"""Build v3 standalone manuscript from retained v2 foundation and v3 evidence."""
from pathlib import Path
import json
ROOT=Path(__file__).resolve().parents[1]
s=(ROOT/'history/v2/MANUSCRIPT.md').read_text()
r=json.loads((ROOT/'results_v3/innovation_report.json').read_text());q=json.loads((ROOT/'results_v3/queries_report.json').read_text())
status=json.loads((ROOT/'STATUS.json').read_text())
s=s.replace('date: "19 September 2026 | Research reference release 2.0.0"','date: "19 September 2026 | Research reference release 3.0.0"')
s=s.replace('subtitle: "Residual-Corrected Persistent World Memory for Unified Neural Graphics"','subtitle: "Certified Innovation Rendering: Persistent Evidence That Removes Physical Sampling Work"')
start=s.index('# 1. Abstract');end=s.index('# 2. Central scientific claim')
s=s[:start]+r'''# 1. Abstract

AUREOLE-R v3 develops **Certified Innovation Rendering**: persistent world memory stores exact physical response together with the domain in which that response remains valid. Verified contributions are removed from the stochastic residual; a neural prior represents only the unresolved complement. This strengthens v2's fallible-memory control variates, which preserved expected linear output but could suffer severe variance after stale-memory sampling.

We prove a common covariance contraction. For an arbitrary frozen vector control and proposal $q$, replacing a set of proposal mass $a$ by its exact current contributions and renormalizing the remaining proposal gives $\Sigma'\preceq(1-a)\Sigma$. The result holds simultaneously for all positive-semidefinite metrics on linear task readouts. A causal sequential estimator assimilates each physical observation only after forming its current correction. Conservative geometric certificates determine the lifetime and spatial reach of exact facts. In a finite deterministic domain, physical query count is bounded by initial unknown terms, newly introduced terms and certificate invalidations, rather than necessarily by the number of displayed frames.

A new twelve-scene held-out study measures 98.40% lower conditional expected linear-RGB MSE during smooth motion than a global cache that resets on every geometry change (95% scene-bootstrap interval 98.12--98.75%). The new method uses 1.203 rather than 2.000 queries per receiver in that phase. It does not outperform that reset baseline after large jumps: the observed 0.15% regression has an interval spanning zero. Certificate construction and dense bookkeeping also make the current CPU implementation slower than the v2 contradiction guard.

On eight additional scenes, a frozen coarse state serves a finer receiver grid, five prescribed times and three known appearance readouts with 76.85% fewer physical queries than fresh shared visibility, including all initialization. The final linear outputs match an independent exact reference in the executed float64 tests; no invalid certificate is accepted in the supported studies. This is spatial visibility refinement and known-time relighting, not general learned SR or frame generation.

The package retains the full original architecture, fifteen earlier scoped results, trained 3,217-parameter prior, original negative evidence and raw experiments. It adds six proved propositions, sixteen executable tests and two new protocols. Classical control variates, visibility caches and kinetic certificates are acknowledged antecedents. The contribution candidate is their explicit validity-and-estimation contract, with falsifiable evidence for work reduction. General transport, cheap production certificates, joint SR/RR/FG, calibrated neural uncertainty and GPU superiority remain open. Scientific maturity is subjectively estimated at 68%; this is a meaningful executable research advance, not 100% completion of unified real-time neural graphics.

**Keywords:** certified innovation rendering; persistent world inference; visibility certificates; neural rendering; physical residual correction; covariance contraction; active sampling; world-space memory; query closure; temporal coherence; renderer co-design.

''' +s[end:]
# New claims are placed before the retained broad architecture.
loc=s.index('# 3. ')
s=s[:loc]+r'''## 2.1 What v3 changes

The most defensible advance is operational: **valid scene evidence deletes physical sampling work**. A record's age does not make it reliable. Its geometric validity domain does. One clearance certificate can support relighting, nearby spatial queries and known intermediate times; unresolved terms retain properly weighted physical correction. The neural prior can be poor without breaking the estimator identity, while a false certificate can break it immediately.

Sections 4--19 retain the broad world-state theory and all earlier proofs. Section 19.9 gives the six new complete arguments. Sections 29.7--29.9 report the new experiments and adverse findings. Historical v1/v2 numbers are explicitly labeled. The original SR/RR/FG ambition is preserved as a research objective, not represented as an executed integrated system.

''' +s[loc:]
loc=s.index('# 4. Formal problem statement')
s=s[:loc]+r'''## 3.1 Closest antecedents for the v3 mechanism

Adaptive Quantization Visibility Caching [R19] and Progressive Visibility Caching [R20] already cache and share visibility queries. Kinetic data structures [R21] provide an established certificate-based view of moving geometry. Adaptive primary-space control variates [R22] already combine approximate integration with unbiased residual sampling. The 2026 3D transport embedding [R10] is a substantially broader learned scene representation than this small direct-light implementation.

Accordingly, neither world memory, geometric certificates, exact visibility reuse, neural residual correction, nor event-driven work reduction is claimed as a standalone invention. C1--C6 make a particular combination explicit and auditable: geometric validity licenses zero residual support; support removal contracts covariance for a shared linear readout; sequential acquisition consolidates evidence without reusing a draw in its own prediction. Novelty beyond this synthesis remains uncertain. Existing world-space visibility and transport caches must be strong baselines in a production evaluation.

The v3 search read primary author pages for [R19--R20] and the primary abstract for [R22]. The large [R19] PDF could not be fetched; [R21] was available only through publisher/government-index search metadata. These access limits are recorded in references.json. The new proofs are self-contained and do not depend on uninspected arguments in those papers.

''' +s[loc:]
# Embed the full new mathematical document, converting it to subordinate sections.
proof=(ROOT/'docs/INNOVATION_THEORY.md').read_text();proof=proof[proof.index('All propositions'):]
proof=proof.replace('\n## ','\n### ')
loc=s.index('# 20. Complexity')
s=s[:loc]+'## 19.9 Certified Innovation Rendering: six new results\n\n'+proof+'\n\n'+s[loc:]
loc=s.index('# 21. GPU implementation')
s=s[:loc]+r'''## 20.2 Executed v3 costs and the certificate bottleneck

The v3 dense memory holds 2,048 receiver addresses, 36 emitters, float32 Boolean evidence, float64 clearance and float64 ledger stamps. It uses 1,524,672 bytes including endpoint and geometry arrays, versus 589,824 bytes for v2's value/epoch tables. The added state is useful provenance, but it is not free compression. A direct 1080p dense extrapolation of the evidence arrays alone would approach 1.5 GB; sparse allocation is mandatory before game integration.

During smooth motion, median measured batch time is approximately 1.88 ms for sequential certificates, 1.97 ms for global-reset elimination, 1.30 ms for certificate iid sampling and 1.18 ms for the v2 guard. These batches contain only 320 receivers. The reported times include certificate construction and validation but exclude common neural-prior preparation, analytic coefficient preparation and exact-reference auditing. They are not end-to-end frame times. The release establishes query and error reductions, not a general speedup.

Certificate creation is a central unresolved cost. The reference evaluates minimum distance to a segment for all three spheres. A hardware shadow ray normally need not return such a margin and can terminate early. General triangle geometry requires conservative BVH distance bounds, primitive distance computation, changed-object dependency tests or another validated adapter. Treating the certificate as free would make the proposed architecture's economics misleading.

A useful deployment test is $C_{\mathrm{validation}}+C_{\mathrm{new\ queries}}+C_{\mathrm{decode}}<C_{\mathrm{baseline}}$ at matched image quality, while meeting p99 deadlines. Fewer rays alone do not establish this inequality. Tightening the global motion ledger to local dependencies is a justified next optimization; it is not part of the measured implementation.

''' +s[loc:]
loc=s.index('# 24. Dataset')
s=s[:loc]+r'''## 23.2 Executed v3 loop

```text
receive authoritative geometry, light coefficients and canonical query IDs
advance the conservative motion ledger
look up exact values whose clearance still covers this query
integrate certified terms exactly; predict the unresolved complement
restrict the sampling proposal to that unresolved complement
for each of the fixed query slots:
    select an unresolved term using the current restricted proposal
    query its physical value and clearance
    form the unbiased correction before assimilating this observation
    mark this term exact and remove it from the next proposal
average the causal corrections, or return the exact sum on full completion
serve other legal linear readouts from the same valid visibility evidence
```

`aureole/certificates.py` and `aureole/innovation.py` implement this boundary.
The API trusts the physical adapter's facts; it is not a theorem prover that
can detect arbitrary engine omissions. Sampling with zero support for an
uncertified residual violates the contract. Full details are in
`docs/CERTIFICATE_API.md`.

''' +s[loc:]
loc=s.index('# 30. Second-order')
new=r'''## 29.7 E10: persistent certificates under known geometry motion

This protocol was fixed before its first run and uses twelve new scenes (400--411), three sampling replicates, 2,048 canonical receivers, a moving 320-receiver viewport and 36 emitters. It includes cold and warm phases, a 500-tick unobserved gap, revisit, relighting, sixteen small-motion frames, a large jump and a return. Every method has at most two physical queries per receiver per frame. Unused queries are counted as savings and are not reassigned. A certificate query includes additional clearance arithmetic.

Unlike E8/E9, the geometry changes are exposed authoritatively. The global-reset baseline uses the same geometry to invalidate its entire visibility cache whenever any sphere changes. It then uses the same sequential residual estimator. The comparison isolates local lifetime from whole-cache reset; it is stronger than comparing only with v2's post-contradiction guard but weaker than a mature locality-aware visibility-cache baseline.

| Method | Smooth-motion MSE | Queries/receiver | Jump MSE | Median smooth batch |
|:--|--:|--:|--:|--:|
'''
for method,label in [('v2_guarded','v2 guard'),('epoch_eliminate','Global reset'),('certificate_iid','Certificate iid'),('certificate_eliminate','Certificate elimination'),('constant_certificate','Constant-prior certificates')]:
 sm=next(x for x in r['summary'] if x['method']==method and x['phase']=='smooth_motion');jump=next(x for x in r['summary'] if x['method']==method and x['phase']=='jump')
 new+=f"| {label} | {sm['expected_mse']:.7f} | {sm['rays_per_receiver']:.3f} | {jump['expected_mse']:.7f} | {sm['cpu_median_ms']:.2f} ms |\n"
new+=r'''
The primary expected-MSE reduction is 98.40% versus global reset, with a 95% scene-bootstrap interval of 98.12--98.75%. Versus the v2 guard, the smooth-motion reduction is 96.08% (95.28--96.98%). These intervals resample twelve independent scene clusters, retaining their sampling replicates. They describe this generator and these prescribed motions.

The sequential method improves expected MSE by 24.44% over certificate iid sampling in the smooth phase (23.05--26.26%), at higher CPU cost. It also learns each term at most once until expiry. Its prior-and-proposal ablation against a constant 0.5 prior gives 8.02% lower error (4.39--13.11%) during smooth motion. This measures the prior and the proposal it induces together; it does not prove that a neural world model is necessary for the main certificate gain.

Large jumps revoke most certificates. Versus global reset, the new method's jump MSE is 0.15% higher, with reduction interval [-0.43%, 0.08%]. Thus the experiment does not establish an advantage in that phase. The return phase also exposes the conservatism of cumulative motion: returning to an old scene does not automatically recover all old certificates.

![Known-motion study. All methods have the same maximum ray budget; actual query counts and computation differ. Error is conditional expected linear RGB MSE.](figures_v3/innovation_results.png){width=95%}

## 29.8 E11: one state for spatial, temporal and appearance queries

A second protocol uses eight further scenes (600--607). It fully observes a 16 by 32 anchor grid with 36 emitters, then freezes that memory. A 32 by 64 query grid requests visibility at five prescribed times along a known affine geometry path. Certificates transfer from the nearest anchor only if their clearance covers both endpoint displacement and geometry motion. Every unresolved query is traced exactly. Each reconstructed visibility field serves three known appearance readouts. No future observations are assimilated into the frozen anchor state.

The result uses 147,456 initialization queries and 535,303 residual queries, totaling 682,759. Fresh visibility shared across the same three readouts would use 2,949,120 queries. Savings including initialization are 76.85%; after initialization alone they are 81.85%. Comparing to three separately traced readouts would exaggerate the gain, so the stronger shared-visibility baseline is the headline. Independent reference evaluation incurs another 2,949,120 audit queries, outside the policy budget.

No accepted certificate disagrees with the independent quadratic intersection oracle, and maximum final linear-RGB difference is zero in these executed float64 computations. This is finite-scene experimental agreement, not a formal numerical proof. Texture and material coefficients are engine-known. The test does not recover unknown high-frequency textures, antialias a camera, infer unknown future motion, or train SR/FG networks. It demonstrates a narrow but actual common response state serving new points, known times and relighting.

![Response to geometry changes and total query count in the independent shared-readout study. Certificate initialization is included in the right panel.](figures_v3/innovation_dynamics.png){width=95%}

![Three known appearance readouts at three of the prescribed query times, from one frozen coarse memory plus exact unresolved queries. Display previews apply gamma; metrics use linear outputs.](figures_v3/query_readouts.png){width=95%}

## 29.9 Negative controls, scope and evidence inventory

An unsafe elimination ablation ignores geometry changes and treats every old observation as exact. During the jump phase it reuses 1,133,574 false visibility entries across all audited frame-method records, with expected MSE 0.0067465. The supported certificate methods have zero false accepted entries in the executed studies. This difference demonstrates why exact-support deletion needs a validity mechanism. If engine changes are withheld, a unit-test counterexample shows that the safe algorithm's premise also fails; certificates do not sense hidden reality by themselves.

The new motion study records 18,144 frame-method observations and 8,202,625 online physical segment queries across its seven methods. The independent query-family study adds 682,759 initialization and residual queries. The two new studies therefore execute 8,885,384 physical segment queries, in addition to the retained v2 experiments. Full reference audits, neural-prior preparation and analytic appearance calculations are extra work, not hidden policy inputs. Their exclusion from reported online ray budgets must not be confused with total authoring or reference-render cost.

Sixteen new tests cover matrix covariance contraction, exact ordered-pair estimator enumeration, complete-state zero-query behavior, finite completion, wrong zero support, geometry validity, spatial transfer, namespace/checkpoint checks, output enclosures, return-motion conservatism and an unreported-change failure. Together with 28 v2 and 14 foundational checks, the release has 58 executable tests. A release validation script checks current files, raw-summary consistency, split separation, scope and reproducibility. It is not a production-readiness certificate.

'''
s=s[:loc]+new+s[loc:]
loc=s.index('# 31. Practical roadmap')
s=s[:loc]+r'''## 30.2 Second-order advance: a budget for physical innovations

C4 offers a more fundamental objective than increasing generated frame count. For a finite persistent query domain, expensive work can scale with newly unresolved scene distinctions and certificate expiry. One physical query can establish a response over a spatial and temporal validity region, after which many readouts are deterministic operations on the common state. The model should choose experiments partly by how much future physical work their *valid conclusions* eliminate.

This is only useful if certificate production, retention and validation cost less than the avoided work. A triangle/BVH adapter and matched-time evaluation are therefore more decisive than a larger neural predictor. The next architectural experiment should compare three systems with equal inputs and total GPU budget: a strong existing visibility cache, that cache with neural residual correction, and the proposed joint validity-and-elimination interface. Failure to exceed the second baseline would mean that the contribution is an effective integration of known parts, not a new general rendering paradigm.

The final conceptual pass removes unnecessary confidence machinery from exact deterministic facts: storing a posterior variance of zero is weaker operationally than storing the condition under which the fact is exact. Uncertain transport still needs a belief and physical correction. A good state can contain both, but they cannot share an undifferentiated trust flag.

''' +s[loc:]
loc=s.index('# 33. Standalone conclusion');end=s.index('# References and source provenance')
conclusion=r'''# 33. Standalone conclusion and research status

The v3 advance is **Certified Innovation Rendering**: persistent scene observations carry validity domains, and valid facts remove terms from the physical sampling problem. The covariance-contraction result explains why this can improve several linear readouts at once even with a fallible neural prior. The sequential estimator turns new rays into consolidated knowledge without same-sample fitting bias. The finite-domain work bound connects cost to unresolved evidence and invalidation events.

The implemented result is substantial within its narrow setting: much lower error under small exposed geometry motion, fewer physical queries on revisits, and exact tested spatial/time/appearance readouts with initialization-inclusive query savings. Equally material limits remain: known geometry is required, large jumps erase most reuse, certificate arithmetic and memory have costs, and the current CPU implementation is slower than v2. The core graphics hypothesis is still unvalidated for full transport and nonlinear SR/RR/FG.

The research therefore advances as a meaningful, reproducible reference. Calling it a final universal breakthrough, 100% mature, or an immediate DLSS replacement would overstate the evidence. A major scientific classification requires independent matched-time results against strong world-space caches and multiple rendering tasks. A transformative classification requires those benefits to survive diverse engines and content.

## Research status and completeness

Percentages are subjective estimates of progress toward the original objective, not measured completion probabilities. Scoped test acceptance is separate. The overall estimate is 68%, compared with 65% in v2; it is not an arithmetic average of the area estimates.

| Area | Estimate | Genuinely established | Still open |
|:--|--:|:--|:--|
'''
labels={'mathematical_core':'Mathematical core','latent_state_theory':'Latent-state theory','world_memory':'World memory','temporal_dynamics':'Temporal dynamics','unified_tasks':'Unified tasks','physics_grounding':'Physics grounding','active_sampling':'Active sampling','real_time_feasibility':'Real-time feasibility','experimental_readiness':'Experimental readiness','novelty_confidence':'Novelty confidence','architecture':'Architecture'}
for a in status['areas']:
 conclusion+=f"| {labels[a['area']]} | {a['completeness_estimate_percent']}% | {a['established']} | {a['remaining']} |\n"
conclusion+='\n**Overall classification: meaningful. Overall maturity estimate: 68%.** Twenty-one scoped mathematical results and a usable CPU reference exist. GPU integration, full task unification, strong production baselines and independent replication remain open. The public package is prepared for research use; external publication has not been performed.\n\n\\newpage\n\n'
s=s[:loc]+conclusion+s[end:]
for ref in json.loads((ROOT/'references.json').read_text()):
 if ref['id'] in ('R19','R20','R21','R22'):
  s+=f"\n**[{ref['id']}]** {ref['citation']} [Primary source]({ref['url']}). Access: {ref['access']}\n"
s=s.replace('physical demo, fifteen scoped mathematical results, original and follow-up evidence.', 'physical demos, twenty-one scoped results, certificate elimination and all retained/new evidence.')
(ROOT/'MANUSCRIPT.md').write_text(s)
print('Assembled v3 standalone manuscript:',len(s.split()),'words')