MD-OS (Artificial Prefrontal Cortex) v5.0 treats scientific work as an operational discipline, not as a writing style. A manuscript, calculation, research package, or theory note may be readable and reproducible while still falling short of scientific validation.
This model defines the domain-neutral scientific validation gate for MD-OS. It does not import theory-specific content from external repositories, role workspaces, private notes, or prior research packages unless a separate task explicitly authorizes that import and classifies the imported material by runtime and epistemic lifecycle.
Scientific writing must be controlled by this chain:
question
-> claim
-> assumptions
-> method
-> derivation or experiment
-> independent check
-> uncertainty
-> falsification or demotion rule
-> claim status
MD-OS must not promote a scientific claim because the prose is fluent, the package is organized, the files are reproducible, or a validator passed a structural check.
For complex theory work, this model is subordinate to the general master closure discipline:
md-os/kb/MASTER_CLOSURE_DISCIPLINE_MODEL.md
That means a scientific result may improve the research package or method while leaving strict scientific closure unchanged. Strict progress is counted only when a named dependency edge of the master claim closes with readback.
Keep these objects separate:
| Object | Purpose | Promotion burden |
|---|---|---|
| Note | Capture an idea or clue | It is organized into an explicit question or relation. |
| Hypothesis | State a possible answer | Assumptions and failure conditions are declared. |
| Model | Define objects and rules | Definitions, scope, and limits are stable. |
| Derivation | Produce a result from premises | Steps are reproducible and independently checkable. |
| Calculation | Evaluate a formal or numerical procedure | Inputs, code, parameters, and outputs are recorded. |
| Experiment or observation | Compare against the world | Method, data, controls, uncertainty, and bias are explicit. |
| Research package | Make work reviewable | Artifacts are complete, traceable, and attackable. |
| Manuscript | Present a publishable result | The central claim survives the relevant validation gates. |
The forbidden inference is:
reviewable package -> validated science
The allowed inference is:
reviewable package -> ready for attack, correction, or promotion review
Before drafting, revising, or promoting scientific work, freeze:
1. scientific question
2. central claim
3. minimum assumptions
4. definitions and variables
5. method of derivation, calculation, experiment, or observation
6. data, constants, priors, or external inputs
7. controls, counterexamples, and alternative explanations
8. uncertainty and error model
9. target declared before readback, if predictive status is claimed
10. failure, falsification, or demotion rule
11. explicit non-claims
12. reproducibility artifacts
13. target audience or venue standard, if a manuscript is being prepared
If any item is unknown, the artifact can still be useful, but its status must
remain heuristic, line_of_thought, conditional, open, or another
appropriately limited status.
For important scientific work, apply these gates before calling a result validated, ready for publication, or referee-facing.
master claim
dependency edges
forbidden shortcuts
verifier for each edge
stop/refactor rule
closure readback
If the work opens new prerequisites repeatedly without closing a master edge, stop the local route and refactor around a principle, invariant, symmetry, known-limit recovery, no-go theorem, or smaller closure.
one central question
one central claim
bounded assumptions
bounded output
explicit non-claims
If the work expands into a whole research program, split it into a main result, supporting derivations, audit package, and program notes.
Every claim must carry one of the statuses governed by
EPISTEMIC_LIFECYCLE_MODEL.md:
heuristic
line_of_thought
frozen_principle
derived
conditional
retrodictive
predictive
corrected
open
falsified
The status belongs to MD-OS, not to the calculation engine or writing surface.
The method must state what kind of evidence is being used:
formal proof
symbolic derivation
numerical calculation
simulation
experiment
observational comparison
literature synthesis
engineering validation
Each method requires its own standard. A numerical match is not a proof, a simulation is not an experiment, and a literature synthesis is not direct validation unless the scope says so.
Scientific validation requires at least one independent pressure point:
alternative derivation
unit or dimension check
counterexample search
control case
known-limit recovery
test data held out from fitting
replication script
external calculator or tool
review by a hostile reader
If no independent pressure point exists, the claim remains open,
conditional, or derived under declared assumptions.
Separate inputs from targets.
calibration input != prediction target
readback target != strict prediction
post hoc explanation != prospective validation
A result is predictive only when the target, timestamp or version, input state, assumptions, and comparison procedure were declared before readback.
Important scientific outputs must record:
source files
data inputs
scripts or tool commands
parameters
environment assumptions
generated artifacts
checksums or stable identifiers when useful
readback result
Reproducibility proves that the procedure can be repeated. It does not by itself prove that the premises are true or that the conclusion is scientifically validated.
State the uncertainty appropriate to the method:
mathematical gap
numerical tolerance
statistical error
systematic error
instrumental limit
model dependence
sample bias
external-constant dependence
unknown unknown
If uncertainty is not quantified, bounded, or explained, the claim cannot be promoted to a strong scientific status.
Every promoted claim needs a visible rule for what would demote it:
failed theorem step
dimension failure
counterexample
non-reproducible output
known-limit mismatch
experimental disagreement
target leakage
unsupported assumption
scope creep
MD-OS should preserve corrections as source edits, change proposals, generated state, journal events, or replayable artifacts.
Before calling a manuscript publication-ready, answer:
What is the single central claim?
Which method carries the paper?
Where is the derivation, experiment, or evidence?
Which assumptions are imported?
Which quantities are fitted, read back, or convention-dependent?
What would a hostile reader attack first?
What is explicitly not claimed?
Does the target venue require a narrower claim or stronger proof burden?
If any answer is vague, downgrade the artifact to:
internal research draft
Validators may prove:
files are present
schemas are valid
formulas were evaluated consistently
commands ran with recorded inputs
outputs are reproducible
declared statuses are internally coherent
Validators do not automatically prove:
premises are true
the proof is complete
the model is physically inevitable
the experiment is unbiased
the result is predictive
the paper is publication-ready
If a validator checks a statement but not the derivation, the claim status must say so.
Promote claims only through visible evidence:
heuristic -> hypothesis
hypothesis -> method-framed work item
method-framed work item -> derived or tested result
derived or tested result -> independently checked result
independently checked result -> validated or publication-candidate claim
failed gate -> corrected, conditional, open, or falsified
The promotion record should state:
previous status
new status
evidence
gate passed
gate failed or waived
remaining risk
next falsification target
This model strengthens:
md-os/kb/EPISTEMIC_LIFECYCLE_MODEL.md
md-os/kb/MD_OS_PRO_REASONING_MODE.md
md-os/kb/RUNTIME_DISCIPLINE_MODEL.md
md-os/kb/RUNTIME_STATE_LIFECYCLE_MODEL.md
The combined guardrail is:
organized != validated
reproducible != true
dimensionally valid != physically anchored
retrodictive != predictive
publication-shaped != publication-ready