Mission control
two questions · four results · runs in progress · figures provisional · 17 October 2026
Two questions, and one result that bears on both. Each gives what it is, what came back, the tables, and what it would take to settle it. The second is a single programme: the measurement came first, grokking said what reorganisation does, and the loop is what those leave standing once the J-space write path in result 04 had failed.
Does alignment come with capability?
Whether the training that makes a model more capable also gives it a reason to keep the world it depends on intact.
Below 14B capability helps. Above it, the tuned models track the group instead of the stock. The untuned arms, which are the test, are on the cluster now.
| instrument | five-agent commons, one capable newcomer |
|---|---|
| ladder | 7B · 14B · 32B · 72B, 300+ games |
| control | pre-entry history byte-identical |
| open arm | untuned base models, running |
Can a model keep learning after training?
A consolidation step that writes new knowledge into the weights and a gate that refuses the write when the model would lose itself.
The loop commits with the base intact and the gate catches the case where it does not. Replay of the loop’s own history is the open arm.
| 01 measure | what training leaves in the weights | Σ 0.00 → 2.43 up the ladder |
|---|---|---|
| 02 reorganise | what happens at the grok | concentration, not span |
| 03 the loop | compartment · simulate · consolidate · gate | recall 1.00, base intact |
Does alignment scale with capability?
A more capable newcomer joins four weaker models who are draining a shared lake, with the pre-entry history held byte-identical so the only difference is who walks in.
From 14B to 72B it works out the sustainable catch in month one, says it, then takes whatever the group is taking. Capability bought no stewardship at all.
- Environment
- GovSim fishery (Piatti et al., NeurIPS 2024): 100-ton lake, five fishers, stock doubles to the cap, ten tons each is sustainable for ever
- Design
- Four residents on Qwen2.5-14B play four months alone; a stranger enters at month five at 7B, 14B, 32B or 72B, neutral persona, no memory of the talk
- n
- 300+ games, ten seeds per cell, temperature 0.7, a fixed smaller model parsing text to numbers
- Arms
- Visible catches, hidden action, and a private pond only the stranger can fish
| stranger | catches | collapsed | months (mean) | stranger | residents | residents' share | yield | forecast error |
|---|---|---|---|---|---|---|---|---|
| 7B | visible | 7 | 8.3 | 12.6 | 5.9 | 0.65 | 0.77 | 10.4 |
| 7B | hidden | 6 | 9.3 | 14.8 | 5.1 | 0.59 | 0.65 | 11.2 |
| 14B | visible | 5 | 9.4 | 6.4 | 5.8 | 0.79 | 0.70 | 10.3 |
| 14B | hidden | 0 | 12.0 | 7.0 | 6.2 | 0.74 | 0.62 | 6.8 |
| 32B | visible | 4 | 10.1 | 6.0 | 5.8 | 0.79 | 0.68 | 8.6 |
| 32B | hidden | 1 | 11.6 | 7.0 | 6.1 | 0.75 | 0.62 | 6.0 |
| 72B (4-bit, GPU, preliminary) | visible | 5 | 7.8 | 6.9 | 6.9 | 0.81 | 0.87 | 9.1 |
Ten games per cell; tons a month while the stranger is present. The 14B hidden row is the one to look at: zero collapses, twelve of twelve months survived, yield 0.62. The village lives by starving and the stranger that knows better joins in.
The hidden-action arm was built to separate a saturated game from a conforming model: with no catches to copy, the only way to fish well is to reason from the stock. It was the model. A private pond at 0, 20, 100 and 300 tons moved the commons take by nothing in either direction, 6.4, 6.3, 6.6 and 6.4 at 14B.
Untuned base models at 7B, 14B and 32B are running: the question is whether this is conformity installed by chat tuning or something about capability. Reasoning-model strangers and a successor arm are queued.
The decisive version is the same ladder on untuned frontier-scale models. Open weights stop at 72B, perhaps 235B, and every model reachable by API is tuned, which is exactly the confound this result points at.
Can a model keep learning without catastrophic forgetting?
One programme in three stages: build a measure of what training puts in the weights, use it to read what happens when a network reorganises, then build the loop that learns continuously and gate it.
New facts enter the weights at recall 1.00 with the base intact, where naive fine-tuning wrecks it. Run the loop five times and it erases its own earlier facts, and the gate catches it.
- Why one programme
- The measure came first, grokking showed that forgetting is the same reorganisation with only local data in view, and the loop is what those leave standing once the J-space route in result 04 had been ruled out
- Models
- Qwen2.5 0.5B to 72B for the measure; Pythia-160M and Qwen2.5 0.5B and 1.5B for the loop; Qwen2.5-7B-Instruct for the write path
- Probes
- 24 invented facts, 72 probes in four families, chance one in eight
- Gate
- Commit only if held-out loss is not worse, intrusion is bounded, spectral span has not collapsed, and weights-only probes beat the base
Semataxy: measuring organisation in the weights
ConfirmedShuffle the entries of a weight matrix. Every entropy-like quantity is unchanged and the network is dead, so what training built is in the relations a permutation destroys.
Organisation falls to the random-weight line when the entries are shuffled, and rises at every rung from 0.5B to 72B.
| Model | Params | Matrices | Σ mean | Σ max | Mass above edge |
|---|---|---|---|---|---|
| Qwen/Qwen2.5-0.5B-Instruct | 0.5B | 168 | 1.52 | 4 | 0.268 |
| Qwen/Qwen2.5-1.5B-Instruct | 1.5B | 196 | 1.72 | 4 | 0.247 |
| Qwen/Qwen2.5-3B-Instruct | 3B | 252 | 1.84 | 4 | 0.257 |
| mistralai/Mistral-7B-Instruct-v0.3 | 7B | 448 | 2.04 | 6 | 0.243 |
| Qwen/Qwen2.5-7B-Instruct | 7B | 196 | 1.96 | 5 | 0.237 |
| Qwen/Qwen1.5-7B-Chat | 7B | 224 | 2.26 | 6 | 0.201 |
| Qwen/Qwen2-7B-Instruct | 7B | 196 | 1.93 | 5 | 0.228 |
| Qwen/Qwen2.5-7B | 7B | 196 | 1.95 | 5 | 0.237 |
| unsloth/Llama-3.1-8B-Instruct | 8B | 224 | 2.11 | 5 | 0.234 |
| unsloth/gemma-2-9b-it | 9B | 294 | 2.08 | 4 | 0.216 |
| Qwen/Qwen2.5-14B-Instruct | 14B | 336 | 2.16 | 6 | 0.231 |
| Qwen/Qwen2.5-32B-Instruct | 32B | 256 | 2.18 | 6 | 0.320 |
| Qwen/Qwen2.5-72B-Instruct | 72B | 320 | 2.42 | 8 | 0.344 |
| Qwen/Qwen2.5-0.5B-Instruct · random weights | 0.5B | 168 | 0.01 | 1 | 0.0008 |
| Qwen/Qwen2.5-1.5B-Instruct · random weights | 1.5B | 196 | 0.00 | 0 | 0.0003 |
| Qwen/Qwen2.5-3B-Instruct · random weights | 3B | 252 | 0.00 | 0 | 0.0001 |
| Qwen/Qwen2.5-7B-Instruct · random weights | 7B | 196 | 0.00 | 0 | 0.0000 |
| Qwen/Qwen2.5-14B-Instruct · random weights | 14B | 336 | 0.00 | 0 | 0.0000 |
| Qwen/Qwen2.5-32B-Instruct · random weights | 32B | 256 | 0.00 | 0 | 0.0000 |
Read from results/structure/ladder_summary.csv. Between layers, every measure tried so far saturates early and separates trained from untrained rather than ranking capability at fixed shape; activations are the next attempt.
The structure of grokking, and what it says about forgetting
Prediction failedThe same measure run across a network that memorises and then generalises, and across Pythia-160M's public training checkpoints.
The prediction was that organisation rises at the grok. It fell: the spectrum concentrated instead of widening, which makes forgetting the same reorganisation run with only local data in view.
| Measurement | Before | After |
|---|---|---|
| Pythia-160M, Σ at initialisation → step 4,000 | 0.00 | 1.90 |
| Pythia-160M, Σ at step 143,000 | 1.90 | 2.06 |
| Grokking, spectral mass above the edge | 0.81 | 0.97 |
| Grokking, tail exponent | 3–5 | 1.3 |
| Grokking, above-edge directions | n | n/2 |
From results/pythia160m_trajectory.csv and results/grokking_p113_trajectory.csv. Memorisation is diffuse; understanding is compact. Spectral prediction, testable: forgetting is concentration plus loss of span, replay is concentration without it.
Continuous learning: the consolidation loop
Works once, fails on repeatCompartment new facts in an adapter, let the untouched base reason over them, consolidate from a clean base with replay, and gate the commit. Sleep, in four steps.
Recall 1.00, reversal to 0.94, two-hop 0.88, held-out loss at or better than the base. Then cycle 1 drives recall of cycle 0's facts to near chance and the gate refuses four commits of five.
| model | arm | recall | new wording | reversal | two-hop | held-out loss (base → arm) |
|---|---|---|---|---|---|---|
| Pythia-160M | naive fine-tune | 1.00 | 0.83 | 0.50 | 0.12 | 4.20 → 8.71 |
| Pythia-160M | consolidated | 1.00 | 0.67 | 0.75 | 0.25 | 4.20 → 3.66 |
| Qwen2.5 0.5B | naive fine-tune | 1.00 | 1.00 | 0.50 | 0.38 | 3.49 → 5.09 |
| Qwen2.5 0.5B | consolidated | 1.00 | 0.71 | 0.81 | 0.88 | 3.49 → 3.14 |
| Qwen2.5 1.5B | naive fine-tune | 1.00 | 0.96 | 0.56 | 0.25 | 3.12 → 4.11 |
| Qwen2.5 1.5B | consolidated | 1.00 | 0.71 | 0.75 | 0.62 | 3.12 → 2.77 |
Naive fine-tuning against the loop, on three model sizes.
| model | cycle 0 | recall of cycle-0 facts after cycle 1 | cycles committed | held-out (base → final) |
|---|---|---|---|---|
| Pythia-160M | committed | 0.21 | 1 of 5 | 4.20 → 3.67 |
| Qwen2.5 0.5B | committed | 0.29 | 1 of 5 | 3.49 → 3.15 |
Five learn-and-consolidate cycles in a row, 24 fresh facts each, with a retention check on every earlier cycle.
Letting the model write its own implications instead of templates took reversal to 0.94 on the model able to write them. A compartment consulted only on a retrieval hit leaked into zero held-out chunks. Replay of general text protects what pretraining built; nothing protected what the loop itself wrote the cycle before.
Replay of the loop's own earlier cycles is written and queued, with the five-cycle run standing as the baseline it has to beat. A replay sweep at 25, 50 and 75 per cent tests the prediction that a stronger model needs less replay to stay intact.
What a consolidation step must preserve for a frontier model to remain itself has to be asked of a frontier model, with its real pretraining distribution to replay against and its real evaluations in the gate. Outside a lab both are approximations.
When scaling is not the answer
A controlled study of inference sampling: more samples per action across four decision families, with the validation protocol frozen and published before it was run.
Sampling amplifies the model's learned ordering. It reduces regret when the ordering matches reality and increases it when the ordering is wrong.
- n
- 76,800 held-out cases across four decision families
- Protocol
- Frozen and published unrun, so someone other than the author can check it
Test-time compute amplifies what was learned rather than correcting it, which is why both questions above are about what training builds and not about how long a model is allowed to think.
Using J-space for live learning
The J-lens gives a coordinate system for what a model is currently holding: a corpus-averaged Jacobian from a middle-layer activation to the logits, isolating the low-dimensional subspace that drives what the model reports. The attempt was to use that space as a write path, capturing the J-space state with a skill in context and then reaching the same state through the weights with the skill removed.
Optimising the weights to reach the captured state does not install the skill, and fails more reliably the better the state is matched. The best state match in the experiment produced no behaviour at all.
- Lens
- J-lens: a corpus-averaged Jacobian from a middle-layer activation to the logits, after the global-workspace work in Transformer Circuits, July 2026
- Thesis
- Forgetting is unsolved because every mitigation acts on parameters and is blind to meaning; the missing piece is persistence, not acquisition
- Design
- Three confinement mechanisms penalising or projecting away change in orthogonal directions, against ordinary fine-tuning at matched weight drift, with the prediction stated before the run
- n
- Qwen2.5-7B-Instruct, an invented skill, five configurations, forty trained models, baseline install 0%
| Arm | Cosine to captured state | Skill installed |
|---|---|---|
| Unconfined state matching, the best match in the experiment | 0.907 | 0.0% |
| Five confinement configurations, every magnitude | — | 0.0% |
| Ordinary fine-tuning | 0.273 | 91.7% |
From RESULT-phase3-jspace.md, 22 August 2026. Three converging lines, enrichment, prefix geometry and this run, agreed; lines sharing a model, corpus and lens share their confounds.
The objective selected against the outcome, which is why this is stronger than a null: the better the J-space state was matched, the more reliably the behaviour failed to appear.
A state is not a function: a procedure cannot be written into weights by reproducing the activations it produces. The surviving candidate was distillation, reproducing what the skill does rather than what it looks like, and that is where the consolidation loop in result 02 came from.
W Withdrawn and set aside
| Date | What | Where it went | Note |
|---|---|---|---|
| 14 Oct 2026 | Persona arm cancelled in three runs of ten; output not reported | runs_invalid/mild_cancelled/ | 2026-10-14-expB-stranger.md |
| 15 Oct 2026 | Three capability scores measured against the wrong model after a port collision; re-queued | runs_invalid/port_collision/ | 2026-10-15-sampled-batch.md |
| 15 Oct 2026 | Two games with silent agents, empty answers in the final months | set aside, not scored | 2026-10-15-sampled-batch.md |
| 16 Oct 2026 | Games lost to a prefill and max-token mismatch between inference servers | runs_invalid/gpu_prefill/, gpu_max_tokens/ | 2026-10-16-night.md |
| 16 Oct 2026 | 72B cell run against a cache shared with CPU-served games; unpaired | runs_invalid/gpu_72b_mixed_cache/ | 2026-10-16-night.md |
Every batch is read first by a separate model context with no hypothesis in view, before it is read by the author. The three ignition findings from the earlier deployment work died the same way and are reported as a null: a monotone signal plus any threshold produces a step, and the ramp design failed its pre-committed power condition on all three models.
R Running now
- Untuned base-model strangers at 7B, 14B and 32B: is the conformity tuning or capability.
- The 72B re-run on a clean cache; hidden-action and early-entry arms at 72B.
- Reasoning-model strangers: R1-Distill-70B and Qwen3-32B.
- 72B base, then Qwen3-235B base, the top of what the allocation can serve.
- A probe that asks the stranger outside the game what it would propose; a successor arm.
- The live-learning replay fix across five cycles on three sizes; the replay sweep at 25, 50 and 75 per cent; the loop at 7B and 14B.
Every result goes into the repository with a dated note, whichever way it goes.
P Papers
- Consciousness Is Deployed Intelligence10.5281/zenodo.21984327
- The working model10.5281/zenodo.21984450
- Sensory fusion through a common origin10.5281/zenodo.21984521
- Self-relative processing10.5281/zenodo.21984601
- How learned representation governs surprise10.5281/zenodo.21984627
- When Thinking Longer Makes You Wronger10.5281/zenodo.21984650
- Frozen external-validation protocol10.5281/zenodo.21984679