# Phase I synthesis — Experiments 000–008 Nine experiments, one synthetic world (12 chains, `body → element → radiation`, seed 1337), one frozen `Qwen3-0.6B` @ `c1899de289a04d12100db370d81485cdf75e47ca`, one evaluation suite (48 recall / 24 composition / 12 control), one host. --- ## Established ### Storage is not the problem Every mechanism permitted to train freely learned the world at ceiling. | mechanism | params | recall | |---|---|---| | LoRA r16 / r8 / r4 | 10.1M / 5.0M / 2.5M | **48/48** each | | isolated residual module (004) | 2.8M | **48/48** | The frozen base scores 0/48 before any memory is attached, verified before every experiment and never once drifting. Detachable parameter deltas can encode previously unknown synthetic facts. This was settled by Experiment 000 and has never been the binding constraint since. ### Detachability works, mechanically Across every mechanism, memory can be attached, removed and restored with the frozen base bit-identical throughout. The base-weight fingerprint has been `52683d0e840fb51e` since Experiment 004 and is re-verified at the start and end of every run. `gate = 0` reproduces the frozen model's logits **exactly** (max |logit diff| `0.000e+00`), proven with randomised non-zero memory parameters so the check is not vacuous, and `gate = 1` demonstrably changes the computation. Removal restores bit-identical logits. The load/unload primitive is viable and is the least equivocal result in Phase I. ### Interference is universal, and does not track parameter count **No mechanism in nine experiments has touched the frozen reasoner without measurable behavioural cost.** | mechanism | params | interference | substantive churn | |---|---|---|---| | LoRA r4 | 2,523,136 | **0.0833** | 8/12 | | LoRA r8 | 5,046,272 | 0.2500 | 10/12 | | LoRA r16 | 10,092,544 | 0.3333 | 9/12 | | isolated residual (004) | 2,811,200 | **0.5000** | 9/12 | | explicit KV (005) | 739,456 | 0.2500 | 8/12 | The smallest mechanism tested (739,456 parameters) is more disruptive than one 3.4× larger. The mechanism that touches no base projection at all is the *most* disruptive measured. **Physical parameter isolation does not imply behavioural isolation** — Experiment 004's central finding, and it has held. Experiment 002 also established that a zero net interference score does not mean unchanged behaviour: the best configuration still changed 11 of 12 control answers, 8 substantively, while scoring 0.0833. ### Composition is the persistent failure Perfect recall does not produce multi-hop reasoning. | mechanism | composition (thinking on) | p vs 1/6 | |---|---|---| | LoRA r16 | **10/24** | **0.0033** | | LoRA r4 | 5/24 | 0.37 | | LoRA r8 | 4/24 | 0.58 | | isolated residual | 0/24 | 1.00 | | explicit KV (005, 006) | 0/24 | 1.00 | **LoRA rank 16 remains the only configuration ever significantly above chance, and it costs 0.3333 interference.** Experiment 000 also established the condition: composition reaches 0.4167 *with thinking enabled* and exactly 0.0000 without it. The chain of thought carries the intermediate; without it the model performs hop 1 and emits the intermediate as its final answer, on all 24 items. Experiment 002 ruled out training exposure as the missing factor — capacity is what matters — and Experiment 003 showed rank 8 is not an intermediate point but rank-16 damage with rank-4 benefit. ### Addressing failed three times, for three different reasons | | why it failed | what it established | |---|---|---| | **005** | the address path received **exactly zero** gradient at step 0 (zero-initialised readout ⇒ `dL/d(read) ≡ 0`) | a differentiable slot softmax collapses to an input-independent argmax; 4,285 of 4,287 accesses selected slot 63 | | **006** | anti-collapse pressure was applied to probability **mass**, not to the **argmax** | ~21 effective soft slots coexisted with **1** dominant slot. Utilisation regularisation cannot reach selection | | **007** | direct 24-way supervision broke the collapse; the representation could not support it | 7 dominant slots, held-out accuracy 0.2292; **hop 48/48, chain 11/48** — the address encoded the question's *type*, not the fact's *identity* | | **008** | — | layer 18 is a near-degenerate region: 24 centroids at mean pairwise cosine **0.9977**. The representation appears abruptly at layer 22 | --- ## Falsified or materially weakened * **"Interference is governed by adapter size."** Suggested by 001–003; contradicted by 004 and 005. What correlates better is how freely a mechanism's output can vary independently of the reasoner's state. * **"Parameter isolation gives behavioural isolation."** Directly falsified by Experiment 004, the most disruptive mechanism measured and the one touching no base weight. * **"Slot-utilisation diversity indicates addressability."** Falsified from both directions: Experiment 006 had ~21 effective soft slots with 1 hard address; Experiment 007 had ~107 effective soft slots with 7. Neither statistic predicts the other. * **"Composition needs more training exposure."** Falsified by Experiment 002 — capacity, not exposure. * **"Layer 18 addressing failure is a property of the model."** Weakened by Experiment 008. The same probe on the same split reaches 0.8958 at layer 22 and 0.2292 at layer 18. The failure was substantially *local*. * **"Experiment 007's trained address head extracted what was there."** Weakened: at layer 18 a zero-parameter class mean matches it exactly — 11/48, identical confusion structure — with **disjoint** successes. 148,608 trained parameters bought nothing over a class mean at that depth. --- ## Still unknown * **Whether a memory attached at layer 22–26 would work.** Experiment 008 measured that fact identity is *legible* there. Legibility is not retrievability, and GOAL_8_FINAL forbids testing it. It is a Phase II question. * **Whether the margins are usable.** Even at the best cell the geometry is razor-thin: within-fact cosine 0.9605 against between-fact 0.9395, mean margin **0.0047**, minimum margin **negative**. A high nearest-centroid accuracy on such margins may not survive a mechanism that has to act on them. * **Whether recursive retrieval is possible at all.** No experiment has yet produced a memory that retrieves hop 1 reliably enough to ask the question. The chain `retrieve A → query with A → retrieve B` has never been observed. * **Whether interference is avoidable in principle**, or intrinsic to writing into a frozen residual stream. * **Everything about generality.** One world, one seed, one 0.6B model, one insertion strategy, 24 facts, 48 held-out probes. `jetson2` has been kept untouched for exactly this reason and has still not been used. --- ## The architectural assumption Phase I should retire > That a *single frozen hidden state at a mid-network insertion point* provides a > stable, paraphrase-invariant address for a fact. Experiments 005–007 all rest on it. Experiment 008 shows it is false where those experiments placed it — layer 18's fact centroids are nearly the same vector — and only marginally true at the best location found, on cosine margins of a few thousandths. What should *not* be retired on this evidence: detachability, which works; the finding that storage is easy; or the possibility that a later-layer address is viable. Experiment 008 did not test that, by instruction.