Still unproven: learned state must control held-out behavior after the lesson is deleted.
Current scientific verdictChanged dynamics is not yet demonstrated.
Research 20 has four runtime-verified INT8 developmental checkpoints. The fifth seed retry was running at the live check; calibration and meta-test remain blocked.
Program snapshot
What works, what does not, and what is still unknown
“Works” names the narrow property actually measured. It never promotes a component result, infrastructure success, or retrieval win into proof of the thesis.
Exact-fact cvec steering, the registered KV recall claim, the two-model mid-layer advantage, hand-authored compounding, and the current latent articulation interface.
What was misleading
Saturated scores, leakage-era 1.00 claims, the 13.5x drift headline, zero-by-construction baselines, and seed-regenerated “compression.”
What remains unknown
Whether a meta-trained fixed-shape cortex can learn an unseen rule online and causally control a frozen language organ after trace deletion.
Nine analytical field notes
Read from thesis to attempt history
The prose explains causality and direction changes. The maps below provide complete numbered coverage; the evidence ledger preserves every dated classifier row.
Open an entry to see why the research existed, what was tried, what happened, the diagnosed cause, and how that evidence changed the roadmap.
01Correction-to-Competence Benchmark v2Can a behavior-only instrument separate architectures that the old scorecard tied at 1.0?Tested null
Why this existed
Can a behavior-only instrument separate architectures that the old scorecard tied at 1.0?
What was tried
Separated exact recall, domain recall, interference, retention, and behavior-per-byte; ran the registered campaign condition.
What happened
Five metrics de-saturated, but mock behavior delta and discrimination both remained 0.0. Exact recall was 0 while domain recall was 1.0.
Why
The system could shift broad semantic posture without acquiring the exact held-out competence the benchmark demanded.
What it changed
The benchmark exposed the distinction it was built to expose. De-saturation itself is not success, and a new v2.2 real-driver baseline is still pending.
02Reserved KV-slot fact injectionCan a fixed KV write force an arbitrary new fact that a residual control vector cannot?Refuted as specified
Why this existed
Can a fixed KV write force an arbitrary new fact that a residual control vector cannot?
What was tried
Compared registered KV-slot injection with residual steering and a direct logit-bias positive control.
What happened
KV rank-1 recall was 0; logit bias reached all three target tokens. The HF successor later reached only 1/3 absolute recall.
Why
A KV splice can reproduce prefix behavior, but that does not guarantee robust token binding or learning. Position and content routing remained the bottleneck.
What it changed
KV content stays useful as a prefix-equivalent delivery mechanism, not evidence of changed dynamics.
03Layer-L hidden extractionDo mid-layer residuals provide a better learning signal than the final-layer mean?Refuted, with one-model closure
Why this existed
Do mid-layer residuals provide a better learning signal than the final-layer mean?
What was tried
Ran the registered layer comparison on Qwen and LFM2.5, then repaired the real-driver transport path and repeated one LFM2.5 closure run.
What happened
S1.4 gaps were -0.083 and +0.058, both below +0.10. The later single-model closure reached +0.109254.
Why
The two-architecture test did not support the assumed mid-layer advantage. The closure proves the repaired path can reproduce one crossing, not that the broader refutation was wrong.
What it changed
Goal 2’s mid-layer assumption remains retired; the closure is reproducibility evidence only.
05Closing the metabolism loopDo repeated corrections compound into cold-state change that causes held-out behavior?Tested null
Why this existed
Do repeated corrections compound into cold-state change that causes held-out behavior?
What was tried
Ran perceive, update, replay, consolidate, and answer as one loop; measured state trajectory and behavioral uptake.
What happened
Four consolidations and a 0.1755 cold-state slope produced metabolism drift delta 0.0 and uptake 0.0.
Why
Internal movement was not aligned with answer behavior. In the successor loop, prefix content was evicted by the fixed token budget and cvec steering corrupted generation.
What it changed
A running update pipeline is not a learning result. The hand-authored loop was replaced by learned update and articulation rules.
06Bounded-growth consolidationCan persistent state remain bounded as experiences accumulate?Engineering positive, thesis caveat
Why this existed
Can persistent state remain bounded as experiences accumulate?
What was tried
Measured persistent footprint across five campaign seeds after an earlier seed-regenerable autoencoder design.
What happened
The later campaign footprint ratio was 0.002079 with zero observed variance and a spread within 20 bytes.
Why
Fixed-shape persistence can be made deterministic. But the earlier A0b design hit its byte target by regenerating a random matrix from a seed, so learned updates could not survive.
What it changed
Storage boundedness is established as an engineering property; compression of learned, behavior-bearing experience is not.
07Conversation world modelCan a learned predictor identify correction and acceptance dynamics better than lexical heuristics?Positive plus null, older claim superseded
Why this existed
Can a learned predictor identify correction and acceptance dynamics better than lexical heuristics?
What was tried
Trained a marker-free predictive path and compared the critic against registered baselines.
What happened
The July campaign recorded marker-free uptake +1.0 but critic AUC improvement 0.0. The earlier June +1.0 headline was separately superseded because its lexical baseline was zero by construction.
Why
A predictor can help the uptake surface, while the stronger critic-improvement claim received no support. Baseline design determines whether a gap is meaningful.
What it changed
Keep the predictive mechanism as a bounded result; do not revive the retired headline or claim a successful world-model critic.
14Organ ablation matrixWhich answer-time organs causally improve held-out behavior?Triage complete; one metricless run
Why this existed
Which answer-time organs causally improve held-out behavior?
What was tried
Combined subtractive GGUF ablations with additive HF tests and a later three-seed M2b run.
What happened
Scope-slot reranking improved S0 +0.667 and S4 +0.250; answer-time hippocampus was exactly 0; DSI was -0.060 in the full stack. M2b exited 0 after about 11,787 seconds but emitted no metric.
Why
Only the retrieval component produced repeatable behavior. The M2b process success did not produce scientific evidence because the scorer emitted nothing.
What it changed
Keep the reranker explicitly as retrieval; archive the other answer-time organs. Treat M2b as metricless, not as a zero effect.
18Consolidation as distillationCan transient context be distilled into LoRA weights, survive deletion, and transfer?Blocked at teacher gate
Why this existed
Can transient context be distilled into LoRA weights, survive deletion, and transfer?
What was tried
Built a teacher/student LoRA comparator, repaired early leakage, ran five seeds, and ran teacher, prompt-contract, and trajectory diagnostics.
What happened
Four of five student holdout deltas were +0.3333, but the teacher was 0.17647 on every seed against a fixed 0.2 gate.
Why
Prompt integrity was clean; the teacher’s task/prompt expressivity ceiling was too weak. Token loss fell while DEV behavior remained weak and unstable.
What it changed
The student signal is diagnostic only. Identical reruns are retired; the threshold was not lowered.
19LM as language organCan learned cortex state control a frozen LM through a latent interface, separately from label-prefix retrieval?Blocked at articulation gate
Why this existed
Can learned cortex state control a frozen LM through a latent interface, separately from label-prefix retrieval?
What was tried
Implemented matched label and latent arms plus zero, swap, shuffle, retrieval, and oracle controls; reached a clean fourth calibration attempt.
What happened
The text oracle ceiling was 0.357143, but latent Arm B did not beat random-cortex C1. No holdout was accessed.
Why
The organ can express the task when given text, but the learned soft interface did not create measurable DEV articulation. A claimed C7 adapter is also absent in the real path.
What it changed
Diagnose the coupler and adapter on DEV before any signoff or claim run.
20Meta-trained cortex over a frozen language organCan a meta-trained writer, reader, consolidator, and coupler learn an unseen rule online without retrieval or online backprop?DEV training active; science blocked
Why this existed
Can a meta-trained writer, reader, consolidator, and coupler learn an unseen rule online without retrieval or online backprop?
What was tried
Built the fixed-shape cortex, frozen Qwen organ, causal controls, INT8 v2 instrument, five-seed developmental campaign, and 90-shard calibration plan.
What happened
Four canonical checkpoints are runtime-verified; the fifth seed’s 8-task retry was running at the live check. Meta-test has not run.
Why
Thirty- and fifteen-task CPU jobs exceeded Kaggle’s 12-hour limit for one seed; retries preserve seed and optimizer identity while reducing only developmental task count.
What it changed
This is the live engineering frontier. Calibration, power analysis, candidate freeze, and explicit human signoff still precede any scientific verdict.
21Cortex-routed frozen specialist organsCan one learned cortex route state into independently frozen language and action specialists?Specification; blocked by R20
Why this existed
Can one learned cortex route state into independently frozen language and action specialists?
This corrects stale repository indexes: Experiment 08 now has a code-backed curriculum, and Experiment 09 has a real DEV implementation and active campaign. Neither implementation alone is a scientific win.
01
Correction-to-Competence
Null
Purpose
Build a behavior-only scorecard that can see beyond saturated internal metrics.
Execution
One Colab campaign run with mock structural control.
Result
Behavior delta 0.0; discrimination 0.0; five sub-metrics de-saturated.
Interpretation
The instrument learned to say “no behavioral transfer.” That is a valid null, not a broken run.
Read all 73 classifications and the full failure lineages.
Search invalidated claims, metricless runs, refutations, provider timeouts, implementation repairs, nulls, positives, and the exact evidence that superseded each headline.