| """Frozen constants for the sparse/semantic attractor probe — see | |
| ../PROTOCOL-SPARSE.md. Pinned before any job runs. | |
| Everything the frozen census pinned (token budget, SEQ, flow steps/seeds, | |
| clustering threshold, count rule, colored floor, lift measurement) is | |
| imported unchanged from `protocol.py` / `wprotocol.py`. This module adds | |
| only: | |
| 1. the domain-dense corpus conditions (Q1) and their document rules, | |
| 2. the depth-spanning block sets for the two censused models, | |
| 3. the tail-concentration flow (Q2) — one changed ingredient, the on-set, | |
| 4. the novelty rule against the models' existing Pile libraries, | |
| 5. the pre-registered verdict rules. | |
| Nothing here is chosen at runtime. | |
| """ | |
| import protocol as P # noqa: F401 (re-exported constants are used verbatim) | |
| # ========================================================================== | |
| # provenance: where the frozen pipeline and the Pile reference come from | |
| CENSUS_RAW_EID = "exp_01m0rjapkveyqr5fytnhtwdzg8" # raw GPT-2 census | |
| CENSUS_RAWNEW_EID = "exp_01m0sy4fjhefv9tv0nwz9cx1nd" # raw Pythia/Qwen census | |
| CENSUS_WGPT2_EID = "exp_01m0tedzxhejkrq70x0pvcsay1" # whitened GPT-2 census | |
| CENSUS_WNEW_EID = "exp_01m0xrq2h1fdertpjmkd9x80cs" # whitened Pythia/Qwen | |
| INTERP_EID = "exp_01m0zrgczdeemty0aep579zjnx" # label/validate pipeline | |
| # ========================================================================== | |
| # 1. the four measures (corpus conditions) | |
| # | |
| # `pile` is the reference measure of every completed census; it is NOT rerun | |
| # here — the stored per-block results are reused (identical slices, identical | |
| # protocol). The three new conditions are built with the census's own slice | |
| # construction: documents in dataset order, fit until >= FIT_TOKENS, held | |
| # from the first document the fit slice never touched, both truncated to a | |
| # multiple of SEQ. A document is a repo file / paper / conversation. | |
| CONDITIONS = ("pile", "code", "math", "chat") | |
| NEW_CONDITIONS = ("code", "math", "chat") | |
| CORPORA = { | |
| "code": dict( | |
| hf="codeparrot/codeparrot-clean", config=None, split="train", | |
| field="content", kind="text", | |
| note="Python repository files; document = one file"), | |
| "math": dict( | |
| hf="math-ai/AutoMathText", config="arxiv-0.70-to-1.00", split="train", | |
| field="text", kind="text", | |
| note="arXiv papers with AutoMathText math score >= 0.70; " | |
| "document = one paper"), | |
| "chat": dict( | |
| hf="HuggingFaceH4/ultrachat_200k", config="default", split="train_sft", | |
| field="messages", kind="chat", | |
| note="UltraChat SFT dialogues rendered as a plain-text transcript; " | |
| "document = one conversation"), | |
| } | |
| # Plain-text chat rendering. Deliberately NOT a model chat template: the | |
| # measure must be the same *text* for both tokenizer families, so the role | |
| # markers are ordinary words rather than special tokens. | |
| CHAT_ROLE_NAMES = {"user": "User", "assistant": "Assistant", | |
| "system": "System"} | |
| CHAT_TURN_SEP = "\n\n" | |
| CHAT_ROLE_SEP = ": " | |
| # Streaming caps: how many documents we are willing to walk per split before | |
| # declaring the corpus too short (guards a silent short slice). | |
| MAX_DOCS_SCAN = 200_000 | |
| MIN_DOC_CHARS = 64 # skip degenerate/empty documents | |
| # ========================================================================== | |
| # 2. models and blocks | |
| # | |
| # The two most-baselined small models, plus their random-init floors. | |
| # Blocks span depth at fixed fractions that reproduce each census's own | |
| # representative blocks: round(L/6), round(L/3), round(L/2), round(5L/6). | |
| MODELS = ("gpt2", "qwen3-0.6b") | |
| FLOOR_OF = {"gpt2": "gpt2-randinit", "qwen3-0.6b": "qwen3-0.6b-randinit"} | |
| TAGS = ("gpt2", "gpt2-randinit", "qwen3-0.6b", "qwen3-0.6b-randinit") | |
| BASE_OF = {"gpt2": "gpt2", "gpt2-randinit": "gpt2", | |
| "qwen3-0.6b": "qwen3-0.6b", "qwen3-0.6b-randinit": "qwen3-0.6b"} | |
| BLOCKS = {"gpt2": (2, 4, 6, 10), "qwen3-0.6b": (5, 9, 14, 23)} | |
| # ========================================================================== | |
| # 3. the tail-concentration flow (Q2) | |
| # | |
| # The frozen census flow ascends | |
| # Delta(v) = E[G(v; h) | h in top-q of a_v] - E[G(v; h)] | |
| # with q = FLOW_Q = 0.1 re-drawn inside each 1024-state batch, i.e. it | |
| # conditions on a ~10% event. The tail flow ascends the SAME pointwise | |
| # conductance under a rare-event conditioning: the on-set is the TAIL_M fit | |
| # states with the largest signed projection under the CURRENT direction. | |
| # Everything else — optimizer, learning rate, steps, batch, clamp, seeds, | |
| # init directions, unit projection — is the frozen flow's, unchanged. | |
| # | |
| # TAIL_MODE selects how that on-set is found. `exact` re-selects the corpus | |
| # top-M every step (one chunked projection of the whole fit slice; a few | |
| # seconds over the 200 steps at census scale), so the flow ascends the | |
| # pre-registered functional itself. `reservoir` is the cheaper persistent | |
| # approximation the plan sketched; it only ever sees the states the flow drew | |
| # (~40% of the corpus at this budget) and is kept as a robustness line. | |
| TAIL_Q = 0.001 # rare-event quantile (0.1%), search and measure | |
| TAIL_M = 399 # on-set size = max(30, round(TAIL_Q * N_fit)) | |
| # pinned for the 399,923-state census slices | |
| # Reporting-only diagnostic (not a decision rule): how many poles concentrate | |
| # their top-M states into fewer than N distinct held sequences. The library | |
| # group is bimodal, so group medians are reported alongside these counts. | |
| CONC_THRESHOLDS = (50, 100, 200) | |
| TAIL_MODE = "exact" # production selection | |
| TAIL_MODE_SENS = "reservoir" # robustness line, one block per model | |
| TAIL_LIFT_Q = TAIL_Q # held-out measurement quantile for the tail arm | |
| TAIL_STEPS = P.FLOW_STEPS # 200, frozen | |
| TAIL_BATCH = P.FLOW_BATCH # 1024 off-set draws per step, frozen | |
| TAIL_LR = P.FLOW_LR # 0.05, frozen | |
| TAIL_CLAMP = P.FLOW_CLAMP # 6.0, frozen | |
| TAIL_SEED = P.FLOW_SEED # 3, frozen (same batch stream as the raw flow) | |
| # Concentration diagnostics measured for every accepted axis (and, as the | |
| # comparison group, for the block's standard attractors). | |
| CONC_TOPK = TAIL_M # states in the measured tail | |
| CONC_SEQ = P.SEQ # a "moment" is a 127-token sequence | |
| # ========================================================================== | |
| # 4. novelty against the existing Pile libraries | |
| # | |
| # An attractor axis found here is NOVEL iff its centroid's largest absolute | |
| # cosine against the union of the model's Pile attractor axes — raw census | |
| # and whitened census, ALL blocks — is below JOIN_COS. The bar is the | |
| # clustering / identity threshold used by every census in this line of work. | |
| NOVEL_COS = 0.9 # = P.CLUSTER_COS = wprotocol.JOIN_COS | |
| NOVEL_FRAMES = ("raw", "whitened") # reported in both; raw is primary | |
| # Secondary (reported, not part of the rule): the join against ALL Pile | |
| # cluster centroids, attractor or not, and against the same block only. | |
| # ========================================================================== | |
| # 5. exemplar extraction + labelling (interpretation pipeline, exp-8) | |
| EXEMPLAR_TOKENS = 2_000_000 # per (family, condition), truncated to SEQ | |
| PILE_EXEMPLAR_START_ROW = 363 # exp-8 NEW_SLICE_START_ROW (untouched rows) | |
| LABEL_CATEGORIES = ("generic_syntactic", "domain_structural", | |
| "semantic_topical") | |
| CATEGORY_DESC = { | |
| "generic_syntactic": | |
| "domain-independent surface syntax: punctuation, sentence or line " | |
| "boundaries, function words, word fragments, whitespace, formulaic " | |
| "continuations — the kinds already catalogued in the Pile library.", | |
| "domain_structural": | |
| "structure specific to this domain's format: LaTeX environments or " | |
| "math mode, code block/indentation/bracket structure, identifiers as " | |
| "syntax, dialogue turn boundaries or role tags, markup.", | |
| "semantic_topical": | |
| "content rather than format: a topic, an entity or entity class, a " | |
| "mathematical concept, a subject-matter category.", | |
| } | |
| # ========================================================================== | |
| # 6. pre-registered verdict rules (../PROTOCOL-SPARSE.md section 6) | |
| # | |
| # Q1 (density). Per condition c, over the 8 cells (2 models x 4 blocks): | |
| # n_novel(cell) = novel attractor axes in the trained model, either | |
| # census arm, passing that block/arm's own count rule | |
| # n_novel_floor(cell) = the same quantity for the random-init model of the | |
| # same architecture under the same condition, joined | |
| # against ITS OWN Pile library | |
| # supported iff sum_cells n_novel >= 1 AND one-sided Wilcoxon signed-rank | |
| # (trained > floor) over the 8 paired cells has p < ALPHA. | |
| # Q1b (content). Every novel pole that earns a detection-validated label is | |
| # categorised into LABEL_CATEGORIES by the judge; counts are reported with | |
| # exact binomial CIs. "Semantics under a domain-dense measure" requires at | |
| # least one validated novel pole categorised semantic_topical. | |
| # Q2 (sparsity in time). Identical shape, with the tail census in place of | |
| # the condition census and the Pile measure throughout. | |
| ALPHA = 0.05 | |
| N_PERM = 10_000 # permutation draws where a permutation null is used | |
| PERM_SEED = 42 | |
| BOOT_N = 10_000 | |
| BOOT_SEED = 42 | |
| WILCOXON_ALT = "greater" # trained > floor, one-sided | |
| # Candidate-matched chance level for the novelty join: under a uniform random | |
| # direction in d dimensions, cos^2 ~ Beta(1/2, (d-1)/2); the reported chance | |
| # level is the expected maximum |cos| over the same number of library | |
| # candidates (order statistic, Monte-Carlo with this seed). | |
| CHANCE_SEED = 7 | |
| CHANCE_DRAWS = 20_000 | |