This model defines a bounded, falsifiable learning accelerator for MD-OS. Its purpose is not to relabel the runtime as AGI. Its purpose is to increase verified learning velocity:
learning velocity
= change in sealed-holdout success
/ independently verified source episodes
The numerator must come from previously unseen cases. The denominator must include every verified episode used for induction. Attempts, regressions, cost, and human intervention are reported separately so a higher score cannot be manufactured by spending more search at evaluation time.
The executable entrypoint is:
cortex agi accelerate --experiment-id <append_only_id>
The command runs one finite experiment. It is not a daemon and does not enable continuous autonomous execution.
A successful run supports only this claim:
Within one declared hypothesis family, MD-OS induced a reusable procedure from
independently verified development episodes and improved on sealed, distinct
holdout cases at a controlled attempt budget.
It does not establish:
open-domain AGI
cross-domain transfer
unbounded invention
continual autonomous self-improvement
parametric training of the host model
long-horizon autonomy
The report therefore carries explicit fields:
narrow_learning_transfer_supported
agi_achieved = false
agi_claim_supported = false
The accelerator implements a complementary two-speed learning system.
Each development run becomes an append-only episode only after an independent oracle verifies at least one candidate. The episode retains:
public learning examples
verified candidates
oracle-rejected hypotheses
provider and benchmark receipts
source split and provenance
Rejected candidates are recorded as prediction_errors. They are learning
signals, not operational failures, and therefore do not corrupt the runtime
failure index.
The hypothesis population is updated only when an observation eliminates at least one currently consistent hypothesis. Non-informative observations do not create a plasticity event.
For an update from H_before to H_after:
surprise_bits = log2(|H_before| / |H_after|)
This makes the update budget proportional to information gained rather than to raw episode volume.
The current bounded family has four binary constraints and therefore sixteen possible conjunctions. Inconsistent hypotheses are inhibited. Consolidation uses a winner-take-all code:
one active hypothesis / sixteen available hypotheses
sparse-code density = 0.0625
The winning unit decodes to the selected structural constraints. This is a computational design inspired by sparse competitive coding; it is not a claim that the implementation reproduces biological neurons.
Verified episodes are replayed in descending prediction-entropy order with a deterministic tie-break. Episodes that divide the surviving hypothesis set most evenly are consolidated first. Replay is limited to development evidence. Validation and holdout cases never enter the induction input.
The slow store receives a parameterized skill only when:
at least two independently verified development episodes exist
+ source cases are distinct
+ one hypothesis is uniquely identified
+ every required constraint has supporting and corrective evidence
The skill transfers structure, not repository-specific strings. On a new repository it infers the source path, variable, delimiter, valid prefix, failure return, and output shape from the bounded public snapshot and visible regression checks.
Consolidation does not imply promotion. A candidate skill is promotable only when it passes:
distinct validation case
+ sealed holdout cohort
+ positive before/after delta
+ equal single-attempt holdout budget
+ zero contamination
+ zero regressions
+ independent verifier readback
+ rollback contract
This gate prevents high plasticity from becoming uncontrolled skill growth.
The first accelerator experiment targets delimited boundary validators. The hypothesis language is:
exact_arity
prefix_match
payload_nonempty
payload_charset
The sixteen candidate hypotheses are the powerset of these constraints. Two development repositories provide complementary counterexamples:
development episode A identifies exact arity and prefix matching
development episode B identifies non-empty payload and allowed payload alphabet
Together they identify one complete grammar. The consolidated skill then infers target-specific parameters on one validation repository and two sealed holdout repositories with different functions, delimiters, prefixes, and return shapes.
The causal sequence is fixed:
1. Run the same provider without skills on one validation case.
2. Run the same provider without skills on two sealed holdout cases.
3. Explore five competing hypotheses on each of two development cases.
4. Write two independently verified episodes.
5. Eliminate inconsistent hypotheses and consolidate one sparse skill.
6. Run the skill once on the distinct validation case.
7. Run the skill once on each sealed holdout case.
8. Audit contamination and equal attempt budgets.
9. Evaluate, gate, and promote only if every closure edge passes.
The provider process receives only its descriptor, request, and implementation file through the Node permission model. It receives no oracle programs, expected outputs, case ground truth, holdout examples, or prior holdout results. The external oracle executes only after candidate generation in an isolated Git worktree.
Every successful report records:
before_success_rate
after_success_rate
absolute_delta
success_delta_per_verified_episode
information_gain_bits_per_episode
hypotheses_eliminated_per_episode
exploration_candidate_count
baseline_holdout_attempts
learned_holdout_attempts
regression_count
human_interventions
total_measured_cost
total_latency_ms
The primary acceptance condition is:
after_success_rate > before_success_rate
and baseline_holdout_attempts = learned_holdout_attempts
and contamination_detected = false
and regression_count = 0
The repository reference run is:
md-os/ops/agi/learning_experiments/neuromorphic_transfer_20260718_v2/report.json
Its expected controlled result is:
verified development episodes: 2
hypotheses: 16 -> 1
information gain: 4 bits
information gain per episode: 2 bits
sealed holdout success before learning: 0 / 2
sealed holdout success after learning: 2 / 2
absolute holdout delta: 1.0
success delta per verified episode: 0.5
holdout attempts before/after: 2 / 2
regressions: 0
human interventions: 0
contamination: false
These numbers are acceptance targets until the corresponding report exists. In a packaged or live workspace, the report is the source of truth.
The experiment writes:
md-os/ops/tasks/
md-os/ops/verifications/
md-os/ops/episodes/
md-os/ops/evals/
md-os/ops/skills/candidates/
md-os/ops/skills/promoted/
md-os/ops/agi/learning_experiments/<experiment_id>/
md-os/ops/agi/neuromorphic_learning_status.json
md-os/ops/agi/neuromorphic_learning_status.md
The report runtime class is defined by:
md-os/schemas/neuromorphic_learning_experiment.schema.json
Each experiment directory is append-only. Reusing an experiment identifier is a hard conflict.
The repository now includes a separate executable suite that closes the next five operational evidence edges in a controlled symbolic environment:
cortex agi prove \
--experiment-id agi_generality_reference_20260718_v3 \
--cycles 96 \
--sessions 6
The v3 suite adds:
cross-domain transfer with disjoint primitive identifiers
+ equal-budget baseline and irrelevant-sketch causal control
+ novel depth-four compositional program synthesis
+ an autonomously selected procedural curriculum
+ cumulative replay, interference detection, and rollback
+ persistent state across fresh process restarts
+ injected transient-fault recovery
+ an append-only hash-chained campaign ledger
+ a complete evidence integrity manifest
Its protocol is defined in:
md-os/kb/AGI_PREREQUISITE_EVIDENCE_MODEL.md
Canonical evidence is written under:
md-os/ops/agi/generality_experiments/<experiment_id>/
Passing the v2 accelerator and the v3 generality suite does not make the AGI claim true. The evidence remains internally authored, finite, and bounded to a symbolic program-synthesis substrate. A broad claim still requires:
independently authored sealed domains
+ external replication by other operators
+ materially open-world tasks and tools
+ repeated seeded cohorts with uncertainty intervals
+ substantially longer wall-clock deployment
+ resource, safety, and regression performance under changing environments
The package therefore records:
operational_agi_prerequisites_supported = true # only when all five v3 gates pass
agi_achieved = false
agi_claim_supported = false