Blame for validation/paper2/RELATEDNESS.json

{
 "what": "the three relatedness measures for paper 2's test 6, fixed before any reading (PAPER-3-Options 2.11 g)",
 "measures": {
  "family": "the inventory's `a_family`; related when equal and not the `builtin` catch-all of 143 names",
  "tokens": "the inventory's `tokens` list -- the surface tokens at the read site and in the cores; related when the two lists intersect",
  "cooccurrence": "both constructs in one function of the calibration corpus (shards 1, 18, 36), observed against independence; related when the ratio exceeds 2 AND at least 20 functions hold both. Below that support the pair is not measured, not unrelated, and is counted separately."
 },
 "rules_fixed_21_Oct": {
  "reading_rule": {
   "statistic": "per measure, the ratio of the observed rate of related rises to the chance rate, with a bootstrap interval and the count of related pairs beside it",
   "bootstrap_unit": "the construct. The programme's intervals resample cores, but a departure's relatedness is fixed by the pair (construct left, construct risen), so every departure of one construct shares its relatedness profile and the construct is the unit that is exchangeable here. Named rather than substituted silently.",
   "branching": "ratio > 1 with the interval excluding 1 on tokens AND on family, and co-occurrence not below 1",
   "displacement": "rises above chance in any construct, not concentrated in related ones",
   "drift": "rises at chance",
   "no_measure_is_promoted_beyond_this": true,
   "pattern_outside_the_rule": "a lift clear of 1 on tokens but not on family is none of the three verdicts and is reported as outside them, not as one of them"
  },
  "diagnostics_fixed_21_Oct_before_pairs_2_and_3": {
   "displacement_number": "REDONE 22 Oct, fixed before computing. The first form was degenerate: the maximum rise over 168 constructs is almost surely positive, so 'did any construct rise' was answered yes by construction (observed 1.0000 at every pair, lift 2.00). Now a rise counts only if it exceeds **one twin unit** of that construct's entry on that neuron -- |E_twin1 - E_twin2| for the same construct, neuron and reference, the surface-noise unit the programme already uses. The same threshold is applied to the matched chance construct, so the comparison is like for like.",
   "tokens_split_by_where_the_token_sits": "the tokens lift computed twice -- for pairs sharing the token AT THE READ SITE, and for pairs sharing one only IN THE CORES -- each with its related-pair count. The first is the construct's own surface, the second its surroundings.",
   "cross_twin": "the tokens and family lifts with the departure taken from twin 1 and the rise from twin 2. A lift that survives the twins is not an artefact of one battery draw."
  },
  "shards": "all three calibration shards (1, 18, 36), fixed before coverage was seen",
  "min_support": 20,
  "related_requires": "ratio > 2 and support >= 20"
 },
 "median_companion_rate": null,
 "population": 169,
 "pairs": 14196,
 "coverage": {
  "family": {
   "pairs_total": 14196,
   "pairs_with_data": 14196,
   "coverage": 1.0
  },
  "tokens": {
   "pairs_total": 14196,
   "pairs_with_data": 14196,
   "coverage": 1.0
  },
  "cooccurrence": {
   "pairs_total": 14196,
   "pairs_with_data": 5596,
   "coverage": 0.39419554804170187
  }
 },
 "base_rate_of_related_pairs": {
  "family": 0.21034094111017188,
  "tokens": 0.0034516765285996054,
  "cooccurrence": 0.42208720514653325
 },
 "note": "co-occurrence covers only the constructs nu's companion table holds; pairs outside it are marked absent and never imputed",
 "provenance": {
  "inputs": [
   "inventory/inventory_v2.json",
   "calibration/companions.json"
  ],
  "reference": null,
  "battery_prompt_hash": "570b42db13f8fe5772b8d1a9ad7d55d6fa1279dae5948ad94e8363789bcbc1ad",
  "code_commit": "7ade464",
  "design": "9-Paper-1/2610-18Oct_PAPER-1-Design.md sha256 47f6a664\u2026"
 },
 "rules_fixed_22_Oct": {
  "b10_model": {
   "population": "ALL pairs of pretraining references (1,404), not neighbours -- neighbouring pairs confound training time with the gap, which is the fault being tested",
   "quantity": "turnover = 1 - |S(t) n S(t+D)| / K, the carrier set at the start against the carrier set at the end, product space, per construct",
   "model": "per start point t, turnover against the step gap D is fitted as a saturating curve y = A(t) * (1 - exp(-r(t) * D)); r(t) is the per-step rate at t and A(t) the asymptote. The slope at D -> 0 is A(t)*r(t), which is the per-step rate the paper's sentence is about.",
   "report": "r(t) and A(t)*r(t) as functions of t over the whole record, with core-bootstrap intervals over constructs; and whether the per-step rate is constant, rising or falling past 84 B tokens (iteration 20,000), which is where true 1,000-step pairs stop existing",
   "fixed_before_fitting": true
  },
  "b11_power": {
   "fixed_before_running": true,
   "noise": "whole ordered residual SERIES, not point-by-point. Each draw takes another construct's residual sequence and applies a random circular shift. Resampling point by point destroys the autocorrelation, and the rule is 'crosses theta and STAYS' -- with autocorrelated residuals an excursion below theta lasts longer, so staying is harder and i.i.d. draws would have overstated the power.",
   "shapes": {
    "step": "base before the date, plateau after -- the easiest shape to detect",
    "exponential": "base + (plateau - base) * (1 - exp(-(t - t0)/tau)) with tau one x1.5 spacing in tokens, the slowest approach the record could still resolve"
   },
   "dates": [
    84000000000.0,
    200000000000.0,
    500000000000.0,
    1000000000000.0,
    1500000000000.0
   ],
   "false_detection": "the same rule on the UNINJECTED series -- base level plus the same resampled noise, no settling -- reported beside the detection rates, so a detection rate is read against the rule's own false-positive rate",
   "reading": "a date whose detection rate is high means '0 of 169' excludes settling before it; a low rate means the null is uninformative there"
  }
 }
}