04 — Context-Scoped Semantic Attractors

Two senses of one token must coexist as distinct basins — correcting one sense must not obliterate the other.

Status: ACCEPTED (2026-07-11, single-run) | Thesis anchor: experiments.txt §7 (energy/attractor memory), §1 (correction-gated cortex), §4 (learned plasticity) | Goal anchor: Goal 3 (organ tensor upgrades; scope-control done-when: "correction of one sense does not obliterate the other because they live in different cortex state regions"), relates Goal 2 (layer-L peek) | Depends on / relates to: 03-layer-l-hidden-extraction, 01-correction-to-competence-benchmark, 06-bounded-growth-consolidation, 05-metabolism-loop-closure

Outcome (2026-07-11): ACCEPTED (single-run caveat). Campaign 0d48130 (colab, commit 537260c): scope_selectivity_index=1.0 — context-addressed slots achieved perfect sense selectivity on the Stage-2 episodes. Caveat: single-run, no cross-seed variance data. The SSI ≥ 0.5 acceptance threshold is met; the single-slot baseline comparison and obliteration_rate were not separately reported in the campaign metrics. Evidence: ../experiments_logs/2026-07-11_campaign_0d48130.md.

Problem

The organism cannot hold two meanings of one word at once. Stage 2 of the curriculum (src/oczy/experiments/organism_curriculum/stages/stage_2_scope.json, 8 episodes: s2_log, s2_file, s2_key, s2_cell, s2_record, s2_branch, s2_model, s2_run) teaches a technical sense of a token already known in a common sense, then tests both. Each episode carries two match_mode="sense" probes: a retention probe in the teaching context and a scope probe in a different context. For s2_log the correction is "No, 'log' here means system error log."; the retention probe "Log the runtime error." expects system error log, and the scope probe "Show the log." expects captain's journal.

The scope test fails 100%. On the real LFM2.5 driver, Stage 2 uptake stayed 0/8 across runs #73–#77. The only variant that moved the needle (run #78, policy_suppresses_fast_answer) reached 0.62 stage-2 uptake (and only 0.50 on the separate scope-discrimination metric) but collapsed Stage 0 retention to 0.62 and Stage 1 transfer to 0.12 — it bought scope by destroying the rest. That trade-off is the whole problem: the cortex has one global state slot, so teaching sense B moves the same vector that encoded sense A.

The mechanism is explicit in the reference cortex. KVCortex holds warm_state and cold_state as single d_cortex vectors (kv_cortex.py:107-108), and observe() updates that one vector by a global EMA: warm = (1-plasticity)*warm + plasticity*tanh(proj_hidden @ h) (kv_cortex.py:173-225). There is no context address. Whatever the current request, the same warm_state projects into the same cvecs. The real-driver curriculum runs this at KVCortexConfig(d_cortex=4) (run_curriculum.py:38-43) — four scalars to hold every sense of every word. One basin, globally reshaped: exactly the overgeneralization the thesis warns about ("If the basin gets reshaped globally, 'profile' might become business vertical everywhere", experiments.txt:427).

The benchmark that should catch this is saturated: code_qa_accuracy=1.0 across runs #75–#79. We need both a new cortex mechanism and a non-saturating scope metric (cross-link 01).

Hypothesis

• H1 (mechanism). If the cortex stores correction-deltas in context-addressed slots — a small associative store keyed by the request's hidden state, read by similarity at articulation time — then two senses of one token settle into two distinct basins, and correcting one sense leaves the other's retrieval intact. Measured: a single-slot baseline scores Sense-Selectivity-Index (SSI, both probes correct per episode) ≤ 0.125, while the context-addressed cortex scores SSI ≥ 0.5 on the same 8 episodes, with retention and scope accuracy both ≥ 0.75 (no run-#78 trade-off).
• H2 (selectivity for free). Because the common sense is the LM's natural prior, a context-addressed read that returns ~zero steering when no basin matches the request will preserve the common sense automatically. Measured: obliteration_rate (taught technical sense leaking into the common-sense context) drops from ~1.0 (single slot) to ≤ 0.25 (context-addressed), without a prefix that bakes the answer in.

Why now / what unblocks it

• The cvec surface already does the right kind of thing for this test. cvec shifts semantic DOMAIN/posture reliably (domain_co_recall 1/1, run #95) even though it cannot force an arbitrary exact token. Stage-2 scoring is match_mode="sense" (token overlap minus stopwords and the ambiguous token, scoring.py / dataset.Episode.ambiguous_token() at dataset.py:96), i.e. a domain-level discrimination. So the failure is not the cvec ceiling — it is the single global slot. Context-addressing is the missing piece, and it is addressable in pure numpy on top of the existing KVCortex/CortexAgent plumbing.
• Thesis §7 is a literal spec for this: energy E(h, context), correction lowers the desired interpretation's energy and raises the wrong one, "the basin must be scoped by context" (experiments.txt:419-427). Modern-Hopfield attention is the read rule. We are implementing the named design, not inventing one.
• Context keys can come from peek_embedding (final-layer mean-pooled, available today). If the two short requests "Log the runtime error." vs "Show the log." do not separate at the final layer, that failure directly motivates and is unblocked by 03-layer-l-hidden-extraction (mid-layer peek_layer). The experiment is designed to diagnose which.

Approach

Make warm/cold cortex state context-addressed instead of global, following thesis §7's energy/attractor framing and §1's correction gate.

• Slot store. Add an associative memory of M slots, each a (key_m ∈ d_embd-or-projected, delta_m ∈ d_cortex) pair, on top of KVCortex (a wrapper in experiments/, leaving the 9/9 reference contract untouched).
• Correction-gated write (basin carving). On observe(hidden, correction_signal), compute the candidate delta tanh(proj_hidden @ hidden) (reuse proj_hidden) and the context key from the request hidden. Find the nearest slot; if max similarity < alloc_threshold (novel context) allocate a new slot, else EMA-update only that slot with the correction-gated plasticity. Writes are local — they cannot reshape a basin that the current context does not address (§7's scoping; §4's plasticity gate).
• Similarity read (settling). At articulation, compute the request key, softmax-attend over slot keys (temperature β), set warm_state = Σ softmax(key·key_m/β) · delta_m, then emit cvecs as today. Gate the read: if max similarity < read_threshold, return zeros → no steering → the LM falls into its natural (common-sense) basin (H2).
• Bounded growth. Cap slot count and merge near-duplicate keys, tying allocation to cross-link 06-bounded-growth-consolidation and the north-star behavior_delta_per_byte_of_persistent_memory (rl_pipeline_design.md:342). consolidate() folds stable slots into cold storage (§ slow change / forgetting raw trace).

Success criteria

Behavioral, on the 8 Stage-2 episodes, real LFM2.5 driver. Replaces the saturated code_qa_accuracy / old binary scope-uptake with a joint metric that cannot be gamed by collapsing to one sense.

• PASS: context-addressed cortex achieves SSI ≥ 0.5 (≥ 4/8 episodes with retention AND scope both correct), with retention_acc ≥ 0.75 AND scope_acc ≥ 0.75, while the matched single-slot baseline scores SSI ≤ 0.125 (consistent with current 0/8). And obliteration_rate ≤ 0.25 (H2).
• Discriminating-by-construction: SSI is the per-episode conjunction. An always-technical cortex fails every scope probe; an always-common cortex fails every retention probe; only genuine per-context selectivity scores. Current value is ~0 (briefs), so there is full headroom — it does not start at 1.0.
• KILL (mechanism): if the oracle-key context-addressed condition (clean orthogonal key per request) cannot beat the baseline SSI by ≥ 0.25, the read/write addressing itself is insufficient and context-addressing is not the lever — pivot away.
• KILL (key quality, hand off to 03): if oracle-key passes but peek_embedding-key SSI ≤ baseline+0.125, the final-layer pooled key cannot separate the two senses → escalate to 03 (mid-layer peek_layer) rather than claim success.
• KILL (growth): if allocated slots exceed 2× the number of distinct request contexts (~16), allocation is uncontrolled → fail (defer to 06).

Risks & open questions

• Final-layer key collision. "Log the runtime error." and "Show the log." are short and share log; final-layer mean-pooled embeddings may not separate them. This is the most likely failure and is the diagnostic that hands the problem to 03.
• cvec answer-path leakage. The curriculum answers via the LM's own decoding. If the LM ignores a weak gated cvec and answers from its prior in both contexts, scope passes trivially but retention fails. The basin must steer hard enough in-context yet read ~zero out-of-context — the read_threshold is the knob; risk it has no clean setting (mirrors the cvec scale cliff, GOALS.md / 2026-06-24 sweep).
• Mock has no semantics. _MockDriver keys (n_embd=16, idx=sum(ord(c))%16) cannot carry meaning; mock is a mechanism-only control (does allocation/read fire correctly), never a semantic pass.
• Open: should slot keys be the raw peek_embedding or a learned proj_key projection? Should basins be per-token or global-by-context? Does consolidate() need a per-slot cold store (vs the single cold_state vector today, kv_cortex.py:354-408)?

Prior evidence

• Stage-2 fails 0/8 on the real driver across runs #73–#77; run #78 reached stage-2 uptake 0.62 (scope-discrimination 0.50) only by collapsing Stage 0→0.62 / Stage 1→0.12 (RL-pipeline brief, 2026-06-26_policy_head_ranking_loop.md).
• The cortex's single global warm_state EMA: kv_cortex.py:107-108 (state vectors), kv_cortex.py:173-225 (observe), real-driver d_cortex=4 (run_curriculum.py:38-43).
• cvec does domain not exact tokens: domain_co_recall 1/1 exact 0/0 (run #95, 2026-06-27_domain_recall_metric.md); five-method cvec exact-token ceiling (2026-06-27_contrastive_cvec_discovery.md).
• Saturated benchmark: code_qa_accuracy=1.0 runs #75–#79 (SUMMARY.md).
• Scoring contract: scoring.matches() sense mode, dataset.Episode.ambiguous_token() (dataset.py:96).
• Thesis §7 energy/attractor + scope-by-context + overgeneralization danger (experiments.txt:419-427).