SRIC: external semantic admission

Download original file

This is the original TeX manuscript source. Use the paper’s PDF for typeset equations.

Read the typeset paper

1 % Options for packages loaded elsewhere
2 \PassOptionsToPackage{unicode}{hyperref}
3 \PassOptionsToPackage{hyphens}{url}
4 \PassOptionsToPackage{dvipsnames,svgnames,x11names}{xcolor}
5 %
6 \documentclass[
7   10pt,
8 ]{article}
9 \usepackage{amsmath,amssymb}
10 \usepackage{iftex}
11 \ifPDFTeX
12   \usepackage[T1]{fontenc}
13   \usepackage[utf8]{inputenc}
14   \usepackage{textcomp} % provide euro and other symbols
15 \else % if luatex or xetex
16   \usepackage{unicode-math} % this also loads fontspec
17   \defaultfontfeatures{Scale=MatchLowercase}
18   \defaultfontfeatures[\rmfamily]{Ligatures=TeX,Scale=1}
19 \fi
20 \usepackage{lmodern}
21 \ifPDFTeX\else
22   % xetex/luatex font selection
23 \fi
24 % Use upquote if available, for straight quotes in verbatim environments
25 \IfFileExists{upquote.sty}{\usepackage{upquote}}{}
26 \IfFileExists{microtype.sty}{% use microtype if available
27   \usepackage[]{microtype}
28   \UseMicrotypeSet[protrusion]{basicmath} % disable protrusion for tt fonts
29 }{}
30 \makeatletter
31 \@ifundefined{KOMAClassName}{% if non-KOMA class
32   \IfFileExists{parskip.sty}{%
33     \usepackage{parskip}
34   }{% else
35     \setlength{\parindent}{0pt}
36     \setlength{\parskip}{6pt plus 2pt minus 1pt}}
37 }{% if KOMA class
38   \KOMAoptions{parskip=half}}
39 \makeatother
40 \usepackage{xcolor}
41 \usepackage[margin=0.78in]{geometry}
42 \usepackage{color}
43 \usepackage{fancyvrb}
44 \newcommand{\VerbBar}{|}
45 \newcommand{\VERB}{\Verb[commandchars=\\\{\}]}
46 \DefineVerbatimEnvironment{Highlighting}{Verbatim}{commandchars=\\\{\}}
47 % Add ',fontsize=\small' for more characters per line
48 \newenvironment{Shaded}{}{}
49 \newcommand{\AlertTok}[1]{\textcolor[rgb]{1.00,0.00,0.00}{\textbf{#1}}}
50 \newcommand{\AnnotationTok}[1]{\textcolor[rgb]{0.38,0.63,0.69}{\textbf{\textit{#1}}}}
51 \newcommand{\AttributeTok}[1]{\textcolor[rgb]{0.49,0.56,0.16}{#1}}
52 \newcommand{\BaseNTok}[1]{\textcolor[rgb]{0.25,0.63,0.44}{#1}}
53 \newcommand{\BuiltInTok}[1]{\textcolor[rgb]{0.00,0.50,0.00}{#1}}
54 \newcommand{\CharTok}[1]{\textcolor[rgb]{0.25,0.44,0.63}{#1}}
55 \newcommand{\CommentTok}[1]{\textcolor[rgb]{0.38,0.63,0.69}{\textit{#1}}}
56 \newcommand{\CommentVarTok}[1]{\textcolor[rgb]{0.38,0.63,0.69}{\textbf{\textit{#1}}}}
57 \newcommand{\ConstantTok}[1]{\textcolor[rgb]{0.53,0.00,0.00}{#1}}
58 \newcommand{\ControlFlowTok}[1]{\textcolor[rgb]{0.00,0.44,0.13}{\textbf{#1}}}
59 \newcommand{\DataTypeTok}[1]{\textcolor[rgb]{0.56,0.13,0.00}{#1}}
60 \newcommand{\DecValTok}[1]{\textcolor[rgb]{0.25,0.63,0.44}{#1}}
61 \newcommand{\DocumentationTok}[1]{\textcolor[rgb]{0.73,0.13,0.13}{\textit{#1}}}
62 \newcommand{\ErrorTok}[1]{\textcolor[rgb]{1.00,0.00,0.00}{\textbf{#1}}}
63 \newcommand{\ExtensionTok}[1]{#1}
64 \newcommand{\FloatTok}[1]{\textcolor[rgb]{0.25,0.63,0.44}{#1}}
65 \newcommand{\FunctionTok}[1]{\textcolor[rgb]{0.02,0.16,0.49}{#1}}
66 \newcommand{\ImportTok}[1]{\textcolor[rgb]{0.00,0.50,0.00}{\textbf{#1}}}
67 \newcommand{\InformationTok}[1]{\textcolor[rgb]{0.38,0.63,0.69}{\textbf{\textit{#1}}}}
68 \newcommand{\KeywordTok}[1]{\textcolor[rgb]{0.00,0.44,0.13}{\textbf{#1}}}
69 \newcommand{\NormalTok}[1]{#1}
70 \newcommand{\OperatorTok}[1]{\textcolor[rgb]{0.40,0.40,0.40}{#1}}
71 \newcommand{\OtherTok}[1]{\textcolor[rgb]{0.00,0.44,0.13}{#1}}
72 \newcommand{\PreprocessorTok}[1]{\textcolor[rgb]{0.74,0.48,0.00}{#1}}
73 \newcommand{\RegionMarkerTok}[1]{#1}
74 \newcommand{\SpecialCharTok}[1]{\textcolor[rgb]{0.25,0.44,0.63}{#1}}
75 \newcommand{\SpecialStringTok}[1]{\textcolor[rgb]{0.73,0.40,0.53}{#1}}
76 \newcommand{\StringTok}[1]{\textcolor[rgb]{0.25,0.44,0.63}{#1}}
77 \newcommand{\VariableTok}[1]{\textcolor[rgb]{0.10,0.09,0.49}{#1}}
78 \newcommand{\VerbatimStringTok}[1]{\textcolor[rgb]{0.25,0.44,0.63}{#1}}
79 \newcommand{\WarningTok}[1]{\textcolor[rgb]{0.38,0.63,0.69}{\textbf{\textit{#1}}}}
80 \usepackage{longtable,booktabs,array}
81 \usepackage{calc} % for calculating minipage widths
82 % Correct order of tables after \paragraph or \subparagraph
83 \usepackage{etoolbox}
84 \makeatletter
85 \patchcmd\longtable{\par}{\if@noskipsec\mbox{}\fi\par}{}{}
86 \makeatother
87 % Allow footnotes in longtable head/foot
88 \IfFileExists{footnotehyper.sty}{\usepackage{footnotehyper}}{\usepackage{footnote}}
89 \makesavenoteenv{longtable}
90 \setlength{\emergencystretch}{3em} % prevent overfull lines
91 \providecommand{\tightlist}{%
92   \setlength{\itemsep}{0pt}\setlength{\parskip}{0pt}}
93 \setcounter{secnumdepth}{5}
94 \usepackage{booktabs}
95 \usepackage{longtable}
96 \usepackage{array}
97 \usepackage{microtype}
98 \usepackage{enumitem}
99 \setlist{nosep,leftmargin=*}
100 \setlength{\parindent}{0pt}
101 \setlength{\parskip}{0.45em}
102 \ifLuaTeX
103   \usepackage{selnolig}  % disable illegal ligatures
104 \fi
105 \usepackage{bookmark}
106 \IfFileExists{xurl.sty}{\usepackage{xurl}}{} % add URL line breaks if available
107 \urlstyle{same}
108 \hypersetup{
109   pdftitle={SRIC Block 24},
110   pdfauthor={Komorebi P.},
111   colorlinks=true,
112   linkcolor={blue},
113   filecolor={Maroon},
114   citecolor={Blue},
115   urlcolor={blue},
116   pdfcreator={LaTeX via pandoc}}
117 
118 \title{SRIC Block 24}
119 \usepackage{etoolbox}
120 \makeatletter
121 \providecommand{\subtitle}[1]{% add subtitle to \maketitle
122   \apptocmd{\@title}{\par {\large #1 \par}}{}{}
123 }
124 \makeatother
125 \subtitle{External Latent-Reasoner Causal-Semantic Admission - Protocol
126 v0.3.0}
127 \author{Komorebi P.}
128 \date{September 3, 2026}
129 
130 \begin{document}
131 \maketitle
132 
133 {
134 \hypersetup{linkcolor=}
135 \setcounter{tocdepth}{3}
136 \tableofcontents
137 }
138 \section{Executive status}\label{executive-status}
139 
140 \textbf{Canonical status:}
141 \texttt{INFRASTRUCTURE\_VALIDATED\_EXTERNAL\_EXECUTION\_PENDING}
142 
143 Block 24 is the first SRIC package whose intended evidence source is an
144 independently authored and independently trained latent-reasoning
145 checkpoint. It translates the synthetic Neuralese program of Blocks
146 16-23 into a source-locked external-model admission protocol.
147 
148 The public checkpoint was \textbf{not executed in the build
149 environment}. The release therefore reports no empirical result about
150 Coconut, PrOntoQA, or any other third-party latent reasoner. It
151 establishes the external adapter, source locks, frozen chronology,
152 causal intervention battery, semantic escalation, post-seal witnesses,
153 engineered positive control, and reproducibility machinery required for
154 that run.
155 
156 A negative real result is fully admissible. The pipeline is designed to
157 classify failure rather than force a positive interpretation.
158 
159 \section{Research strategy}\label{research-strategy}
160 
161 The next major gap is external validity. Another package-local positive
162 control would contribute less evidence than exposing the architecture to
163 an independently authored recurrence and checkpoint. The design
164 therefore prioritizes source fidelity and causal clarity over model
165 size.
166 
167 The conservative defaults are:
168 
169 \begin{enumerate}
170 \def\labelenumi{\arabic{enumi}.}
171 \tightlist
172 \item
173   GPT-2 multimode Coconut on PrOntoQA is the primary target.
174 \item
175   PrOntoQA supports a \texttt{CONTENT\_SPECIFIC} claim, not a
176   different-algorithm claim.
177 \item
178   Explicit-CoT is revealed only after the semantic quotient is sealed.
179 \item
180   ProsQA is the first structure-sensitive \texttt{PROCEDURE\_SPECIFIC}
181   follow-up.
182 \item
183   Smoke, primary, and full profiles are run in that order without moving
184   scientific thresholds.
185 \end{enumerate}
186 
187 \section{Part I - Brainstorm and design
188 decisions}\label{part-i---brainstorm-and-design-decisions}
189 
190 \subsection{Why the next gap is
191 empirical}\label{why-the-next-gap-is-empirical}
192 
193 Blocks 16-23 progressively established a synthetic theory of latent
194 causal auditing, semantic charts, cross-architecture transport,
195 obstruction-guided repair, unpaired alignment, and endogenous semantic
196 quotient construction. Another synthetic positive control would add less
197 evidence than testing whether the same gates survive contact with a
198 checkpoint designed by an independent group.
199 
200 The central next question is therefore not whether another controlled
201 fixture can be made to pass. It is whether a source-locked external
202 latent reasoner lands in one of SRIC's predeclared epistemic classes
203 without altering the gates after seeing the result.
204 
205 \subsection{Candidate external models
206 considered}\label{candidate-external-models-considered}
207 
208 \subsubsection{Multimode GPT-2 Coconut on
209 PrOntoQA}\label{multimode-gpt-2-coconut-on-prontoqa}
210 
211 Advantages:
212 
213 \begin{itemize}
214 \tightlist
215 \item
216   small enough for an 8 GB laptop GPU at batch size one;
217 \item
218   public repository and raw checkpoint;
219 \item
220   latent, direct-answer, and explicit-CoT modes in one trained wrapper;
221 \item
222   recurrent latent vectors are reinjected at identifiable latent-token
223   positions;
224 \item
225   logic answers and proof chains are externally checkable;
226 \item
227   semantic twins and counterfactual query interventions are easy to
228   construct.
229 \end{itemize}
230 
231 Limitation:
232 
233 \begin{itemize}
234 \tightlist
235 \item
236   the task family is dominated by fixed-form unary chains. It supports a
237   clean first causal audit but not a strong claim that different latent
238   traces encode different reasoning algorithms.
239 \end{itemize}
240 
241 Decision: use this checkpoint as the primary source-faithful causal
242 admission target, with maximum causal status \texttt{CONTENT\_SPECIFIC}
243 before semantic escalation.
244 
245 \subsubsection{ProsQA multimode Coconut}\label{prosqa-multimode-coconut}
246 
247 Advantages:
248 
249 \begin{itemize}
250 \tightlist
251 \item
252   more varied proof structures;
253 \item
254   better setting for structure-near versus structure-far donor
255   contrasts;
256 \item
257   same external research codebase and model family.
258 \end{itemize}
259 
260 Limitations:
261 
262 \begin{itemize}
263 \tightlist
264 \item
265   answer parsing and proof-structure extraction require a separate
266   adapter path;
267 \item
268   harder to debug simultaneously with the first source-faithful
269   recurrence adapter.
270 \end{itemize}
271 
272 Decision: reserve as Tier 2. It is the first dataset on which
273 \texttt{PROCEDURE\_SPECIFIC} should be enabled.
274 
275 \subsubsection{SLPO Coconut}\label{slpo-coconut}
276 
277 Advantages:
278 
279 \begin{itemize}
280 \tightlist
281 \item
282   exposes adaptive stopping information;
283 \item
284   provides latent lengths and gate outputs.
285 \end{itemize}
286 
287 Limitations:
288 
289 \begin{itemize}
290 \tightlist
291 \item
292   additional runtime and checkpoint family;
293 \item
294   adaptive stopping introduces another mechanism before basic replay is
295   proven;
296 \item
297   full training experiments exceed the current 8 GB hardware target.
298 \end{itemize}
299 
300 Decision: use after the fixed-length causal audit is stable.
301 
302 \subsubsection{Hyper 3B}\label{hyper-3b}
303 
304 Advantages:
305 
306 \begin{itemize}
307 \tightlist
308 \item
309   directly recurrent latent reasoning;
310 \item
311   larger and more ambitious external stress.
312 \end{itemize}
313 
314 Limitations:
315 
316 \begin{itemize}
317 \tightlist
318 \item
319   at the hardware boundary;
320 \item
321   adapter and memory failures could be confused with theoretical
322   failure;
323 \item
324   current halt behavior is not a strong adaptive-depth positive control.
325 \end{itemize}
326 
327 Decision: later stress test, not the first empirical target.
328 
329 \subsection{Major methodological
330 choices}\label{major-methodological-choices}
331 
332 \subsubsection{1. Adapter fidelity before model
333 ambition}\label{adapter-fidelity-before-model-ambition}
334 
335 A large checkpoint is scientifically useless if the intervention adapter
336 does not reproduce the model's native recurrence. Block 24 therefore
337 gives exact replay and source-code parity priority over scale.
338 
339 \subsubsection{2. Teacher-forced answer loss as the primary effect
340 measure}\label{teacher-forced-answer-loss-as-the-primary-effect-measure}
341 
342 Free generation is useful descriptively but introduces parsing,
343 decoding, and stochasticity. The main causal effects use teacher-forced
344 protected answer loss while holding the prompt and suffix constant.
345 Free-generation accuracy is a secondary behavioral witness.
346 
347 \subsubsection{3. Content specificity is not procedure
348 specificity}\label{content-specificity-is-not-procedure-specificity}
349 
350 A same-answer donor rules out pure final-answer encoding. It does not
351 automatically prove a different algorithm is represented. On PrOntoQA
352 the result is called content/proof-instance specificity. Procedure
353 specificity is reserved for a task family with varied proof topology and
354 an explicit far-versus-near structural contrast.
355 
356 \subsubsection{4. Explicit CoT is post-seal
357 only}\label{explicit-cot-is-post-seal-only}
358 
359 The multimode checkpoint can be given gold proof steps in verbalized
360 mode. Allowing those traces into semantic construction would leak the
361 target ontology. Therefore explicit CoT is revealed only after the
362 quotient seal as a comparison witness.
363 
364 \subsubsection{5. Covariance calibration is
365 disjoint}\label{covariance-calibration-is-disjoint}
366 
367 Covariance-matched noise must not estimate its nuisance distribution
368 from the same rows used to measure the final effect. Twenty percent of
369 proof groups are deterministically reserved for covariance calibration.
370 
371 \subsubsection{6. Semantic final traces are not computed before
372 sealing}\label{semantic-final-traces-are-not-computed-before-sealing}
373 
374 The previous prototype split the records correctly but captured every
375 semantic orbit before choosing the quotient. The v0.3 implementation now
376 captures discovery, then calibration, seals the selected quotient, and
377 only then captures final semantic twins and counterfactuals.
378 
379 \subsubsection{7. Fail closed}\label{fail-closed}
380 
381 Replay failure terminates the causal claim. Missing source objects
382 terminate setup. Dirty or wrong upstream commits terminate source
383 verification. No semantic candidate produces
384 \texttt{ONTOLOGY\_UNDERIDENTIFIED}, not a threshold adjustment.
385 
386 \subsubsection{8. Smallest valid
387 quotient}\label{smallest-valid-quotient}
388 
389 Semantic construction is validity first, then minimum rank. Full hidden
390 state is a descriptive upper control and is ineligible for selection.
391 
392 \subsection{Alternatives rejected for the first
393 run}\label{alternatives-rejected-for-the-first-run}
394 
395 \begin{itemize}
396 \tightlist
397 \item
398   Training a new latent model: confounds model construction with
399   external validation.
400 \item
401   Beginning with a 3B checkpoint: raises adapter and memory risk without
402   improving causal clarity.
403 \item
404   Using explicit CoT during quotient discovery: leaks proof semantics.
405 \item
406   Calling any same-answer transplant procedure-specific: overstates what
407   fixed-chain tasks identify.
408 \item
409   Estimating covariance from the final audit: weakens the intervention's
410   independence.
411 \item
412   Selecting only tasks where latent recurrence helps: creates
413   post-selection bias. The full causal population is reported first.
414 \item
415   Lowering gates after failure: invalidates the sealed experiment.
416 \end{itemize}
417 
418 \subsection{Remaining judgment calls}\label{remaining-judgment-calls}
419 
420 The user was asked whether to prioritize a small clean causal proof or
421 the strongest latent model. No answer was available before
422 implementation, so the conservative default was chosen: small and
423 source-faithful first.
424 
425 The user was also asked whether explicit CoT should help construct
426 semantics or remain post-seal. The conservative default is post-seal
427 only.
428 
429 The next strategic choice after the PrOntoQA primary result is whether
430 to proceed immediately to ProsQA or first expand the PrOntoQA final
431 population. The recommended default is to inspect the primary receipt,
432 then use ProsQA only if replay and at least useful ordered recurrence
433 are supported.
434 
435 \newpage
436 
437 \section{Part II - Canonical
438 protocol}\label{part-ii---canonical-protocol}
439 
440 \subsection{External Latent-Reasoner Causal-Semantic
441 Admission}\label{external-latent-reasoner-causal-semantic-admission}
442 
443 \subsubsection{1. Purpose}\label{purpose}
444 
445 Block 24 is the first SRIC protocol designed to admit or reject a
446 third-party latent-reasoning checkpoint under a frozen causal and
447 semantic standard. The primary target is the multimode GPT-2 Coconut
448 model trained on PrOntoQA. The protocol treats a negative result as
449 complete evidence when the failure is reproducible and correctly
450 localized.
451 
452 \subsubsection{2. Primary scientific
453 question}\label{primary-scientific-question}
454 
455 Does the recurrent latent trace of the external checkpoint carry
456 content-specific, ordered, structured information that causally
457 contributes to its answer and supports a post-seal low-rank semantic
458 quotient?
459 
460 The protocol does not initially ask whether every latent step has a
461 unique English translation. It asks whether the latent computation has
462 operationally identifiable causal content.
463 
464 \subsubsection{3. Source lock}\label{source-lock}
465 
466 The external model is admitted only when all of the following match the
467 bundled lock:
468 
469 \begin{itemize}
470 \tightlist
471 \item
472   repository: \texttt{connordilgren/are-lrms-easily-interpretable};
473 \item
474   commit: \texttt{32f413d8d55239d9bc54bb6b6ec37b0630891ed4};
475 \item
476   multimode wrapper Git blob:
477   \texttt{d0b353cd78f98ab7aae0d20ce4e8859d9fd9958e};
478 \item
479   Coconut recurrence Git blob:
480   \texttt{646acb08f149afec075eefb212c8aa1f3ac8ceda};
481 \item
482   requirements Git blob:
483   \texttt{4a63ea4b92f0dd565f5d632241e36299695237e4};
484 \item
485   PrOntoQA test-data Git blob:
486   \texttt{7222832226883a545328f2398dc4976b1bb85c7d};
487 \item
488   checkpoint repository revision:
489   \texttt{4ec759a9f9b8a8e7e1a8d1e3e2d43eef969ac83f};
490 \item
491   checkpoint filename: \texttt{checkpoint\_37};
492 \item
493   GPT-2 base revision:
494   \texttt{607a30d783dfa663caf39e06633721c8d4cfcd7e}.
495 \end{itemize}
496 
497 The downloaded checkpoint is hashed in full. The complete local
498 base-model payload is hashed as a path-size-file-digest tree with
499 download metadata excluded.
500 
501 \subsubsection{4. Adapter contract}\label{adapter-contract}
502 
503 The adapter must expose the exact vectors reinjected at latent-token
504 positions. It must support:
505 
506 \begin{itemize}
507 \tightlist
508 \item
509   ordinary latent capture;
510 \item
511   exact supplied-trace replay;
512 \item
513   direct-answer mode;
514 \item
515   post-seal verbalized mode;
516 \item
517   teacher-forced protected answer scoring.
518 \end{itemize}
519 
520 The trace engine is compared against a separately written reference
521 implementation of the reviewed source recurrence. Exact replay is then
522 tested on every audit row. A replay error above the frozen tolerance
523 returns \texttt{ADAPTER\_INVALID} and terminates all later
524 interpretation.
525 
526 \subsubsection{5. Population and
527 grouping}\label{population-and-grouping}
528 
529 Tasks are selected deterministically from the pinned public data. Splits
530 are made by \texttt{group\_id}, not by transformed text row, so semantic
531 twins and related records cannot cross evidence partitions.
532 
533 For the primary 64-task profile:
534 
535 \begin{itemize}
536 \tightlist
537 \item
538   20\% of groups are reserved for covariance calibration;
539 \item
540   the remainder forms the causal audit population;
541 \item
542   if causal admission succeeds, that population is split into 50\%
543   semantic discovery, 25\% semantic calibration, and 25\% semantic final
544   support.
545 \end{itemize}
546 
547 The exact row counts depend on group sizes and are frozen in the
548 pre-audit seal.
549 
550 \subsubsection{6. Causal interventions}\label{causal-interventions}
551 
552 For task \texttt{i} with clean trace \texttt{H\_i} and clean answer loss
553 \texttt{L\_i}, the protocol computes:
554 
555 \paragraph{Exact replay}\label{exact-replay}
556 
557 \texttt{r\_i\ =\ max\_abs(logits(clean)\ -\ logits(replay(H\_i)))}.
558 
559 \paragraph{Latent necessity}\label{latent-necessity}
560 
561 \texttt{n\_i\ =\ L\_i(zero\ trace)\ -\ L\_i}.
562 
563 \paragraph{Step order}\label{step-order}
564 
565 \texttt{o\_i\ =\ L\_i(permuted\ trace)\ -\ L\_i}.
566 
567 \paragraph{Norm-matched structure}\label{norm-matched-structure}
568 
569 \texttt{c\_i\ =\ L\_i(norm-matched\ noise)\ -\ L\_i}.
570 
571 \paragraph{Covariance-matched
572 structure}\label{covariance-matched-structure}
573 
574 A nuisance covariance is estimated only on the reserved calibration
575 support. The audit effect is:
576 
577 \texttt{c\_cov\_i\ =\ L\_i(covariance-matched\ noise)\ -\ L\_i}.
578 
579 \paragraph{Same-answer content
580 transplantation}\label{same-answer-content-transplantation}
581 
582 A donor has the same final answer, preferentially the same proof depth,
583 a different orbit, and maximal proof-signature distance within the
584 frozen donor population:
585 
586 \texttt{s\_i\ =\ L\_i(H\_donor\_same\_answer)\ -\ L\_i}.
587 
588 This is a content/proof-instance specificity test. On PrOntoQA it is not
589 labeled a different-algorithm procedure test.
590 
591 \paragraph{Different-answer
592 transplantation}\label{different-answer-transplantation}
593 
594 This is a weaker positive control showing sensitivity to answer-bearing
595 task differences.
596 
597 \paragraph{Semantic-twin
598 transplantation}\label{semantic-twin-transplantation}
599 
600 The premise order is deterministically changed without changing the
601 protected answer, proof chain, group, or proof signature:
602 
603 \texttt{t\_i\ =\ L\_i(H\_twin)\ -\ L\_i}.
604 
605 A strong result requires content sensitivity together with twin
606 stability.
607 
608 \paragraph{Truncation}\label{truncation}
609 
610 The model is scored with zero through K retained latent steps to
611 estimate when near-clean performance appears.
612 
613 \subsubsection{7. Causal status ordering}\label{causal-status-ordering}
614 
615 The status is assigned lexicographically:
616 
617 \begin{enumerate}
618 \def\labelenumi{\arabic{enumi}.}
619 \tightlist
620 \item
621   replay;
622 \item
623   necessity;
624 \item
625   order;
626 \item
627   structured corruption;
628 \item
629   semantic-twin stability;
630 \item
631   declared specificity tier.
632 \end{enumerate}
633 
634 The first failed gate determines the terminal causal status. This
635 prevents a strong later statistic from compensating for an invalid
636 adapter or unused latent loop.
637 
638 \subsubsection{8. Mode controls}\label{mode-controls}
639 
640 Direct-answer mode is permitted during causal admission because it uses
641 no gold proof. Verbalized mode appends the protected proof steps and is
642 therefore inaccessible before the semantic quotient seal.
643 
644 The explicit-CoT result is descriptive. It may help interpret a sealed
645 quotient, but it may not determine its basis, rank, thresholds, or
646 acceptance.
647 
648 \subsubsection{9. Semantic construction
649 interface}\label{semantic-construction-interface}
650 
651 Semantic escalation occurs only after \texttt{CONTENT\_SPECIFIC} or
652 \texttt{PROCEDURE\_SPECIFIC} causal admission.
653 
654 The constructor sees:
655 
656 \begin{itemize}
657 \tightlist
658 \item
659   current latent state;
660 \item
661   next latent state;
662 \item
663   answer-residual future state;
664 \item
665   anonymous query-polarity intervention response;
666 \item
667   final answer label;
668 \item
669   semantic-twin relation.
670 \end{itemize}
671 
672 It does not see the protected proof-instance witness or explicit-CoT
673 trace.
674 
675 \subsubsection{10. Candidate grammar}\label{candidate-grammar}
676 
677 The finite grammar contains:
678 
679 \begin{itemize}
680 \tightlist
681 \item
682   answer-only rank-one direction;
683 \item
684   PCA directions;
685 \item
686   predictive directions;
687 \item
688   intervention-response directions;
689 \item
690   joint predictive-interventional directions;
691 \item
692   equal-rank random directions;
693 \item
694   full-state descriptive upper control.
695 \end{itemize}
696 
697 Candidate bases are canonicalized for deterministic signs and
698 orthonormality.
699 
700 \subsubsection{11. Candidate gates}\label{candidate-gates}
701 
702 A candidate must exceed frozen thresholds for:
703 
704 \begin{itemize}
705 \tightlist
706 \item
707   future-state \texttt{R\^{}2};
708 \item
709   answer-residual future-state \texttt{R\^{}2};
710 \item
711   intervention-response \texttt{R\^{}2};
712 \item
713   final-answer accuracy;
714 \item
715   semantic-twin drift;
716 \item
717   maximum rank.
718 \end{itemize}
719 
720 Among valid candidates, the minimum rank is selected. Ties are broken by
721 weakest normalized margin and then a frozen method order. Full state is
722 not eligible.
723 
724 \subsubsection{12. Semantic chronology}\label{semantic-chronology}
725 
726 The order is enforced in code and tested:
727 
728 \begin{enumerate}
729 \def\labelenumi{\arabic{enumi}.}
730 \tightlist
731 \item
732   declare group-disjoint discovery, calibration, and final IDs;
733 \item
734   capture discovery semantic orbits;
735 \item
736   capture calibration semantic orbits;
737 \item
738   construct candidates;
739 \item
740   evaluate candidates only on calibration;
741 \item
742   select quotient;
743 \item
744   hash basis, center, split IDs, and thresholds into the quotient seal;
745 \item
746   capture final semantic orbits only after that seal exists;
747 \item
748   fit heads on discovery plus calibration;
749 \item
750   evaluate final construction metrics;
751 \item
752   reveal post-seal witness;
753 \item
754   run causal subspace patching;
755 \item
756   reveal explicit-CoT comparison.
757 \end{enumerate}
758 
759 \subsubsection{13. Post-seal proof-instance
760 witness}\label{post-seal-proof-instance-witness}
761 
762 The current PrOntoQA witness includes answer polarity, recurrent
763 progress, and an audit-only proof-instance coordinate derived from the
764 protected proof signature. It is deliberately described as
765 proof-instance content, not an algorithmic procedure ontology. A
766 stronger ProsQA witness will use explicit graph-structural features.
767 
768 \subsubsection{14. Causal subspace
769 patching}\label{causal-subspace-patching}
770 
771 For a selected basis \texttt{B}, the projector is
772 \texttt{P\ =\ B\ B\^{}T}. A donor semantic packet is inserted while
773 preserving the orthogonal complement:
774 
775 \texttt{H\_patch\ =\ H\_recipient\ +\ (H\_donor\ -\ H\_recipient)\ P}.
776 
777 The selected packet is compared with:
778 
779 \begin{itemize}
780 \tightlist
781 \item
782   full-trace transplant;
783 \item
784   answer-only packet;
785 \item
786   random equal-rank packet;
787 \item
788   PCA packet as a descriptive control.
789 \end{itemize}
790 
791 The causal gate requires a minimum recovery ratio, positive-task
792 fraction, and superiority over answer-only and random shortcuts.
793 
794 \subsubsection{15. Statistical reporting}\label{statistical-reporting}
795 
796 Effects are reported for the complete causal audit population. Bootstrap
797 intervals resample task rows. For final publication, proof-template or
798 group-level resampling should be preferred whenever public metadata
799 support it.
800 
801 Correct and incorrect model predictions must be reported separately in
802 the real-result summary. A necessity-positive stratum may be reported
803 only as a secondary predeclared analysis; it may not replace the full
804 population.
805 
806 \subsubsection{16. Result interpretation}\label{result-interpretation}
807 
808 \begin{itemize}
809 \tightlist
810 \item
811   \texttt{LATENT\_UNUSED}: no evidence that the checkpoint relies
812   materially on its latent loop in this population.
813 \item
814   \texttt{LATENT\_GENERIC}: evidence for useful ordered latent
815   computation, but not enough same-answer content specificity.
816 \item
817   \texttt{CONTENT\_SPECIFIC}: the latent trace is tied to proof-instance
818   content beyond final-answer identity.
819 \item
820   \texttt{ONTOLOGY\_UNDERIDENTIFIED}: no small quotient satisfied the
821   observable causal-predictive contract.
822 \item
823   \texttt{SEMANTIC\_CANDIDATE}: a quotient passed construction but not
824   the post-seal proof-instance witness.
825 \item
826   \texttt{SEMANTICALLY\_IDENTIFIED}: the witness passed, but subspace
827   editing did not.
828 \item
829   \texttt{CAUSALLY\_VALIDATED}: the sealed quotient passed construction,
830   post-seal witness, and causal patching.
831 \end{itemize}
832 
833 No status means that the trace is a complete or unique transcript of
834 hidden thought.
835 
836 \subsubsection{17. Engineered positive
837 control}\label{engineered-positive-control}
838 
839 The package includes a synthetic adapter whose trace contains a low-rank
840 causal content packet and high-variance nuisance dimensions. The fixture
841 validates the transaction and analysis mechanics. It reaches
842 \texttt{CAUSALLY\_VALIDATED} by design and is always marked
843 \texttt{engineered\_positive\_control:\ true}.
844 
845 \subsubsection{18. Threats to validity}\label{threats-to-validity}
846 
847 \paragraph{Adapter mismatch}\label{adapter-mismatch}
848 
849 Mitigation: source-lock verification, recurrence parity test, exact
850 replay, fixed tokenization and suffix.
851 
852 \paragraph{Final-support leakage}\label{final-support-leakage}
853 
854 Mitigation: final semantic orbits are not captured until after quotient
855 sealing; explicit CoT is post-seal only.
856 
857 \paragraph{Answer shortcut}\label{answer-shortcut}
858 
859 Mitigation: same-answer donors and answer-residual future prediction.
860 
861 \paragraph{Surface shortcut}\label{surface-shortcut}
862 
863 Mitigation: premise-order twins and later distractor/renaming orbit
864 extensions.
865 
866 \paragraph{Procedure overclaim}\label{procedure-overclaim}
867 
868 Mitigation: PrOntoQA status is content-specific; procedure-specific
869 status is reserved for varied structural tasks.
870 
871 \paragraph{Noise miscalibration}\label{noise-miscalibration}
872 
873 Mitigation: covariance is estimated on group-disjoint calibration rows.
874 
875 \paragraph{Unsupported semantic edit}\label{unsupported-semantic-edit}
876 
877 Mitigation: causal patching is compared with equal-rank shortcut
878 controls and can fail independently of probe scores.
879 
880 \paragraph{Post-result threshold
881 changes}\label{post-result-threshold-changes}
882 
883 Mitigation: config, model, source, dataset, donors, splits, and output
884 contract are sealed before execution.
885 
886 \subsubsection{19. Hardware plan}\label{hardware-plan}
887 
888 The primary target is intentionally small. The recommended configuration
889 for an 8 GB RTX 3070 Ti is batch size one, deterministic inference,
890 short prompts, capture caching, and moving saved traces to CPU. The
891 package does not retrain the external model.
892 
893 \subsubsection{20. Escalation path}\label{escalation-path}
894 
895 \begin{enumerate}
896 \def\labelenumi{\arabic{enumi}.}
897 \tightlist
898 \item
899   Run 12-task smoke solely to validate source parity and replay.
900 \item
901   Run 64-task primary without changing gates.
902 \item
903   Inspect the canonical status and failure coordinate.
904 \item
905   Run the 256-task profile only if the adapter remains valid.
906 \item
907   Add ProsQA only after PrOntoQA produces a stable causal result.
908 \item
909   Add CODI or another latent method for cross-model transport only after
910   one checkpoint has a valid semantic object.
911 \item
912   Add SLPO for adaptive-depth analysis.
913 \item
914   Add Hyper as a scale stress.
915 \end{enumerate}
916 
917 \subsubsection{21. Stopping rules}\label{stopping-rules}
918 
919 \begin{itemize}
920 \tightlist
921 \item
922   Replay failure: stop and repair the adapter; do not interpret latent
923   interventions.
924 \item
925   Out-of-memory during smoke: reduce runtime memory, not scientific
926   thresholds.
927 \item
928   \texttt{LATENT\_UNUSED}: preserve as the primary result; do not filter
929   to successful tasks as the headline.
930 \item
931   \texttt{LATENT\_GENERIC}: do not run semantic escalation.
932 \item
933   No valid quotient: report \texttt{ONTOLOGY\_UNDERIDENTIFIED}.
934 \item
935   Causal patch failure: do not call the quotient causally validated.
936 \end{itemize}
937 
938 \subsubsection{22. Current execution
939 status}\label{current-execution-status}
940 
941 The source-locked external checkpoint was not executed in the
942 environment used to build this package. That environment had CPU-only
943 PyTorch, no Transformers installation, no external checkout, no
944 checkpoint, no public dataset, and no CUDA device. The package's
945 validated result is the engineered fixture only.
946 
947 \subsubsection{23. Decision after the first real
948 receipt}\label{decision-after-the-first-real-receipt}
949 
950 The first real receipt should determine the next move:
951 
952 \begin{itemize}
953 \tightlist
954 \item
955   adapter invalid -\textgreater{} repair only adapter fidelity;
956 \item
957   latent unused -\textgreater{} test a second published checkpoint or
958   task, without rewriting the result;
959 \item
960   latent generic -\textgreater{} compare latent and direct modes and
961   stop semantic claims;
962 \item
963   content specific -\textgreater{} execute sealed quotient construction;
964 \item
965   semantic candidate -\textgreater{} improve independent witness design
966   in a new protocol version;
967 \item
968   causally validated -\textgreater{} proceed to ProsQA structure
969   sensitivity and cross-model transport.
970 \end{itemize}
971 
972 \newpage
973 
974 \section{Engineered positive-control
975 result}\label{engineered-positive-control-result}
976 
977 The included fixture is deterministic and deliberately contains a
978 low-rank causal semantic packet. It validates the mechanics only; it is
979 not a trained language model.
980 
981 \begin{longtable}[]{@{}lr@{}}
982 \toprule\noalign{}
983 Quantity & Frozen result \\
984 \midrule\noalign{}
985 \endhead
986 \bottomrule\noalign{}
987 \endlastfoot
988 Fixture tasks & 48 \\
989 Covariance calibration tasks & 8 \\
990 Causal audit tasks & 40 \\
991 Causal status & \texttt{CONTENT\_SPECIFIC} \\
992 Semantic status & \texttt{CAUSALLY\_VALIDATED} \\
993 Selected quotient & \texttt{joint}, rank 3 \\
994 Replay maximum & 0 \\
995 Necessity & 5.814125 {[}5.655045, 5.964764{]} \\
996 Order & 0.829070 {[}0.784764, 0.874355{]} \\
997 Norm-matched structure & 7.342361 {[}7.027469, 7.659845{]} \\
998 Covariance-matched structure & 12.649418 {[}12.020959, 13.362811{]} \\
999 Same-answer content transplant & 4.408597 {[}3.409079, 5.523186{]} \\
1000 Different-answer transplant & 14.829357 {[}13.318282, 16.259007{]} \\
1001 Semantic-twin transplant & 0.000000 {[}0.000000, 0.000000{]} \\
1002 Final future R2 & 0.751820 \\
1003 Final answer-residual future R2 & 0.636851 \\
1004 Final intervention R2 & 0.984125 \\
1005 Post-seal proof-witness R2 & 0.954829 \\
1006 Causal patch passed & \texttt{true} \\
1007 \end{longtable}
1008 
1009 \begin{Shaded}
1010 \begin{Highlighting}[]
1011 \NormalTok{fixture pre{-}audit seal: 68b04ea973f5030a7085d2b1891e87bedbf058db007d004fefee047beaee70b9}
1012 \NormalTok{fixture result receipt: 924157159d14b8d3715f2f886c2ad294b55f5cac15a3b63f1552df1b3731e620}
1013 \NormalTok{independent verification: e4e0c5806022faa79e41a363a2100695a8b9d0f33a85c931014f8c732b84c9ca}
1014 \end{Highlighting}
1015 \end{Shaded}
1016 
1017 This result validates the decision logic, evidence chronology, and
1018 intervention plumbing. It establishes nothing about the public
1019 checkpoint.
1020 
1021 \section{Current runtime boundary}\label{current-runtime-boundary}
1022 
1023 \begin{Shaded}
1024 \begin{Highlighting}[]
1025 \FunctionTok{\{}
1026   \DataTypeTok{"checkpoint\_exists"}\FunctionTok{:} \KeywordTok{false}\FunctionTok{,}
1027   \DataTypeTok{"cuda\_available"}\FunctionTok{:} \KeywordTok{false}\FunctionTok{,}
1028   \DataTypeTok{"data\_exists"}\FunctionTok{:} \KeywordTok{false}\FunctionTok{,}
1029   \DataTypeTok{"external\_repo\_exists"}\FunctionTok{:} \KeywordTok{false}\FunctionTok{,}
1030   \DataTypeTok{"ready"}\FunctionTok{:} \KeywordTok{false}\FunctionTok{,}
1031   \DataTypeTok{"torch"}\FunctionTok{:} \StringTok{"2.10.0+cpu"}\FunctionTok{,}
1032   \DataTypeTok{"transformers\_error"}\FunctionTok{:} \StringTok{"No module named \textquotesingle{}transformers\textquotesingle{}"}
1033 \FunctionTok{\}}
1034 \end{Highlighting}
1035 \end{Shaded}
1036 
1037 The third-party checkpoint result remains unknown.
1038 
1039 \section{Part III - Source interface and
1040 chronology}\label{part-iii---source-interface-and-chronology}
1041 
1042 The adapter is valid only when it satisfies all of the following.
1043 
1044 \subsection{Trace identity}\label{trace-identity}
1045 
1046 \texttt{capture(task)} returns the vectors actually reinjected at the
1047 model's latent-token positions, in recurrent order, with shape
1048 \texttt{{[}latent\_steps,\ hidden\_size{]}}.
1049 
1050 \subsection{Supplied-trace replay}\label{supplied-trace-replay}
1051 
1052 \texttt{score\_with\_trace(task,\ trace)} runs the same teacher-forced
1053 prompt and answer suffix while replacing only the reinjected latent
1054 vectors. Exact own-trace replay must reproduce answer logits within the
1055 frozen tolerance.
1056 
1057 \subsection{Source parity}\label{source-parity}
1058 
1059 The package includes a separate reference recurrence derived from the
1060 pinned upstream source. The adapter engine must match it exactly on a
1061 controlled PyTorch fixture.
1062 
1063 \subsection{Fixed surfaces}\label{fixed-surfaces}
1064 
1065 Across clean and intervention arms, the following remain fixed:
1066 
1067 \begin{itemize}
1068 \tightlist
1069 \item
1070   checkpoint and base model;
1071 \item
1072   tokenizer and special tokens;
1073 \item
1074   prompt text for the recipient task;
1075 \item
1076   token positions;
1077 \item
1078   attention mask;
1079 \item
1080   answer suffix;
1081 \item
1082   teacher-forced label mask;
1083 \item
1084   decoding settings.
1085 \end{itemize}
1086 
1087 \subsection{Modes}\label{modes}
1088 
1089 \begin{itemize}
1090 \tightlist
1091 \item
1092   \texttt{latent}: admissible for causal and semantic construction.
1093 \item
1094   \texttt{direct}: admissible as a pre-seal mode control.
1095 \item
1096   \texttt{verbalized}: contains gold proof steps and is post-seal only.
1097 \end{itemize}
1098 
1099 \subsection{Fingerprint}\label{fingerprint}
1100 
1101 The adapter fingerprint commits:
1102 
1103 \begin{itemize}
1104 \tightlist
1105 \item
1106   upstream repository and commit;
1107 \item
1108   source-verification receipt;
1109 \item
1110   checkpoint repository, revision, filename, and full SHA-256;
1111 \item
1112   local base-model path, revision, and tree SHA-256;
1113 \item
1114   latent-step count;
1115 \item
1116   device type.
1117 \end{itemize}
1118 
1119 \subsection{Failure}\label{failure}
1120 
1121 Any replay, source-lock, cache representation, latent-position, or
1122 trace-shape mismatch returns an adapter failure. It must not be treated
1123 as evidence about the model's semantics.
1124 
1125 \begin{Shaded}
1126 \begin{Highlighting}[]
1127 \NormalTok{T0  package code, thresholds, claim tier, source lock frozen}
1128 \NormalTok{T1  external repository checked out at exact commit}
1129 \NormalTok{T2  required Git blobs verified and dirty tree rejected}
1130 \NormalTok{T3  pinned checkpoint and GPT{-}2 revision downloaded}
1131 \NormalTok{T4  full checkpoint and local GPT{-}2 payload hashed}
1132 \NormalTok{T5  deterministic task population and group IDs frozen}
1133 \NormalTok{T6  covariance{-}calibration and causal{-}audit groups frozen}
1134 \NormalTok{T7  donor plans and semantic split IDs frozen}
1135 \NormalTok{T8  pre{-}audit seal written}
1136 \NormalTok{T9  calibration traces captured and covariance committed}
1137 \NormalTok{T10 causal audit executed on disjoint audit rows}
1138 \NormalTok{T11 causal status assigned}
1139 \NormalTok{T12 if admitted, semantic discovery traces captured}
1140 \NormalTok{T13 semantic calibration traces captured}
1141 \NormalTok{T14 candidate quotient chosen}
1142 \NormalTok{T15 quotient basis, center, thresholds, and splits sealed}
1143 \NormalTok{T16 semantic final traces captured for the first time}
1144 \NormalTok{T17 post{-}seal proof{-}instance witness revealed}
1145 \NormalTok{T18 causal subspace patching executed}
1146 \NormalTok{T19 explicit{-}CoT comparison revealed}
1147 \NormalTok{T20 canonical result receipt written}
1148 \end{Highlighting}
1149 \end{Shaded}
1150 
1151 No T16-T19 object may influence T14. The implementation records and
1152 tests the semantic chronology.
1153 
1154 The external assets are intentionally not redistributed.
1155 
1156 \subsection{Source repository}\label{source-repository}
1157 
1158 \begin{itemize}
1159 \tightlist
1160 \item
1161   \texttt{connordilgren/are-lrms-easily-interpretable}
1162 \item
1163   pinned commit: \texttt{32f413d8d55239d9bc54bb6b6ec37b0630891ed4}
1164 \end{itemize}
1165 
1166 The package verifies the wrapper, latent runtime, requirements, and
1167 selected dataset Git blobs listed in
1168 \texttt{external\_source\_lock.json}.
1169 
1170 \subsection{Primary checkpoint}\label{primary-checkpoint}
1171 
1172 \begin{itemize}
1173 \tightlist
1174 \item
1175   repository: \texttt{connordilgren/gpt2-prontoqa-multimode-coconut}
1176 \item
1177   revision: \texttt{4ec759a9f9b8a8e7e1a8d1e3e2d43eef969ac83f}
1178 \item
1179   file: \texttt{checkpoint\_37}
1180 \end{itemize}
1181 
1182 \subsection{Base model}\label{base-model}
1183 
1184 \begin{itemize}
1185 \tightlist
1186 \item
1187   repository: \texttt{openai-community/gpt2}
1188 \item
1189   revision: \texttt{607a30d783dfa663caf39e06633721c8d4cfcd7e}
1190 \end{itemize}
1191 
1192 \subsection{Runtime}\label{runtime}
1193 
1194 The isolated real environment pins PyTorch 2.5.1 and Transformers
1195 4.46.2, matching the reviewed external requirements where exact pins
1196 were available.
1197 
1198 \subsection{Secondary planned
1199 checkpoint}\label{secondary-planned-checkpoint}
1200 
1201 ProsQA source and checkpoint revisions are recorded in
1202 \texttt{external\_source\_lock.json}. They are not executed by the
1203 primary v0.3.0 run path.
1204 
1205 \newpage
1206 
1207 \section{Part IV - Real execution status and
1208 runbook}\label{part-iv---real-execution-status-and-runbook}
1209 
1210 \subsection{Canonical status}\label{canonical-status}
1211 
1212 \texttt{EXTERNAL\_EXECUTION\_NOT\_RUN}
1213 
1214 The build environment used for v0.3.0 reported:
1215 
1216 \begin{itemize}
1217 \tightlist
1218 \item
1219   CPU-only PyTorch available;
1220 \item
1221   CUDA unavailable;
1222 \item
1223   Transformers unavailable;
1224 \item
1225   pinned upstream checkout absent;
1226 \item
1227   checkpoint absent;
1228 \item
1229   public dataset absent.
1230 \end{itemize}
1231 
1232 No attempt was made to substitute a different model or fabricate
1233 external metrics. The external checkpoint remains an open empirical
1234 test.
1235 
1236 \subsection{What was executed}\label{what-was-executed}
1237 
1238 \begin{itemize}
1239 \tightlist
1240 \item
1241   engineered fixture generation;
1242 \item
1243   engineered pre-audit sealing;
1244 \item
1245   causal and semantic pipeline on the fixture;
1246 \item
1247   automated test suite;
1248 \item
1249   source-lock unit tests;
1250 \item
1251   recurrence parity fixture;
1252 \item
1253   chronology and split-isolation tests;
1254 \item
1255   independent recomputation of fixture means and semantic-basis
1256   properties;
1257 \item
1258   package closure and manifest verification.
1259 \end{itemize}
1260 
1261 \subsection{What must happen on the user's
1262 machine}\label{what-must-happen-on-the-users-machine}
1263 
1264 \begin{enumerate}
1265 \def\labelenumi{\arabic{enumi}.}
1266 \tightlist
1267 \item
1268   Run \texttt{Setup\_External\_PrOntoQA\_Coconut.ps1}.
1269 \item
1270   Run the smoke profile.
1271 \item
1272   Inspect replay and source verification.
1273 \item
1274   Run the primary profile without modifying the frozen config.
1275 \item
1276   Preserve the complete \texttt{frozen}, \texttt{results},
1277   \texttt{capture\_cache}, and environment records.
1278 \end{enumerate}
1279 
1280 \subsection{Failure policy}\label{failure-policy}
1281 
1282 The first real status is the result. An adapter failure permits an
1283 adapter-only revision because no semantic claim is yet admissible. Any
1284 scientific threshold revision requires a new protocol and new final
1285 support.
1286 
1287 \subsection{Before setup}\label{before-setup}
1288 
1289 \begin{itemize}
1290 \tightlist
1291 \item[$\square$]
1292   At least 25 GB free disk space.
1293 \item[$\square$]
1294   NVIDIA driver operational.
1295 \item[$\square$]
1296   Python 3.10, 3.11, or 3.12 available through \texttt{py}.
1297 \item[$\square$]
1298   Git available in PowerShell.
1299 \item[$\square$]
1300   Network access to GitHub, PyPI, PyTorch wheels, and Hugging Face.
1301 \end{itemize}
1302 
1303 \subsection{After setup}\label{after-setup}
1304 
1305 \begin{itemize}
1306 \tightlist
1307 \item[$\square$]
1308   \texttt{runtime\_paths.json} exists.
1309 \item[$\square$]
1310   upstream commit equals the source lock.
1311 \item[$\square$]
1312   source blob verification passes.
1313 \item[$\square$]
1314   checkpoint revision and full SHA-256 are recorded.
1315 \item[$\square$]
1316   GPT-2 revision and local tree SHA-256 are recorded.
1317 \item[$\square$]
1318   \texttt{pip\ check} passes.
1319 \item[$\square$]
1320   CUDA is visible to the isolated environment.
1321 \end{itemize}
1322 
1323 \subsection{Smoke run}\label{smoke-run}
1324 
1325 \begin{itemize}
1326 \tightlist
1327 \item[$\square$]
1328   Seal is written before model audit.
1329 \item[$\square$]
1330   Exact replay passes all 12 rows.
1331 \item[$\square$]
1332   Prompt, suffix, latent positions, and answer mask are stable.
1333 \item[$\square$]
1334   No out-of-memory or cache-format error occurs.
1335 \end{itemize}
1336 
1337 \subsection{Primary run}\label{primary-run}
1338 
1339 \begin{itemize}
1340 \tightlist
1341 \item[$\square$]
1342   Configuration file unchanged from the sealed copy.
1343 \item[$\square$]
1344   Covariance calibration rows are group-disjoint from audit rows.
1345 \item[$\square$]
1346   Donor plans cover every causal audit row once.
1347 \item[$\square$]
1348   Explicit CoT remains unavailable before semantic seal.
1349 \item[$\square$]
1350   Final semantic orbits are captured only after quotient seal.
1351 \item[$\square$]
1352   Complete result and environment records are retained even on failure.
1353 \end{itemize}
1354 
1355 \subsection{After result}\label{after-result}
1356 
1357 \begin{itemize}
1358 \tightlist
1359 \item[$\square$]
1360   Run \texttt{verify-result}.
1361 \item[$\square$]
1362   Generate summary.
1363 \item[$\square$]
1364   Copy the complete result directory before any exploratory analysis.
1365 \item[$\square$]
1366   Do not relabel content specificity as algorithmic procedure
1367   specificity.
1368 \end{itemize}
1369 
1370 \subsection{Objective}\label{objective}
1371 
1372 Execute a source-locked causal-semantic audit of the public multimode
1373 GPT-2 Coconut PrOntoQA checkpoint without changing the v0.3.0 gates
1374 after result access.
1375 
1376 \subsection{Recommended first command}\label{recommended-first-command}
1377 
1378 \begin{Shaded}
1379 \begin{Highlighting}[]
1380 \OperatorTok{.}\NormalTok{\textbackslash{}Setup\_External\_PrOntoQA\_Coconut}\OperatorTok{.}\FunctionTok{ps1}
1381 \end{Highlighting}
1382 \end{Shaded}
1383 
1384 Then:
1385 
1386 \begin{Shaded}
1387 \begin{Highlighting}[]
1388 \OperatorTok{.}\NormalTok{\textbackslash{}Run\_Real\_LRM\_Pilot}\OperatorTok{.}\FunctionTok{ps1} \OperatorTok{{-}}\NormalTok{Profile smoke }\OperatorTok{{-}}\NormalTok{Device cuda}
1389 \end{Highlighting}
1390 \end{Shaded}
1391 
1392 The smoke run is an adapter and replay qualification. It has semantic
1393 escalation disabled.
1394 
1395 \subsection{Advancement rule}\label{advancement-rule}
1396 
1397 Proceed to \texttt{primary} only when:
1398 
1399 \begin{itemize}
1400 \tightlist
1401 \item
1402   source verification passes;
1403 \item
1404   the checkpoint and base-model hashes are recorded;
1405 \item
1406   exact replay is below tolerance;
1407 \item
1408   no cache or token-position discrepancy is observed;
1409 \item
1410   GPU memory is stable.
1411 \end{itemize}
1412 
1413 \subsection{Scientific rule}\label{scientific-rule}
1414 
1415 Do not alter \texttt{configs/prontoqa\_multimode\_coconut\_gpt2.json}
1416 after staging. If the primary result is negative, preserve it. A new
1417 hypothesis requires a new version and seal.
1418 
1419 \subsection{Returned files to
1420 preserve}\label{returned-files-to-preserve}
1421 
1422 \begin{itemize}
1423 \tightlist
1424 \item
1425   \texttt{external\_runtime/environment.json}
1426 \item
1427   \texttt{external\_runtime/runtime\_paths.json}
1428 \item
1429   \texttt{frozen/real\_primary\_preaudit\_seal.json}
1430 \item
1431   donor plans and experiment plan beside the seal
1432 \item
1433   \texttt{results/real\_primary\_block24.json}
1434 \item
1435   \texttt{results/real\_primary\_summary.md}
1436 \item
1437   \texttt{results/capture\_cache/} if present
1438 \end{itemize}
1439 
1440 \subsection{Current package evidence}\label{current-package-evidence}
1441 
1442 The engineered positive control passed. The external model has not been
1443 executed. No real-model conclusion should be inferred from the fixture.
1444 
1445 \newpage
1446 
1447 \section{Part V - Structure-sensitive
1448 follow-up}\label{part-v---structure-sensitive-follow-up}
1449 
1450 \subsection{Purpose}\label{purpose-1}
1451 
1452 PrOntoQA is the source-faithful causal admission tier. ProsQA is the
1453 intended structure-sensitive tier. It should be activated only after the
1454 PrOntoQA adapter passes replay.
1455 
1456 \subsection{Stronger claim}\label{stronger-claim}
1457 
1458 A ProsQA result may use \texttt{PROCEDURE\_SPECIFIC} only when all of
1459 the following are implemented before final audit:
1460 
1461 \begin{enumerate}
1462 \def\labelenumi{\arabic{enumi}.}
1463 \item
1464   proof graphs or derivation structures are extracted independently of
1465   text surface;
1466 \item
1467   donors are matched on final answer and broad difficulty;
1468 \item
1469   a structure-near donor and structure-far donor are assigned for every
1470   recipient;
1471 \item
1472   the primary statistic is the contrast
1473 
1474   \texttt{loss(far\ donor)\ -\ loss(near\ donor)};
1475 \item
1476   the contrast has a positive frozen interval and positive-task
1477   fraction;
1478 \item
1479   semantic twins preserve proof topology while changing wording, entity
1480   names, and premise order;
1481 \item
1482   final proof structures are held out by template family.
1483 \end{enumerate}
1484 
1485 \subsection{Why the contrast matters}\label{why-the-contrast-matters}
1486 
1487 A harmful arbitrary donor can reflect task identity, vocabulary, or
1488 context formatting. The far-minus-near contrast asks whether disruption
1489 grows with formal solution-structure distance after answer and
1490 difficulty are controlled.
1491 
1492 \subsection{Additional adapter work}\label{additional-adapter-work}
1493 
1494 ProsQA answers are not limited to Boolean labels. A separate answer
1495 parser and protected loss contract are required. The current PrOntoQA
1496 adapter must not be silently reused as if these formats were identical.
1497 
1498 \subsection{Frozen source references}\label{frozen-source-references}
1499 
1500 The external source lock records the ProsQA test-data Git blob,
1501 checkpoint repository, checkpoint revision, and checkpoint filename.
1502 Those entries are planning locks, not evidence that the ProsQA
1503 checkpoint has been executed.
1504 
1505 \section{Part VI - Claim boundary and decision
1506 policy}\label{part-vi---claim-boundary-and-decision-policy}
1507 
1508 \subsection{Current release status}\label{current-release-status}
1509 
1510 \texttt{INFRASTRUCTURE\_VALIDATED\_EXTERNAL\_EXECUTION\_PENDING}
1511 
1512 The package validates a source-faithful intervention architecture, a
1513 staged chronology, and an engineered causal-semantic positive control.
1514 The real public checkpoint was not run in the build environment.
1515 
1516 \subsection{Supported claims}\label{supported-claims}
1517 
1518 The release supports the following statements:
1519 
1520 \begin{enumerate}
1521 \def\labelenumi{\arabic{enumi}.}
1522 \tightlist
1523 \item
1524   The adapter contains a separately tested implementation of the
1525   reviewed Coconut recurrence and exposes supplied-trace replay.
1526 \item
1527   Exact source commit and required Git blobs are pinned and verifiable.
1528 \item
1529   Checkpoint and base-model revisions are pinned by setup; local payload
1530   hashes are sealed at staging.
1531 \item
1532   Covariance-matched noise uses a group-disjoint calibration support in
1533   the Block 24 pipeline.
1534 \item
1535   Semantic discovery, calibration, quotient sealing, final capture,
1536   proof witness, causal patching, and explicit-CoT comparison occur in
1537   the declared order.
1538 \item
1539   The engineered fixture reaches \texttt{CAUSALLY\_VALIDATED} and all
1540   included automated tests pass.
1541 \item
1542   PrOntoQA's primary maximum specificity claim is content/proof-instance
1543   specificity, not proof of multiple internal reasoning algorithms.
1544 \end{enumerate}
1545 
1546 \subsection{Unsupported claims}\label{unsupported-claims}
1547 
1548 The release does not establish that the real Coconut checkpoint:
1549 
1550 \begin{itemize}
1551 \tightlist
1552 \item
1553   relies on latent recurrence;
1554 \item
1555   carries ordered latent computation;
1556 \item
1557   represents proof-instance content beyond answer identity;
1558 \item
1559   is stable under semantic twins;
1560 \item
1561   exposes a low-rank semantic quotient;
1562 \item
1563   supports causal semantic editing;
1564 \item
1565   matches its explicit natural-language proof mode;
1566 \item
1567   generalizes to ProsQA or any other task;
1568 \item
1569   is safe, conscious, sentient, or recursively self-improving.
1570 \end{itemize}
1571 
1572 The engineered fixture cannot be used as evidence for any external
1573 model.
1574 
1575 \subsection{Interpretation boundary}\label{interpretation-boundary}
1576 
1577 Even a real \texttt{CAUSALLY\_VALIDATED} result would mean only that,
1578 within the pinned model, task population, intervention language, and
1579 frozen thresholds, a compact latent subspace predicts and causally
1580 mediates declared proof-instance consequences. It would not reveal a
1581 verbatim chain of thought or prove that the discovered coordinates are
1582 the unique correct semantics.
1583 
1584 \subsection{Revision boundary}\label{revision-boundary}
1585 
1586 Any threshold, donor rule, counterfactual generator, source revision,
1587 checkpoint, model mode, or claim tier changed after a real result must
1588 produce a new protocol version and a new pre-audit seal. A failed run
1589 remains part of the evidence lineage.
1590 
1591 The next action is the 12-task source-parity and replay smoke run on the
1592 user's RTX 3070 Ti.
1593 
1594 After the primary PrOntoQA result:
1595 
1596 \begin{itemize}
1597 \tightlist
1598 \item
1599   \texttt{ADAPTER\_INVALID}: repair adapter fidelity only.
1600 \item
1601   \texttt{LATENT\_UNUSED}: preserve the result and consider a different
1602   public latent checkpoint.
1603 \item
1604   \texttt{LATENT\_GENERIC}: stop semantic escalation; compare execution
1605   modes descriptively.
1606 \item
1607   \texttt{CONTENT\_SPECIFIC}: complete the sealed semantic escalation.
1608 \item
1609   \texttt{CAUSALLY\_VALIDATED}: proceed to the ProsQA
1610   structure-sensitive protocol before any broader claim.
1611 \end{itemize}
1612 
1613 Default recommendation: do not proceed directly to Hyper or training.
1614 Resolve the smallest external checkpoint first.
1615 
1616 \end{document}

Back to source controls

Back to SRIC: external semantic admission