⇒  Detection_t = 0
O^auth_t = 0  ⇒  R^Ω_t = 0
O^act_t  = 0  ⇒  A^cap_t = 0

Feedback Dynamics

The pipeline is not one-shot. The system's prior actions reshape the conditions under which it will later detect, revise, and act. Three feedback loops are distinguished:

1. Opportunity loop. Prior actions reshape future access to signal, authority, and execution.
2. Environment loop. Prior actions reshape the environment the system will face, including its volatility.
3. Capacity loop. Prior revisions and actions reshape future revision and action capacity.

Intervention Points

Measurement

Three Levels

Each variable may be measured at one of three levels:

1. External. Measured by an observer, independently of the system's own reporting or behaviour.
2. Internal. Measured through the system's own reporting or internal state.
3. Behavioral. Inferred from observable behaviour.

The protocol prefers external measurement, accepts internal measurement as secondary, and uses behavioral inference only when the other two are unavailable.

Time and Stability

τ = α · T_int ,   α ∈ ℝ_{>0}

where T_int is a pre-specified characteristic time scale of the system.
The estimator for T_int is chosen before data collection from:

  1. response-time constant
  2. natural update-cycle duration
  3. characteristic stabilization time

The stability window δ is expressed as a fraction of T_int:
  δ = 0.1 · T_int   (default, operational not theoretical)

Thresholds

Measurement Invariance

Invariance is tested at three levels:

• Configural. Same structural meaning across systems.
• Metric. Same units of measurement across systems.
• Scalar. Same zero point across systems.

Each level is a precondition for the next. Without at least configural invariance, cross-system comparison is not valid.

Experimental Design

Design Principle

Test environments are designed to separate variables, not to correlate them. Target separations:

R   ⊥ A   | controls
R^K ⊥ R^Ω | controls
R   ⊥ O   | design
σ_E ⊥ O   | design

Environment Types

1. Synthetic. Fully controlled. Preferred for causal identification.
2. Semi-synthetic. Partially controlled. Accepted when synthetic environments are unrepresentative.
3. Natural. Fully observed, not controlled. Secondary validation only.

Required Environments

Failure Matrix

Failure is not binary. It is classified by the first layer that fails.

Statistical Models

The protocol specifies three nested models:

• M0. Baseline. Strongest model without R or A.
• M1. Full. Adds R^K, R^Ω, A^cap to M0.
• M2. Interaction. Adds R × σ_E, A × σ_E to M1.

Model M0

logit( P(S_i = 1) )
    = β0
    + β1 · PAcc_i
    + β2 · O^sig_i
    + β3 · O^auth_i
    + β4 · O^act_i
    + β5 · sigma_E_i
    + u_system[i]
    + v_env[i]

Model M1

logit( P(S_i = 1) )
    = M0
    + β6 · R^K_i
    + β7 · R^Ω_i
    + β8 · A^cap_i

The coefficients of interest are β6, β7, β8.

Model M2

logit( P(S_i = 1) )
    = M1
    + γ1 · ( R^K_i   × sigma_E_i )
    + γ2 · ( R^Ω_i   × sigma_E_i )
    + γ3 · ( A^cap_i × sigma_E_i )

The coefficients of interest are γ1, γ2, γ3.

Model Comparison

M1 is compared with M0 by four criteria:

1. Out-of-sample predictive performance (primary). Cross-validation, by a pre-specified margin.
2. Information criteria (secondary). AIC and BIC.
3. Likelihood ratio test (secondary). 3 df.
4. Coefficient significance (secondary). After multiple-comparison correction.

Falsification

Falsification operates at four levels:

See Falsification page for the full criteria, failure matrix, counterexample types, and open questions.

Limitations

The protocol has structural limitations, not defects. Limitations are constraints on valid use.

Types of Limitations

Reproducibility

Frozen and Living Layers

The protocol distinguishes two states:

• Frozen. The file is a reference. Any change is recorded in protocol-core/CHANGELOG.md as amendment, correction, or extension.
• Living. The file is a working record. It may change freely. Living files are: CHANGELOG.md, notes/open_questions.md, notes/rejected_paths.md, notes/interpretive_hypotheses.md.

Executable Audit

The protocol includes an executable audit that verifies its own structural integrity. The audit runs on every push and merge request through continuous integration.

Run the audit

# From the repository root
bash tests/audit.sh
bash tests/run_all.sh

What is not yet reproducible

• Empirical results (none exist).
• Simulation results (no simulation has been run).
• Data (none have been collected).
• Fitted models (no data to fit).

Repository Structure

adaptive-survival-protocol/
│
├── README.md
├── LICENSE
├── CITATION.cff
├── AUTHORS.md
├── CONTRIBUTING.md
├── VERSION
├── .gitignore
├── .gitattributes
├── .gitlab-ci.yml
│
├── tests/
│   ├── README.md
│   ├── common.R
│   ├── audit.sh
│   ├── test_M0_baseline.R
│   ├── test_M1_full.R
│   ├── test_M2_interactions.R
│   └── run_all.sh
│
├── PAPER/
│   ├── README.md
│   ├── outline.md
│   ├── paper.md
│   └── sections/
│
├── PUBLICATION/
│   ├── README.md
│   └── Netlify/
│       ├── netlify.toml
│       └── public/
│           ├── index.html
│           ├── dashboard.html
│           ├── results.html
│           ├── documentation.html
│           ├── assets/style.css
│           └── data/
│               ├── protocol_status.json
│               └── falsification_status.json
│
├── DOCS/
│   ├── README.md
│   ├── glossary.md
│   └── faq.md
│
└── protocol-core/
    ├── README.md
    ├── CHANGELOG.md
    ├── foundations/      (2 files)
    ├── variables/        (7 files)
    ├── pipeline/         (3 files)
    ├── measurement/      (3 files)
    ├── experiments/      (4 files)
    ├── models/           (3 files)
    ├── falsification/    (2 files)
    └── notes/            (3 files)

Frozen vs Living

• Frozen files: 25
• Living files: 4
• Executable checks: 4 scripts, 200+ individual checks

Author & Citation

Citation (BibTeX)

@misc{baladi2026adaptive,
  author       = {Baladi, Samir},
  title        = {Adaptive Survival Protocol:
                  A Measurement Protocol for Survival
                  Under Model--Reality Divergence},
  year         = {2026},
  version      = {0.1.0-protocol},
  howpublished = {Protocol manuscript},
  note         = {Pre-empirical stage. No claims made.},
  doi          = {10.17605/OSF.IO/PT2K6}
}

Citation (APA)

Baladi, S. (2026). Adaptive Survival Protocol: A Measurement
Protocol for Survival Under Model-Reality Divergence (Version
0.1.0-protocol) [Protocol manuscript].
https://doi.org/10.17605/OSF.IO/PT2K6

Status