The machine-readable results retain the controlled fixtures, selected pair constraints, optimization witnesses, development-only weights, uncertainty intervals and falsifying controls. The five PNG/SVG pairs and summary table are generated from those recorded results. Earlier study artifacts are not rewritten.
+
Post-certificate structural extensions for Jev graph synthesis
+
Timothy Wayne Gregg | AI-assisted author-review research extension | September 18, 2026; corrected September 28, 2026
+
Abstract
+
Five precommitted controlled extensions were executed after the dependence-aware certificate study. Recorded frozen-target outcomes: H1: met; H2: not met; H3: met; H4: met; H5: met. These are separate algorithmic conjunctions, not a pooled success percentage. The experiments address induced-width repair, transactional delta maintenance, query resilience, finite endpoint-scenario review and n-ary conflict constraints. H4 tests endpoint scenarios only; it does not establish minimax review over continuous probability intervals. An interior counterexample below rejects that broader interpretation. No new semantic-accuracy observations, deployment guarantees or external-system superiority are established.
+
1. Frozen protocol and evidence boundary
+
The protocol was committed as 8187cc4e4b98e40f3146ce787b71f262b62b3abd against baseline 9d2e4a60a6c79e616ab1d352cd5da9fe414e6c98 before implementation and execution. Its bytes, fixture definitions, caps, seed 20260923 and thresholds remain unchanged. Fresh Jev calls: 0. New scientific documents: 0. This is exploratory follow-up, not independent preregistration.
+
The correction record distinguishes repaired implementation/control defects from the preserved protocol. H4's frozen Method explicitly specifies interval endpoint vectors; its former broader interval-minimax description is withdrawn. The supplemental negative control is outside the frozen 128-fixture population. Treewidth-aware dynamic programming, incremental view maintenance, query resilience, finite-scenario minimax decisions and hypergraph optimization have prior art; these are repository experiments, not worldwide novelty claims.
+
+
+
+
Hypothesis
+
Frozen target
+
Recorded outcome
+
+
+
+
+
H1: induced-width repair
+
zero oracle/permutation failures; solve 8 large chains; stage controls; >=90% fewer counted states
+
met
+
+
+
H2: transactional delta
+
zero snapshot/atomicity failures; >=80% fewer local evaluations
+
not met
+
+
+
H3: query resilience
+
zero oracle/witness failures; solve 8 analytic families; stage 6 controls
+
met
+
+
+
H4: endpoint-scenario review
+
zero endpoint-oracle discrepancies; never worse than midpoint on endpoint scenarios; strict gain on >=20%
+
met, endpoint scenarios only
+
+
+
H5: n-ary conflicts
+
zero oracle/permutation failures; solve 8 chains; strict gain vs pairwise projection; stage controls
+
met
+
+
+
+
A failed conjunction remains not met regardless of individual favorable endpoints.
+
2. H1 - induced-width exact conflict optimization
+
The solver computes deterministic min-fill order from the actual graph, checking width and allocation limits before constructing exponential tables. The shortcut based on numeric identifier distance has been removed. Of 160 small fixtures, 160 match the exhaustive objective; recorded failures: 0; permutation failures: 0. Width distribution: {'1': 33, '2': 45, '3': 53, '4': 29}.
+
For the frozen 12-16 vertex subset, counted elimination evaluations are 14,660, versus 1,986,560 full assignments: 99.26% reduction. This original work counter is not total memory allocation or CPU time; a separate cumulative table-allocation guard enforces the frozen cap. Large triangle chains use an independent prefix dynamic-programming oracle. Per-fixture objectives, widths and staged controls remain in results.json.
+
+
3. H2 - transactional batch delta maintenance
+
Transactions validate revisions and updates, evaluate an overlay in computed topological order, and return it only after successful evaluation. Unknown dependencies and cycles are rejected. Across 640 accepted transactions, independent full recomputation finds 0 snapshot mismatches. A control passes only when the expected exception is raised and original values/revision remain unchanged: injected evaluation 63/64, stale revision 64/64, malformed update 64/64. Returned revision mismatches: 0. The frozen injected-failure conjunction is not met: one selected primitive has no affected derived node, so no failure was injected. That case remains in the denominator; the exact control record is retained in results.json.
+
Local evaluations: 16,205, versus 163,840 for full recomputation: 90.11% reduction. Connected controls separately show 0.00% mean saving. These are in-memory atomic batches, not crash durability or concurrent database isolation.
+
+
4. H3 - exact query-resilience certificates
+
For supplied monotone DNF lineage and removal costs, incidence decomposition and bounded branch-and-bound find a minimum-cost deletion set hitting each active proof clause. 160/160 small fixtures match the exhaustive oracle. Recorded oracle, witness or control failures: 0. All 8 analytic-family results and 6 over-cap controls are retained. This is conditional structural resilience, not factual truth or calibrated review effort.
+
+
5. H4 - endpoint-scenario query review
+
The selector minimizes worst-case expected residual Bernoulli variance over the finite Cartesian product of interval endpoints, with two reviews. Production uses Shannon recursion; the independent checker enumerates Boolean worlds over the same endpoint scenarios. Neither searches interval interiors.
+
Across 128 frozen fixtures, endpoint-oracle discrepancies: 0; cases worse than midpoint on endpoint scenarios: 0. Strict endpoint gains: 45/128 (35.16%), comprising 13 random and 32 engineered fixtures. Point-interval controls pass 16/16.
+
Continuous-interval claim: unsupported. For singleton queries with intervals 0=[0.1,0.9], 1=[0.3,0.4], 2=[0.5,0.5], the supplemental control selects reviews [1, 2] and reports endpoint worst loss 0.09. The allowed interior witness p0=0.5 yields 0.25. Reviewing [0, 2] has analytic continuous worst loss 0.24. The endpoint choice is therefore not continuous-interval minimax. This negative evidence is separate from the preserved endpoint target. Independent primitives are assumed; no dependence or semantic-calibration guarantee follows.
+
+
6. H5 - n-ary conflict constraints
+
A forbidden hyperedge rejects only its all-selected assignment; pairwise conflict is the arity-two case. 160/160 small fixtures match exhaustive optimization; recorded failures: 0; permutation failures: 0. Eight chains compare against n - floor(n/3), with objectives and staging controls retained. At the largest chain, retained n-ary utility is 1366, versus 683 under deliberately over-conservative pairwise projection.
+
+
7. Interpretation
+
Every result is conditional on supplied graphs, lineage, costs, constraints or finite probability scenarios. Work counters do not measure wall-clock/API latency or cloud cost. The H4 interior counterexample rejects the broader interval-minimax interpretation. None establishes correct extracted entities, relations, qualifiers or probabilities. No production graph policy changes here.
results.json retains outcomes and controls; summary.csv and five SVGs derive from those measurements. Manual validation also checks protocol ancestry/bytes, archive integrity, deterministic HTML and current-paper build hashes. The correction record documents implementation and interpretation changes; the frozen protocol is unchanged.
Do additional Jev calls improve graph edges?
Timothy Wayne Gregg — exploratory research addendum, September 18, 2026
Design and evidence boundary
@@ -4455,7 +4532,7 @@
Manuscript status
This is a detailed empirical manuscript draft grounded in the completed run-20260918 artifacts. It is suitable for author review and development into a paper, not a claim of submission readiness. Author names, affiliations, contributions, funding, conflicts of interest, venue format, and final data-distribution permissions require the authors' confirmation. No independent replication, new adjudicated dataset, or additional experiment is implied by packaging the study.
Typed semantic decision services offer bounded outputs, but the quality and repeatability of those outputs must be measured on the intended task. We investigate whether changes to request formulation improve a fixed Jev model without replacing or fine-tuning it. A locally frozen protocol compares four formulations per task, selects one alternative using 60 training-derived development examples, fits scalar temperature calibration on a separate split, and evaluates the retained alternative against a corrected baseline. On 413 identity-disjoint DBLP–ACM bibliographic pairs, a bundled intervention comprising explicit identity instructions, a typed Choice contract, and six training demonstrations raises operational macro-F1 from 0.9605 to 0.9859. The paired 95% percentile bootstrap interval for the difference is [0.0010, 0.0540]. Predictions of identity on different-entity pairs fall from 8/63 to 2/63, while missed matches increase from 0/350 to 1/350. On 339 SciFact claim–cited-abstract examples, selected-formulation macro-F1 changes from 0.8508 to 0.8527, with an interval spanning zero. Calibration has task- and metric-dependent effects. Cached replay reconstructs all 3,644 predictions exactly; three fresh calls on 20 fixed examples per task preserve selected-formulation labels while probabilities vary. The study uses 3,405 HTTP calls and 4.79 million input tokens, with approximately $0.20 estimated provider charges. These results support further independent evaluation of the entity-matching formulation, not general Jev superiority, a causal claim about demonstrations alone, or a guarantee of safe graph mutation.
diff --git a/manuscript/paper-current.md b/manuscript/paper-current.md
index 7aca305..eb49251 100644
--- a/manuscript/paper-current.md
+++ b/manuscript/paper-current.md
@@ -22,6 +22,7 @@ Each entry links to its study and its in-document protocol or evidence boundary.
| [Structural frontiers for evidence-preserving Jev graph synthesis](#structural-frontiers-for-evidence-preserving-jev-graph-synthesis); [Protocol and scope](#1-motivation-novelty-scope-and-frozen-design) | Controlled algorithms and decision models | Review misses; four structural targets meet. | 2026-09-18 |
| [Dependence-aware certificates for Jev graph synthesis](#dependence-aware-certificates-for-jev-graph-synthesis); [Protocol and scope](#1-motivation-novelty-boundary-and-frozen-methodology) | Saved responses; controlled algorithms | Forecasting misses; four controlled targets meet, with documented overlap. | 2026-09-18 |
| [Five post-certificate improvements for Jev graph synthesis](#five-post-certificate-improvements-for-jev-graph-synthesis); [Protocol and scope](#1-frozen-methodology-and-novelty-boundary) | Saved responses; controlled algorithms | Source-risk stacking misses; four controlled targets meet. | 2026-09-18 |
+| [Post-certificate structural extensions for Jev graph synthesis](#post-certificate-structural-extensions-for-jev-graph-synthesis); [Protocol and scope](#1-frozen-protocol-and-evidence-boundary) | Controlled algorithms; finite endpoint scenarios | H2 misses an injected-failure control; H4 passes only its endpoint-scenario target, not continuous interval minimax. | 2026-09-28 |
| [Do additional Jev calls improve graph edges?](#do-additional-jev-calls-improve-graph-edges); [Protocol and scope](#design-and-evidence-boundary) | 3,024 fresh Jev calls | No added-call policy establishes a matched-volume advantage. | 2026-09-18 |
| [Fresh Jev execution and graph-synthesis falsification](#fresh-jev-execution-and-graph-synthesis-falsification); [Protocol and scope](#evidence-scope) | Fresh same-data Jev repeat; synthetic challenge | Original entity improvement is unresolved on fresh macro-F1 interval. | 2026-09-18 |
| [Follow-up: five evidence-driven improvements](#follow-up-five-evidence-driven-improvements); [Protocol and scope](#summary-of-fixed-operational-targets) | Saved responses; controlled algorithms | Mixed target outcomes; no independent semantic validation. | 2026-09-18 |
@@ -1714,6 +1715,85 @@ The machine-readable results retain the controlled fixtures, selected pair const
+
+
+# Post-certificate structural extensions for Jev graph synthesis
+
+Timothy Wayne Gregg | AI-assisted author-review research extension | September 18, 2026; corrected September 28, 2026
+
+## Abstract
+
+Five precommitted controlled extensions were executed after the dependence-aware certificate study. Recorded frozen-target outcomes: **H1: met; H2: not met; H3: met; H4: met; H5: met.** These are separate algorithmic conjunctions, not a pooled success percentage. The experiments address induced-width repair, transactional delta maintenance, query resilience, finite endpoint-scenario review and n-ary conflict constraints. **H4 tests endpoint scenarios only; it does not establish minimax review over continuous probability intervals.** An interior counterexample below rejects that broader interpretation. No new semantic-accuracy observations, deployment guarantees or external-system superiority are established.
+
+## 1. Frozen protocol and evidence boundary
+
+The protocol was committed as `8187cc4e4b98e40f3146ce787b71f262b62b3abd` against baseline `9d2e4a60a6c79e616ab1d352cd5da9fe414e6c98` before implementation and execution. Its bytes, fixture definitions, caps, seed `20260923` and thresholds remain unchanged. **Fresh Jev calls: 0. New scientific documents: 0.** This is exploratory follow-up, not independent preregistration.
+
+[The correction record](../graph_synthesis/post_certificate/CORRECTIONS.md) distinguishes repaired implementation/control defects from the preserved protocol. H4's frozen Method explicitly specifies interval endpoint vectors; its former broader interval-minimax description is withdrawn. The supplemental negative control is outside the frozen 128-fixture population. Treewidth-aware dynamic programming, incremental view maintenance, query resilience, finite-scenario minimax decisions and hypergraph optimization have prior art; these are repository experiments, not worldwide novelty claims.
+
+| Hypothesis | Frozen target | Recorded outcome |
+|---|---|---|
+| H1: induced-width repair | zero oracle/permutation failures; solve 8 large chains; stage controls; >=90% fewer counted states | **met** |
+| H2: transactional delta | zero snapshot/atomicity failures; >=80% fewer local evaluations | **not met** |
+| H3: query resilience | zero oracle/witness failures; solve 8 analytic families; stage 6 controls | **met** |
+| H4: endpoint-scenario review | zero endpoint-oracle discrepancies; never worse than midpoint on endpoint scenarios; strict gain on >=20% | **met**, endpoint scenarios only |
+| H5: n-ary conflicts | zero oracle/permutation failures; solve 8 chains; strict gain vs pairwise projection; stage controls | **met** |
+
+A failed conjunction remains not met regardless of individual favorable endpoints.
+
+## 2. H1 - induced-width exact conflict optimization
+
+The solver computes deterministic min-fill order from the actual graph, checking width and allocation limits before constructing exponential tables. The shortcut based on numeric identifier distance has been removed. Of 160 small fixtures, 160 match the exhaustive objective; recorded failures: **0**; permutation failures: **0**. Width distribution: {'1': 33, '2': 45, '3': 53, '4': 29}.
+
+For the frozen 12-16 vertex subset, counted elimination evaluations are **14,660**, versus **1,986,560** full assignments: **99.26% reduction**. This original work counter is not total memory allocation or CPU time; a separate cumulative table-allocation guard enforces the frozen cap. Large triangle chains use an independent prefix dynamic-programming oracle. Per-fixture objectives, widths and staged controls remain in `results.json`.
+
+
+
+## 3. H2 - transactional batch delta maintenance
+
+Transactions validate revisions and updates, evaluate an overlay in computed topological order, and return it only after successful evaluation. Unknown dependencies and cycles are rejected. Across **640** accepted transactions, independent full recomputation finds **0 snapshot mismatches**. A control passes only when the expected exception is raised and original values/revision remain unchanged: injected evaluation **63/64**, stale revision **64/64**, malformed update **64/64**. Returned revision mismatches: **0**. The frozen injected-failure conjunction is not met: one selected primitive has no affected derived node, so no failure was injected. That case remains in the denominator; the exact control record is retained in `results.json`.
+
+Local evaluations: 16,205, versus 163,840 for full recomputation: **90.11% reduction**. Connected controls separately show **0.00% mean saving**. These are in-memory atomic batches, not crash durability or concurrent database isolation.
+
+
+
+## 4. H3 - exact query-resilience certificates
+
+For supplied monotone DNF lineage and removal costs, incidence decomposition and bounded branch-and-bound find a minimum-cost deletion set hitting each active proof clause. **160/160** small fixtures match the exhaustive oracle. Recorded oracle, witness or control failures: **0**. All 8 analytic-family results and 6 over-cap controls are retained. This is conditional structural resilience, not factual truth or calibrated review effort.
+
+
+
+## 5. H4 - endpoint-scenario query review
+
+The selector minimizes worst-case expected residual Bernoulli variance **over the finite Cartesian product of interval endpoints**, with two reviews. Production uses Shannon recursion; the independent checker enumerates Boolean worlds over the same endpoint scenarios. Neither searches interval interiors.
+
+Across 128 frozen fixtures, endpoint-oracle discrepancies: **0**; cases worse than midpoint on endpoint scenarios: **0**. Strict endpoint gains: **45/128 (35.16%)**, comprising 13 random and 32 engineered fixtures. Point-interval controls pass **16/16**.
+
+**Continuous-interval claim: unsupported.** For singleton queries with intervals `0=[0.1,0.9]`, `1=[0.3,0.4]`, `2=[0.5,0.5]`, the supplemental control selects reviews [1, 2] and reports endpoint worst loss 0.09. The allowed interior witness `p0=0.5` yields **0.25**. Reviewing [0, 2] has analytic continuous worst loss **0.24**. The endpoint choice is therefore not continuous-interval minimax. This negative evidence is separate from the preserved endpoint target. Independent primitives are assumed; no dependence or semantic-calibration guarantee follows.
+
+
+
+## 6. H5 - n-ary conflict constraints
+
+A forbidden hyperedge rejects only its all-selected assignment; pairwise conflict is the arity-two case. **160/160** small fixtures match exhaustive optimization; recorded failures: **0**; permutation failures: **0**. Eight chains compare against `n - floor(n/3)`, with objectives and staging controls retained. At the largest chain, retained n-ary utility is **1366**, versus **683** under deliberately over-conservative pairwise projection.
+
+
+
+## 7. Interpretation
+
+Every result is conditional on supplied graphs, lineage, costs, constraints or finite probability scenarios. Work counters do not measure wall-clock/API latency or cloud cost. The H4 interior counterexample rejects the broader interval-minimax interpretation. None establishes correct extracted entities, relations, qualifiers or probabilities. No production graph policy changes here.
+
+## 8. Reproducibility
+
+```bash
+python -B -m unittest discover -s graph_synthesis/post_certificate/tests -v
+python -B -m graph_synthesis.post_certificate.run --check
+python -B -m graph_synthesis.post_certificate.report --check --update-paper
+```
+
+`results.json` retains outcomes and controls; `summary.csv` and five SVGs derive from those measurements. Manual validation also checks protocol ancestry/bytes, archive integrity, deterministic HTML and current-paper build hashes. The correction record documents implementation and interpretation changes; the frozen protocol is unchanged.
+
+
# Do additional Jev calls improve graph edges?
Timothy Wayne Gregg — exploratory research addendum, September 18, 2026
@@ -2429,3 +2509,5 @@ A fixed Jev service benefits from a selected bundle of explicit identity instruc
# References
The package bibliography supplies the citation entries used above. The source-verification notes distinguish interface documentation, original papers, dataset provenance, and remaining licensing or metadata questions.
+
+
diff --git a/manuscript/paper-current.pdf b/manuscript/paper-current.pdf
index fb4dd45..541adcd 100644
Binary files a/manuscript/paper-current.pdf and b/manuscript/paper-current.pdf differ
