RESULT INCLUSION AND SOURCE/PROOF LEDGER
9 November 2026. Final source reviewed; compiled PDF and portable checks verified. Central publication decision remains separate.

COMPARISON BASIS
C: Diagnosing While Repairing, companion v0.2.5, 8 November 2026.
S: Causal Self-Diagnosis with Noisy Tests, expanded reviewed v0.3, 5 November 2026.
The shorter original baseline was also searched; the reviewed v0.3 master
includes the later law-oracle and task-recovery sections. Do not confuse it
with the original baseline's earlier-version heading.
A: From Observable States to Acquirable Predictive Models, 7 October 2026.

The local public and deliverable companion PDFs match SHA256:
a341324cf2d5b607216875f85d350de91b3f9a1bf701f018deb112827bd6e161.
The recorded deployment/live checks identified this unchanged 46-page PDF.
Fresh web opening of C's HTML/PDF was unavailable. This is local publication
identity plus recorded live evidence, not a fresh live-fetch claim.

1. FINITE RISK/COST FRONTIERS — integration-ready appendix.
Existing relatives: C Section 5 Prop 5.1(p 19), F_charged_frontier.tex 51–99;
Section 9 Prop 9.1(p 28), J_statewise_cost.tex 1–23, 44–85. Those are repair-origin
error versus uniform-prior read cost after resets. Our six-piece N3 risk/loss
and seven-piece error/worst-state-read frontiers include separate static carrier
failure. Addendum Appendix A; data/risk_curve_n3.json, cost_curve_n3.json.
Proof: exact endpoint mixtures plus supporting duals, directly recomputed by
code/verify.py. No replacement of C's frontier or claim of new LP technique.

2. BAYES, STATEWISE AND HC CONTRACTS — cross-reference plus specific HC addition.
Existing: C Section 6 G_minimax.tex 1–38, Section 9 J_statewise_cost.tex 1–14;
S03_causal.tex 154–207 already distinguishes one-use chance gates, statewise
unlocking and confidence coverage. Addendum Sections 2–4 specify after-selection,
before-execution public-history risk of FIRST exclusion now or later conditional
on survival. Equations 3–9 and Proposition 2 give complete realization constraints.
Proof: disjoint first events, exact conditional ratios, nonnegative realization
flows. code/fresh_check.py independently reconstructs 85 HC rows.

3. SEQUENTIAL EXCLUSIONS — integration-ready addendum.
S03_causal.tex 200–207 already warns that per-use caps do not bound episode risk;
S02_cumulative.tex 4–22, 47–51 instead changes cumulative order on a full carrier.
Our D ledger and two-exclusion N5 witness do not occur there. Addendum Sections 3,
4, 6: full-posterior cost/error, disjoint first-exclusion identity, risks 3/325 and
1/513. No risk resets, physical hazard interpretation or observed-survival input.

4. RANDOMIZATION — useful cross-reference and contract clarification.
Existing: C Props 6.1(p 21), 7.1(p 24), 8.1(p 26), Cor 6.3(p 23);
G_minimax.tex 22–75, 197–230, H_terminal_report.tex, I_acquisition_randomness.tex.
Our N3 carrier example has a separate first-r probability obstruction and
randomized loss 352229/780000. Addendum Sections 4–5. Distinguish public behavioral
information from conditioning on a known root seed. All shown HC components
individually pass; no general equivalence under different filtrations is claimed.

5. ALL LEGAL CARRIERS — inherited structural rule plus additional HC effect.
A Section 2 Lemma 1 already supplies the endpoint criterion and non-heredity.
The 19-action list is a direct specialization, not a structural contribution.
Addendum Equation 1 and Sections 4, 6 retain all actions; expansion witness changes
future conditional risk from 0 to 1/17. Exact benchmark has NO maximal-carrier
optimum gap. code/verify.py enumerates all 19 for the complete-class dual.

6. N5 OPTIMUM 267963/640625 — integration-ready central benchmark.
Absent from C/S/A; closest methods are C Props 5.1/6.1, not matching contracts.
Addendum Section 5 has exact dual weights, mixture weights and metrics;
data/certificates.json has complete trees. Optional episode multipliers zero
show the value is not silently strengthened by episode rows. Three distinct
checkers agree. This is composite loss, not companion pure error 6353/30625.

7. IMMEDIATE CAP IMPLIES HC THROUGH N6 — bounded integration-ready result.
No matching result in C/S/A. Addendum Section 6 Equation 12 and per-depth table;
code/check_horizons.py plus fresh_check.py. All reachable likelihood/intersection
states checked, not just root risk. Only this prior, BSC, library, alpha 1/50 and
horizons<=6. No longer-horizon equivalence or separation is inferred.

8. INTERSECTION EXAMPLE — correction/omission, not a contribution.
A Lemma 1 already makes q/xyz illegal. Earlier noise prose incorrectly used it
as one of two legal carriers with illegal intersection. Addendum Section 8
withdraws that illustration; valid q/H versus q/xz non-heredity remains.
All computed legal lists excluded xyz, so no numeric result changes. No such
error found in C/A. Nonempty intersection closure here is not a general theorem.

9. TAILORED TIME-UNIFORM GATE — integration-ready addendum.
S03_causal.tex 204–207 already states generic confidence-set coverage.
Addendum Section 7 adds actual fourth-carrier contraction to xy after 111, persistent
intersection, statewise ever failure<=1/125 and maximum HC1/65. Every maximal
optimal component fits. This proves only the specified raw PPR core pays the
measured penalty; it disproves an inference that all confidence sequences do.

OMIT AS SEPARATE MANUSCRIPT MATERIAL
Broad W(H) oracle/amplification development adapts S02_cumulative and 11_law_oracle;
the elementary hereditary-monotonicity aside is not needed. Unknown sensor-law
calibration, continuous timing and structural thinning are outside this scope.

CLOSEST PRIMARY SOURCES AND ACCESS
A: https://www.b-theory.com/manuscripts-html/acquirable-predictive-models/
Opened 9 November; Section 2 Lemma 1 read. Structural criterion attributed.
C: https://www.b-theory.com/manuscripts-html/diagnosing-while-repairing/
https://www.b-theory.com/manuscripts/diagnosing-while-repairing-companion-v0.2.5.pdf
Fresh opens unavailable; preserved publication sources and recorded evidence used.
S: preserved reviewed v0.3; no public URL or independent publication status invented.

Santana, Thiebaux, Williams, RAO*, AAAI2016:
https://people.csail.mit.edu/psantana/data/files/articles/2016_AAAI_rao.pdf
Opened 9 November; Sections 2–3, equations 6–17, especially safe-observation recursion
10–12 inspected. Antecedent for survival-conditioned execution risk; no physical
safety theorem imported into certificate coverage.

Fauss, Zoubir, A Linear Programming Approach to Sequential Hypothesis Testing:
https://arxiv.org/pdf/1501.05870
Opened 9 November; Section 4, Theorems 4.1–4.2 inspected. Classical scalarization and
LP antecedent. Our finite weak-duality proof is direct, without importing their
general regularity, differentiability or attainment assumptions.

Waudby-Smith, Ramdas, Confidence sequences for sampling without replacement:
https://papers.nips.cc/paper_files/paper/2020/file/e96c7de8f6390b1e6c71556e4e0a4959-Paper.pdf
Earlier 4 November reading: Section 2.1 Proposition 2.1. Fresh 9 November open timed
out and was not retried. Classical PPR attribution retained; the positive finite-
likelihood martingale/Ville proof is given directly in the addendum.

von Stengel, Extensive Games and the Sequence Form, author slides:
https://conferences.mpi-inf.mpg.de/adfocs-24/material/Bernhard/2ext-adfocs.pdf
Opened 9 November; PDF pages 25–29 on realization flows and behavioral ratios read.
Sequence form is classical; the present HC coefficient proof is explicit.

Targeted source assessment is not systematic historical-priority clearance.
Complete private review records and local-source locators remain separately
preserved; no private paths or internal identifiers belong in the public package.
