""" hv-reader ========= The reading experience, in one call. Given a text, produce a ReadingProfile: pace, memory, passes, slip, and wall — the five axes that describe what it is like to read a text. This unifies seven component models: hv-tempo → pace hv-forget → memory hv-ttu → ttu_s (total time) hv-fold → passes hv-slip → slip hv-hunger → hunger (internal, feeds slip and wall) hv-wall → wall The profile is the artifact. The axes are the readings. The signature is what predicts whether a text gets finished. Pure stdlib. No dependencies. Author: zeechimp License: Apache-2.0 """ from __future__ import annotations import argparse import json import math import os import re import sys from collections import Counter, defaultdict from dataclasses import dataclass, asdict, field from typing import Any, Dict, List, Optional, Set, Tuple # ============================================================================ # Lexicons # ============================================================================ COMMON_WORDS = frozenset(""" the be to of and a in that have i it for not on with he as you do at this but his by from they we say her she or an will my one all would there their what so up out if about who get which go me when make can like time no just him know take people into year your good some could them see other than then now look only come its over think also back after use two how our work first well way even new want because any these give day most us is are was were been being has had having does did doing will would shall should can could may might must man woman child water fire earth air sun moon star light dark hand head eye ear mouth nose foot leg arm body face heart mind life death food bread milk meat fish tree flower grass leaf root seed farm field hill mountain river sea lake boat ship road street city town house room door window wall floor roof bed chair table book page word line letter number name place thing part side end start middle top bottom front back left right high low long short big small old new young hot cold wet dry clean dirty light heavy soft hard fast slow easy true false good bad happy sad love hate fear hope help hurt win lose give take send bring buy sell pay cost money price work play run walk jump sit stand sleep wake eat drink cook wash read write speak hear see feel know think learn teach ask answer tell show hide open close push pull carry hold drop throw catch break fix build make do try use move turn stop start keep leave stay wait meet join save spend show thank want wish walk stop continue morning climb row push pull begin finish start end remain rest return arrive depart leave enter exit follow lead sit stand lie rise fall drop still quiet calm slow fast soft loud bright dark warm cool fresh clean never always often sometimes rarely usually speak spoke spoken take took taken give gave given see saw seen know knew known think thought thought come came come go went gone say said said tell told told find found found hold held held bring brought brought buy bought bought teach taught taught catch caught caught build built built send sent sent spend spent lose lost lost lead led led meet met met read read read write wrote written run ran run swim swam swum thing things word words time times year years day days man men woman women child children person people place places work works way ways life lives hand hands eye eyes part parts end ends line lines side sides name names head heads house houses friend friends family families group groups country countries world worlds city cities school schools """.split()) ABSTRACT_SUFFIXES = ( "tion", "sion", "ism", "ity", "ness", "ance", "ence", "ship", "hood", "ment", "ology", "itude", "acy", ) SUBORDINATORS = frozenset(""" which that because although though while whereas since when if unless provided assuming given whenever wherever whoever whichever """.split()) HEDGES = frozenset(""" may might maybe perhaps possibly probably typically usually often generally roughly approximately about somewhat rather """.split()) CONDITIONALS = frozenset(""" if when unless provided assuming suppose supposing """.split()) NEGATIONS = frozenset(""" not no never none without cannot can't don't doesn't won't isn't aren't wasn't weren't nor neither """.split()) BE_FORMS = frozenset(""" is are was were be been being am """.split()) REENTRY_MARKERS = frozenset(""" above below previous preceding following aforementioned noted mentioned discussed described stated referred earlier later """.split()) ARTICLES = frozenset(""" the this that these those such said aforementioned """.split()) UNIQUE_REFERENTS = frozenset(""" sun moon earth world sky ground horizon morning afternoon evening night noon midnight dawn dusk """.split()) CERTAINTY_MARKERS = frozenset(""" definitely definitively certainly obviously clearly undoubtedly unquestionably absolutely surely plainly evidently undeniably unmistakably decidedly categorically conclusively decisively resolutely proves proven prove proved proof impossible must always never guaranteed """.split()) CONCLUSION_MARKERS = ( "therefore", "thus", "hence", "consequently", "accordingly", "it follows that", "we conclude", "we can conclude", "this shows", "this demonstrates", "this proves", "in conclusion", "as a result", ) CONTRAST_MARKERS = ( "however", "but", "yet", "although", "though", "nevertheless", "nonetheless", "conversely", "on the contrary", "in contrast", "on the other hand", "by contrast", "notwithstanding", "despite this", "even so", ) EVIDENCE_MARKERS = ( "according to", "studies show", "studies suggest", "research shows", "research suggests", "data show", "data suggest", "evidence indicates", "we measured", "we observed", "we found", "for example", "for instance", "specifically", "namely", "for one", "in fact", "as measured", ) QUESTION_RAISERS = ( "why", "how", "whether", "what caused", "the reason", "unclear", "unknown", "remains to be determined", "remains unclear", "puzzling", "mysterious", "unexplained", "open question", "puzzle", ) ANSWER_MARKERS = ( "because", "since", "as a result", "due to", "explained by", "the reason is", "this explains", "the cause", "attributable to", "results from", "arises from", "the mechanism is", ) DIRECTION_GROUPS = { "up": frozenset(""" increase increases increased increasing rise rises rose risen grow grows grew grown growth expand expands expanded expansion raise raises raised raising improve improves improved improving gain gains gained gaining positive higher highest more most upward up climb climbs climbed climbing """.split()), "down": frozenset(""" decrease decreases decreased decreasing fall falls fell fallen shrink shrinks shrank shrunk contract contracts contracted lower lowers lowered lowering reduce reduces reduced reducing worsen worsens worsened worsening lose loses lost losing negative lower lowest less least downward down decline declines declined declining drop drops dropped dropping diminish """.split()), "cause": frozenset(""" cause causes caused causing produce produces produced producing create creates created creating induce induces induced trigger triggers triggered triggering generate """.split()), "prevent": frozenset(""" prevent prevents prevented preventing avoid avoids avoided block blocks blocked blocking inhibit inhibits inhibited prohibit prohibits prohibited prohibiting stop stops stopped """.split()), "support": frozenset(""" support supports supported supporting confirm confirms confirmed agree agrees agreed approve approves approved accept accepts accepted affirm affirms affirmed """.split()), "oppose": frozenset(""" oppose opposes opposed opposing deny denies denied contradict contradicts contradicted disagree disagrees disagreed reject rejects rejected refuse refuses refused """.split()), } GROUP_OPPOSITES = frozenset([ ("up", "down"), ("down", "up"), ("cause", "prevent"), ("prevent", "cause"), ("support", "oppose"), ("oppose", "support"), ]) DISCOURSE_MARKERS = frozenset(""" therefore thus hence consequently accordingly however but yet although though nevertheless nonetheless conversely meanwhile similarly moreover furthermore additionally """.split()) # ============================================================================ # Regexes # ============================================================================ _WORD_RE = re.compile(r"[A-Za-z][A-Za-z'\-]*") _SENT_SPLIT_RE = re.compile(r"(?<=[.!?])\s+(?=[A-Z\"'(])") _NUMBER_RE = re.compile(r"\b\d+(?:[.,]\d+)*\b") _STANDALONE_DEMON_RE = re.compile( r'^\s*(this|that|these|those)\s*[.,!?;:]*\s*$', re.IGNORECASE ) # ============================================================================ # Tokenization helpers # ============================================================================ def _stem(w: str) -> str: w = w.lower() if len(w) <= 4: return w for suffix in ("ingly", "edly", "ing", "ed", "ly", "es", "s"): if w.endswith(suffix) and len(w) - len(suffix) >= 3: base = w[: -len(suffix)] if len(base) >= 2 and base[-1] == base[-2] and base[-1] not in "aeiou": base = base[:-1] return base return w def _words(text: str) -> List[str]: return _WORD_RE.findall(text) def _content_words(text: str) -> List[str]: return [ w.lower() for w in _words(text) if w.lower() not in COMMON_WORDS and w.lower() not in DISCOURSE_MARKERS and len(w) >= 3 ] def _sentences(text: str) -> List[str]: return [s.strip() for s in _SENT_SPLIT_RE.split(text) if s.strip()] def _is_common(w: str) -> bool: lw = w.lower() if lw in COMMON_WORDS: return True return _stem(lw) in COMMON_WORDS def _is_rare(w: str) -> bool: return len(w) >= 7 and not _is_common(w) def _has_conclusion_marker(text: str) -> bool: low = text.lower().strip() for m in CONCLUSION_MARKERS: if low.startswith(m): return True if f" {m} " in f" {low} ": return True return False def _has_contrast_marker(text: str) -> Tuple[bool, str]: low = text.lower().strip() for m in CONTRAST_MARKERS: if low.startswith(m): return True, m if f" {m} " in f" {low} ": return True, m return False, "" def _has_evidence_marker(text: str) -> bool: low = text.lower() return any(m in low for m in EVIDENCE_MARKERS) # ============================================================================ # Config # ============================================================================ @dataclass class HVReaderConfig: # Pace (hv-tempo weights) baseline_wpm: float = 220.0 w_sentence_len_excess: float = 0.40 w_clause_rate: float = 0.08 w_rare_rate: float = 0.70 w_abstract_rate: float = 0.40 w_digit_rate: float = 0.30 w_negation_rate: float = 0.40 w_hedge_rate: float = 0.40 w_conditional_rate: float = 0.60 w_passive_rate: float = 0.30 w_list_bonus: float = -0.50 max_log_slowdown: float = 1.5 # Memory (hv-forget) memory_target_days: float = 7.0 memory_stability_base: float = 1.0 memory_stability_density: float = 5.0 memory_stability_rare: float = 3.0 memory_stability_salience: float = 2.0 # Fold (hv-fold) fold_demon_weight: float = 2.0 fold_definite_only: float = 0.5 fold_reentry: float = 1.0 fold_forward: float = 0.5 fold_resolved: float = 0.1 exempt_first_sentence: bool = False # Slip (hv-slip) monotone_window: int = 3 slip_monotone_weight: float = 0.4 slip_repetition_weight: float = 0.3 slip_absence_weight: float = 0.3 # Hunger (hv-hunger) hunger_question_weight: float = 1.0 hunger_answer_weight: float = 1.0 # Wall (hv-wall) wall_threshold: float = 0.35 certainty_scale: float = 6.0 hedge_scale: float = 5.0 contradiction_scale: float = 0.4 specificity_scale: float = 5.0 gap_confidence_jump_weight: float = 0.5 contrast_bonus: float = 0.15 contradiction_min: float = 0.30 overclaim_min: float = 0.30 gap_min: float = 0.25 # Reporting min_sentence_words: int = 3 top_reasons: int = 3 version: str = "0.1.1" # ============================================================================ # Report dataclasses # ============================================================================ @dataclass class SpanProfile: id: int text: str span: Tuple[int, int] n_words: int wpm: float slowdown: float density: float stability_days: float retention_week: float fold_load: float slip: float hunger_delta: float hunger: float wall: float wall_type: str reasons: List[str] = field(default_factory=list) def to_dict(self) -> dict: return { "id": self.id, "text": self.text, "span": list(self.span), "n_words": self.n_words, "wpm": round(self.wpm, 2), "slowdown": round(self.slowdown, 3), "density": round(self.density, 4), "stability_days": round(self.stability_days, 3), "retention_week": round(self.retention_week, 4), "fold_load": round(self.fold_load, 3), "slip": round(self.slip, 3), "hunger_delta": round(self.hunger_delta, 3), "hunger": round(self.hunger, 3), "wall": round(self.wall, 4), "wall_type": self.wall_type, "reasons": list(self.reasons), } @dataclass class ReadingProfile: text: str n_sentences: int n_words: int pace: float memory: float passes: float slip: float wall: float ttu_s: float mean_wpm: float hunger_final: float spans: List[SpanProfile] slowest_span: Optional[SpanProfile] wall_span: Optional[SpanProfile] summary: str def to_dict(self) -> dict: return { "text": self.text, "n_sentences": self.n_sentences, "n_words": self.n_words, "profile": { "pace": round(self.pace, 4), "memory": round(self.memory, 4), "passes": round(self.passes, 4), "slip": round(self.slip, 4), "wall": round(self.wall, 4), }, "ttu_s": round(self.ttu_s, 2), "mean_wpm": round(self.mean_wpm, 2), "hunger_final": round(self.hunger_final, 3), "slowest_span_id": self.slowest_span.id if self.slowest_span else None, "wall_span_id": self.wall_span.id if self.wall_span else None, "summary": self.summary, "spans": [s.to_dict() for s in self.spans], } # ============================================================================ # Pace (hv-tempo) # ============================================================================ def _count_list_markers(text: str) -> int: bullets = len(re.findall(r"(?:^|\n)\s*[-*•]\s+\S", text)) numbered = re.findall(r"(?:^|[\s;:.])\d+[.)]\s+\S", text) n_numbered = len(numbered) if len(numbered) >= 2 else 0 return max(bullets, n_numbered) def _pace_features(sentence: str) -> Dict[str, float]: words = _words(sentence) n_words = len(words) if n_words == 0: return {k: 0.0 for k in [ "n_words", "mean_sentence_len", "sentence_len_signed", "clauses_per_sentence", "rare_rate", "abstract_rate", "digit_rate", "negation_rate", "hedge_rate", "conditional_rate", "passive_rate", "list_rate", ]} n_sentences = max(1, len([s for s in _SENT_SPLIT_RE.split(sentence) if s.strip()])) mean_len = n_words / n_sentences signed_len = (mean_len - 15.0) / 10.0 n_clauses = ( sentence.count(",") + sentence.count(";") + sentence.count(":") + sum(1 for w in words if w.lower() in SUBORDINATORS) ) clauses_per_sentence = n_clauses / n_sentences n_rare = sum(1 for w in words if len(w) >= 7 and not _is_common(w)) rare_rate = n_rare / n_words n_abstract = sum( 1 for w in words if len(w) > 5 and w.lower().endswith(ABSTRACT_SUFFIXES) ) abstract_rate = n_abstract / n_words digit_rate = len(_NUMBER_RE.findall(sentence)) / n_words negation_rate = sum(1 for w in words if w.lower() in NEGATIONS) / n_words hedge_rate = sum(1 for w in words if w.lower() in HEDGES) / n_words conditional_rate = sum(1 for w in words if w.lower() in CONDITIONALS) / n_words n_passive = 0 for i, w in enumerate(words): if w.lower() in BE_FORMS and i + 1 < len(words): nxt = words[i + 1].lower() if (nxt.endswith("ed") and len(nxt) > 3) or nxt in ( "gone", "seen", "written", "taken", "made", "known", "found", "given", "held", "sent", "left", "kept", ): n_passive += 1 passive_rate = n_passive / n_sentences n_list = _count_list_markers(sentence) list_rate = n_list / n_sentences return { "n_words": float(n_words), "mean_sentence_len": mean_len, "sentence_len_signed": signed_len, "clauses_per_sentence": clauses_per_sentence, "rare_rate": rare_rate, "abstract_rate": abstract_rate, "digit_rate": digit_rate, "negation_rate": negation_rate, "hedge_rate": hedge_rate, "conditional_rate": conditional_rate, "passive_rate": passive_rate, "list_rate": list_rate, } def _pace_slowdown( f: Dict[str, float], cfg: HVReaderConfig ) -> Tuple[float, Dict[str, float]]: contrib = { "sentence_length": cfg.w_sentence_len_excess * f["sentence_len_signed"], "clause_density": cfg.w_clause_rate * f["clauses_per_sentence"], "rare_words": cfg.w_rare_rate * f["rare_rate"], "abstract_terms": cfg.w_abstract_rate * f["abstract_rate"], "numerals": cfg.w_digit_rate * f["digit_rate"], "negation": cfg.w_negation_rate * f["negation_rate"], "hedging": cfg.w_hedge_rate * f["hedge_rate"], "conditionals": cfg.w_conditional_rate * f["conditional_rate"], "passive_voice": cfg.w_passive_rate * f["passive_rate"], "list_structure": cfg.w_list_bonus * f["list_rate"], } log_slowdown = sum(contrib.values()) log_slowdown = max( -cfg.max_log_slowdown, min(cfg.max_log_slowdown, log_slowdown) ) return math.exp(log_slowdown), contrib # ============================================================================ # Memory (hv-forget) # ============================================================================ def _memory_stability( sentence: str, density: float, cfg: HVReaderConfig ) -> Tuple[float, float]: words = _words(sentence) n_content = max(1, len(_content_words(sentence))) n_rare = sum(1 for w in words if _is_rare(w)) rare_ratio = n_rare / n_content n_digits = len(_NUMBER_RE.findall(sentence)) n_proper = sum( 1 for i, w in enumerate(words) if i > 0 and w[0].isupper() and w.lower() not in COMMON_WORDS and w.lower() not in DISCOURSE_MARKERS and len(w) >= 3 ) salience = min(1.0, (n_digits + n_proper) / n_content) stability = ( cfg.memory_stability_base + cfg.memory_stability_density * density + cfg.memory_stability_rare * rare_ratio + cfg.memory_stability_salience * salience ) return stability, salience def _memory_retention(stability: float, days: float) -> float: if stability <= 0: return 0.0 return math.exp(-days / stability) # ============================================================================ # Density # ============================================================================ def _density_per_sentence(sentences: List[str]) -> List[float]: seen: Set[str] = set() out: List[float] = [] for i, s in enumerate(sentences): content = [_stem(w) for w in _content_words(s)] if not content: out.append(0.0) continue if i == 0: out.append(1.0) else: novel = [w for w in content if w not in seen] out.append(len(novel) / len(content)) seen.update(content) return out # ============================================================================ # Fold (hv-fold) # ============================================================================ def _definite_nps(sentence: str) -> List[Tuple[str, str]]: out: List[Tuple[str, str]] = [] words = list(_WORD_RE.finditer(sentence)) for idx, w in enumerate(words): if w.group(0).lower() not in ARTICLES: continue tail = words[idx + 1: idx + 3] content = [ tw.group(0).lower() for tw in tail if tw.group(0).lower() not in COMMON_WORDS and len(tw.group(0)) >= 3 ] if not content: continue if any(c in UNIQUE_REFERENTS for c in content): continue head = content[-1] end = tail[-1].end() if tail else w.end() out.append((sentence[w.start():end], head)) return out def _fold_loads(sentences: List[str], cfg: HVReaderConfig) -> List[float]: n = len(sentences) word_sentences: Dict[str, Set[int]] = defaultdict(set) for i, s in enumerate(sentences): for w in _words(s): lw = w.lower() if lw not in COMMON_WORDS and len(lw) >= 3: word_sentences[_stem(lw)].add(i) loads: List[float] = [] for i, s in enumerate(sentences): load = 0.0 if _STANDALONE_DEMON_RE.match(s.strip()): load += cfg.fold_demon_weight for _, head in _definite_nps(s): stem = _stem(head) occ = word_sentences.get(stem, set()) prior = [j for j in occ if j < i] later = [j for j in occ if j > i] if prior: load += cfg.fold_resolved elif later: delay = min(later) - i load += cfg.fold_forward * delay else: if i == 0 and cfg.exempt_first_sentence: pass else: load += cfg.fold_definite_only if any(w.lower() in REENTRY_MARKERS for w in _words(s)): load += cfg.fold_reentry loads.append(load) return loads # ============================================================================ # Slip (hv-slip) # ============================================================================ def _slip_per_sentence( sentences: List[str], cfg: HVReaderConfig ) -> List[float]: n = len(sentences) if n == 0: return [] lengths = [len(_words(s)) for s in sentences] slips: List[float] = [] prev_content: Set[str] = set() for i, s in enumerate(sentences): lo = max(0, i - cfg.monotone_window) hi = min(n, i + cfg.monotone_window + 1) window = lengths[lo:hi] if len(window) > 1: mean = sum(window) / len(window) var = sum((x - mean) ** 2 for x in window) / len(window) std = math.sqrt(var) monotone = max(0.0, 1.0 - std / 8.0) else: monotone = 0.0 content = {_stem(w) for w in _content_words(s)} if prev_content and content: overlap = len(content & prev_content) / max(1, len(content)) else: overlap = 0.0 prev_content = content n_digits = len(_NUMBER_RE.findall(s)) words_s = _words(s) n_proper = sum( 1 for j, w in enumerate(words_s) if j > 0 and w[0].isupper() and w.lower() not in COMMON_WORDS and len(w) >= 3 ) absence = 1.0 if (n_digits + n_proper) == 0 else 0.0 slip = ( cfg.slip_monotone_weight * monotone + cfg.slip_repetition_weight * overlap + cfg.slip_absence_weight * absence ) slips.append(max(0.0, min(1.0, slip))) return slips # ============================================================================ # Hunger (hv-hunger) # ============================================================================ def _hunger_deltas(sentences: List[str], cfg: HVReaderConfig) -> List[float]: deltas: List[float] = [] for s in sentences: low = s.lower() raised = sum(1 for m in QUESTION_RAISERS if m in low) answered = sum(1 for m in ANSWER_MARKERS if m in low) delta = ( cfg.hunger_question_weight * raised - cfg.hunger_answer_weight * answered ) deltas.append(delta) return deltas # ============================================================================ # Wall (hv-wall) # ============================================================================ def _direction_of(word: str) -> List[str]: w = word.lower() sw = _stem(w) out = [] for g, words in DIRECTION_GROUPS.items(): if w in words or sw in words: out.append(g) return out def _closest_topic(words: List[str], idx: int) -> Optional[int]: best = None best_key: Tuple[int, int] = (10, 1) for j in range(max(0, idx - 3), min(len(words), idx + 4)): if j == idx: continue cand = words[j] if cand in COMMON_WORDS or len(cand) < 4: continue if _direction_of(cand): continue if cand in DISCOURSE_MARKERS: continue after = 0 if j > idx else 1 key = (abs(j - idx), after) if key < best_key: best_key = key best = j return best def _direction_pairs(sentence: str) -> List[Tuple[str, str, str]]: words = [w.lower() for w in _words(sentence)] out: List[Tuple[str, str, str]] = [] for i, w in enumerate(words): groups = _direction_of(w) if not groups: continue j = _closest_topic(words, i) if j is None: continue topic = _stem(words[j]) for g in groups: out.append((topic, g, w)) return out def _wall_confidence( sentence: str, cfg: HVReaderConfig ) -> Tuple[float, int, int]: words = [w.lower() for w in _words(sentence)] n = max(1, len(words)) cert = sum(1 for w in words if w in CERTAINTY_MARKERS) hedg = sum(1 for w in words if w in HEDGES) conf = min(1.0, cfg.certainty_scale * cert / n) hedge = min(1.0, cfg.hedge_scale * hedg / n) return max(0.0, conf - 0.5 * hedge), cert, hedg def _wall_evidence( sentence: str, cfg: HVReaderConfig ) -> Tuple[float, float, bool]: words = _words(sentence) content = _content_words(sentence) n_content = max(1, len(content)) n_digits = len(_NUMBER_RE.findall(sentence)) n_proper = sum( 1 for i, w in enumerate(words) if i > 0 and w[0].isupper() and w.lower() not in COMMON_WORDS and w.lower() not in DISCOURSE_MARKERS and len(w) >= 3 ) n_rare = sum(1 for w in content if len(w) >= 8) spec = ( 0.5 * (n_digits / n_content) + 0.3 * (n_proper / n_content) + 0.2 * (n_rare / n_content) ) spec = min(1.0, cfg.specificity_scale * spec) attribution = _has_evidence_marker(sentence) evidence = 0.6 * spec + 0.4 * (1.0 if attribution else 0.0) return evidence, spec, attribution def _wall_per_sentence( sentences: List[str], cfg: HVReaderConfig ) -> List[Tuple[float, str, List[str]]]: n = len(sentences) out: List[Tuple[float, str, List[str]]] = [] prior_sentences: List[str] = [] prior_conf: List[float] = [] prior_evid: List[float] = [] for i, s in enumerate(sentences): confidence, cert_count, _ = _wall_confidence(s, cfg) evidence, _spec, attribution = _wall_evidence(s, cfg) n_words = len(_words(s)) overclaim = ( max(0.0, confidence - evidence) if n_words >= cfg.min_sentence_words else 0.0 ) # Contradiction via direction-group conflicts. contradiction = 0.0 contra_idx: Optional[int] = None pairs: List[Tuple[str, str]] = [] if prior_sentences: curr_pairs = _direction_pairs(s) prior_dirs: Dict[str, List[Tuple[int, str, str]]] = {} for j, p in enumerate(prior_sentences): for topic, g, src in _direction_pairs(p): prior_dirs.setdefault(topic, []).append((j, g, src)) for topic, curr_g, curr_src in curr_pairs: for prior_idx, prior_g, prior_src in prior_dirs.get(topic, []): if (prior_g, curr_g) in GROUP_OPPOSITES: pairs.append(( f"{topic}:{prior_src}↔{curr_src}", f"({prior_g} vs {curr_g})", )) contra_idx = prior_idx if pairs: contradiction = min( 1.0, cfg.contradiction_scale * math.sqrt(len(pairs)) ) # Gap. gap = 0.0 has_conclusion = _has_conclusion_marker(s) if has_conclusion and prior_evid: prior_ev_mean = sum(prior_evid) / len(prior_evid) ev_deficit = max(0.0, confidence - prior_ev_mean) prior_conf_mean = ( sum(prior_conf) / len(prior_conf) if prior_conf else 0.0 ) conf_jump = max(0.0, confidence - prior_conf_mean) gap = ev_deficit + cfg.gap_confidence_jump_weight * conf_jump # Priority ordering. if contradiction >= cfg.contradiction_min: wall, wtype = contradiction, "contradiction" elif overclaim >= cfg.overclaim_min: wall, wtype = overclaim, "overclaim" elif gap >= cfg.gap_min: wall, wtype = gap, "gap" else: best = max(overclaim, contradiction, gap) if best <= 0.0: wall, wtype = 0.0, "neutral" else: wall = best if best == contradiction: wtype = "contradiction" elif best == overclaim: wtype = "overclaim" else: wtype = "gap" wall = min(1.0, wall) has_contrast, _contrast_word = _has_contrast_marker(s) if has_contrast and wall > 0.05: wall = min(1.0, wall + cfg.contrast_bonus) reasons: List[str] = [] if wtype == "contradiction" and pairs: reasons.append(f"conflict with sentence {contra_idx}: {pairs[0][0]}") elif wtype == "overclaim": if cert_count: matched = [ w.lower() for w in _words(s) if w.lower() in CERTAINTY_MARKERS ] reasons.append( f"certainty markers: " f"{', '.join(repr(m) for m in matched[:3])}" ) if not attribution: reasons.append("no attribution marker") elif wtype == "gap": if has_conclusion: reasons.append("conclusion marker with weak prior evidence") out.append((wall, wtype if wall > 0.0 else "neutral", reasons)) prior_sentences.append(s) prior_conf.append(confidence) prior_evid.append(evidence) return out # ============================================================================ # The model # ============================================================================ class HVReader: """Unified reading-experience model.""" def __init__(self, config: Optional[HVReaderConfig] = None): self.config = config or HVReaderConfig() self._obs = 0 def __repr__(self) -> str: return ( f"HVReader(baseline_wpm={self.config.baseline_wpm}, " f"wall_threshold={self.config.wall_threshold}, " f"version={self.config.version})" ) def analyze(self, text: str) -> ReadingProfile: if not text or not text.strip(): return self._empty(text) sentences = _sentences(text) n = len(sentences) if n == 0: return self._empty(text) spans: List[Tuple[int, int]] = [] cursor = 0 for s in sentences: i = text.find(s, cursor) if i < 0: i = cursor spans.append((i, i + len(s))) cursor = i + len(s) pace_feats = [_pace_features(s) for s in sentences] slowdowns: List[float] = [] wpms: List[float] = [] for f in pace_feats: sd, _ = _pace_slowdown(f, self.config) slowdowns.append(sd) wpms.append( self.config.baseline_wpm / sd if sd > 0 else self.config.baseline_wpm ) densities = _density_per_sentence(sentences) fold_loads = _fold_loads(sentences, self.config) slips = _slip_per_sentence(sentences, self.config) hunger_deltas = _hunger_deltas(sentences, self.config) hunger_cumulative: List[float] = [] h = 0.0 for d in hunger_deltas: h += d hunger_cumulative.append(h) stabilities: List[float] = [] retentions: List[float] = [] for s, dens in zip(sentences, densities): stab, _ = _memory_stability(s, dens, self.config) stabilities.append(stab) retentions.append( _memory_retention(stab, self.config.memory_target_days) ) wall_per = _wall_per_sentence(sentences, self.config) span_profiles: List[SpanProfile] = [] for i in range(n): wall_score, wall_type, reasons = wall_per[i] n_words_span = int(pace_feats[i]["n_words"]) span_profiles.append(SpanProfile( id=i, text=sentences[i], span=spans[i], n_words=n_words_span, wpm=wpms[i], slowdown=slowdowns[i], density=densities[i], stability_days=stabilities[i], retention_week=retentions[i], fold_load=fold_loads[i], slip=slips[i], hunger_delta=hunger_deltas[i], hunger=hunger_cumulative[i], wall=wall_score, wall_type=wall_type, reasons=reasons[: self.config.top_reasons], )) n_words_total = sum(s.n_words for s in span_profiles) mean_wpm = ( sum(s.wpm * s.n_words for s in span_profiles) / max(1, n_words_total) ) pace = max(0.0, min(1.0, mean_wpm / 400.0)) memory = sum(retentions) / n if n else 0.0 total_load = sum(fold_loads) passes = 1.0 + total_load / max(1, n) slip = sum(slips) / n if n else 0.0 wall_scores = [s.wall for s in span_profiles] wall_max = max(wall_scores) if wall_scores else 0.0 wall = wall_max wps = mean_wpm / 60.0 ttu_s = n_words_total / wps if wps > 0 else 0.0 hunger_final = hunger_cumulative[-1] if hunger_cumulative else 0.0 slowest = ( min(span_profiles, key=lambda s: s.wpm) if span_profiles else None ) wall_span: Optional[SpanProfile] = None if wall > 0.15: wall_span = max(span_profiles, key=lambda s: s.wall) summary = self._summary( span_profiles, pace, memory, passes, slip, wall ) self._obs += 1 return ReadingProfile( text=text, n_sentences=n, n_words=n_words_total, pace=pace, memory=memory, passes=passes, slip=slip, wall=wall, ttu_s=ttu_s, mean_wpm=mean_wpm, hunger_final=hunger_final, spans=span_profiles, slowest_span=slowest, wall_span=wall_span, summary=summary, ) def _empty(self, text: str) -> ReadingProfile: return ReadingProfile( text=text, n_sentences=0, n_words=0, pace=0.0, memory=0.0, passes=1.0, slip=0.0, wall=0.0, ttu_s=0.0, mean_wpm=0.0, hunger_final=0.0, spans=[], slowest_span=None, wall_span=None, summary="Empty text.", ) @staticmethod def _summary( spans: List[SpanProfile], pace: float, memory: float, passes: float, slip: float, wall: float, ) -> str: if not spans: return "Empty text." parts = [] parts.append(f"reads at {pace * 400:.0f} WPM (pace {pace:.2f})") parts.append(f"memory after a week: {memory * 100:.0f}%") parts.append(f"requires {passes:.2f} passes") parts.append(f"slip probability: {slip:.2f}") if wall > 0.15: wall_span = max(spans, key=lambda s: s.wall) parts.append( f"wall at sentence {wall_span.id} " f"({wall_span.wall_type}, {wall:.2f})" ) else: parts.append("no wall") return ". ".join(parts).capitalize() + "." def render(self, profile: ReadingProfile) -> str: lines: List[str] = [] bar = "=" * 72 lines.append(bar) lines.append("hv-reader — the reading experience") lines.append(bar) lines.append("") lines.append(f" text : {profile.n_sentences} sentences, " f"{profile.n_words} words") lines.append("") lines.append(" READING PROFILE") lines.append(" " + "-" * 68) lines.append(f" pace {profile.pace:>6.3f} " f"({profile.mean_wpm:.0f} WPM)") lines.append(f" memory {profile.memory:>6.3f} " f"(fraction surviving 1 week)") lines.append(f" passes {profile.passes:>6.3f} " f"(reads needed)") lines.append(f" slip {profile.slip:>6.3f} " f"(attention-lapse probability)") lines.append(f" wall {profile.wall:>6.3f} " f"(reader-refusal probability)") lines.append("") lines.append(f" ttu : {profile.ttu_s:.1f} s " f"(total reading time)") lines.append(f" hunger : {profile.hunger_final:+.2f} " f"(unresolved questions)") lines.append("") if not profile.spans: lines.append(" (no content)") return "\n".join(lines) lines.append(" PER-SENTENCE") lines.append(" " + "-" * 68) lines.append( f" {'id':>3} {'wpm':>5} {'dens':>5} {'ret':>5} " f"{'fold':>5} {'slip':>5} {'wall':>5} type" ) for s in profile.spans: marker = ( "*" if (profile.wall_span and s.id == profile.wall_span.id) else " " ) lines.append( f" {marker}{s.id:>2} {s.wpm:>5.0f} {s.density:>5.2f} " f"{s.retention_week:>5.2f} {s.fold_load:>5.2f} " f"{s.slip:>5.2f} {s.wall:>5.2f} {s.wall_type}" ) lines.append("") if profile.slowest_span: s = profile.slowest_span lines.append(" SLOWEST SPAN") lines.append(" " + "-" * 68) lines.append(f" [{s.id}] {s.wpm:.0f} WPM " f"(slowdown {s.slowdown:.2f}x)") lines.append(f" \"{self._shorten(s.text, 60)}\"") lines.append("") if profile.wall_span: s = profile.wall_span lines.append(" WALL SPAN") lines.append(" " + "-" * 68) lines.append(f" [{s.id}] {s.wall_type} " f"(score {s.wall:.3f})") lines.append(f" \"{self._shorten(s.text, 60)}\"") for r in s.reasons: lines.append(f" - {r}") lines.append("") lines.append(" SUMMARY") lines.append(" " + "-" * 68) lines.append(f" {profile.summary}") lines.append("") return "\n".join(lines) @staticmethod def _shorten(s: str, n: int) -> str: s = s.strip().replace("\n", " ") if len(s) <= n: return s return s[: n - 1].rsplit(" ", 1)[0] + "…" def save_pretrained(self, save_dir: str) -> None: os.makedirs(save_dir, exist_ok=True) payload = { "config": asdict(self.config), "observations": self._obs, } with open(os.path.join(save_dir, "config.json"), "w") as f: json.dump(payload, f, indent=2) @classmethod def from_pretrained(cls, save_dir: str) -> "HVReader": with open(os.path.join(save_dir, "config.json"), "r") as f: payload = json.load(f) cfg_dict = payload.get("config", {}) known = {f.name for f in HVReaderConfig.__dataclass_fields__.values()} cfg_dict = {k: v for k, v in cfg_dict.items() if k in known} cfg = HVReaderConfig(**cfg_dict) obj = cls(config=cfg) obj._obs = int(payload.get("observations", 0)) return obj # ============================================================================ # Demo # ============================================================================ SAMPLE_FICTION = ( "The old man walked slowly to the boat. He stopped, looked at the " "water, and then continued. The sea was quiet that morning. He " "pushed the boat into the water and climbed in. The oars were cold " "in his hands. He rowed out past the harbor and into the open sea." ) SAMPLE_ACADEMIC = ( "A black hole is a region of spacetime where gravity is so strong " "that nothing — no particles or even electromagnetic radiation such " "as light — can escape from it. The theory of general relativity " "predicts that a sufficiently compact mass can deform spacetime to " "form a black hole. The boundary of the region from which no escape " "is possible is called the event horizon. Although the event horizon " "has profound effects on the fate of an object that crosses it, it " "has no locally detectable features. A black hole acts as a perfect " "black body, and moreover, it emits Hawking radiation." ) SAMPLE_OVERCLAIM = ( "The data suggests some correlation between the policy and the outcome. " "Results appear to indicate a modest effect in some subpopulations. " "The mechanism remains unclear, and further work is needed to establish " "causality. " "Therefore, the policy definitively causes the outcome in all cases, " "and this is unquestionably proven by the evidence." ) SAMPLE_MONOTONE = ( "The system processes the input. The system processes the data. " "The system processes the output. The system processes the result. " "The system processes the value. The system processes the record." ) def _demo(output_dir: str = "./hv_reader_output") -> None: os.makedirs(output_dir, exist_ok=True) m = HVReader() samples = [ ("Fiction", SAMPLE_FICTION), ("Academic", SAMPLE_ACADEMIC), ("Overclaim", SAMPLE_OVERCLAIM), ("Monotone", SAMPLE_MONOTONE), ] for name, text in samples: print() print("#" * 72) print(f"# {name}") print("#" * 72) print(m.render(m.analyze(text))) print() print("=" * 72) print("Summary across samples") print("=" * 72) print( f" {'sample':<12} {'words':>6} {'pace':>6} {'mem':>6} " f"{'pass':>5} {'slip':>6} {'wall':>5} {'ttu':>6}" ) print(" " + "-" * 68) for name, text in samples: r = m.analyze(text) print( f" {name:<12} {r.n_words:>6} {r.pace:>6.3f} " f"{r.memory:>6.3f} {r.passes:>5.2f} " f"{r.slip:>6.3f} {r.wall:>5.2f} {r.ttu_s:>5.1f}s" ) print() print("=" * 72) print("Save / load round trip") print("=" * 72) save_path = os.path.join(output_dir, "hv_reader_model") m.save_pretrained(save_path) m2 = HVReader.from_pretrained(save_path) print(f" saved to : {save_path}") print(f" reloaded : {m2!r}") a = m.analyze(SAMPLE_OVERCLAIM) b = m2.analyze(SAMPLE_OVERCLAIM) print(f" pace : {a.pace:.4f}") print(f" wall : {a.wall:.4f}") print(f" identical : " f"{abs(a.pace - b.pace) < 1e-9 and abs(a.wall - b.wall) < 1e-9}") print() # ============================================================================ # CLI # ============================================================================ def _cli() -> None: p = argparse.ArgumentParser( description="hv-reader: the reading experience in one call." ) p.add_argument("--text", type=str, default="", help="text to analyze (or '-' to read stdin)") p.add_argument("--json", action="store_true", help="output JSON instead of a rendered report") p.add_argument("--profile", action="store_true", help="print only the five-axis profile") p.add_argument("--save-to", type=str, default="", help="save the model to this directory") p.add_argument("--outdir", type=str, default="./hv_reader_output") args = p.parse_args() text = sys.stdin.read() if args.text == "-" else args.text m = HVReader() if args.save_to: m.save_pretrained(args.save_to) print(f"saved to {args.save_to}", file=sys.stderr) if not text: _demo(args.outdir) return profile = m.analyze(text) if args.profile: print(f"pace {profile.pace:.3f}") print(f"memory {profile.memory:.3f}") print(f"passes {profile.passes:.3f}") print(f"slip {profile.slip:.3f}") print(f"wall {profile.wall:.3f}") return if args.json: print(json.dumps(profile.to_dict(), indent=2, ensure_ascii=False)) else: print(m.render(profile)) if __name__ == "__main__": _cli()