FEP_Lean v1.5.0 — 168-topic catalogue, finite information program and streamlined release gate

Release commit: 195d4e9c340e588534603780a7afb1f470a9d501 · hosted CI run green on that exact commit (Python suite, 15-cell distribution matrix on Linux/macOS/Windows × Python 3.10–3.14, Lean build/serial-Lean suite, accepted manuscript render). Local: 2,578 tests passed (91.95% coverage) and 521 serial-Lean tests passed.

Assets: wheel and sdist built from the release commit (uv build); fep_lean-manuscript-1.5.0.pdf is the CI-accepted render for the release commit (hash matches its render artifact manifest); SHA256SUMS covers all three.

Evidence boundary: this release changes no scientific claim. Not established here: the frozen H3 primary study (seeds unopened), full seven-stage publication capture and archive parity, the installed-wheel runtime matrix, Q7 generated-runner execution, and current Hermes/OpenGauss provider runs. These remain open research lanes in TODO.md. Concept DOI: 10.5281/zenodo.19699233; the version DOI is minted by Zenodo from this release.

---

1.5.0 — 2026-10-06 — 168-topic catalogue, finite information program and streamlined release gate

First published release since v1.3.0 (the 1.4.0 candidate prepared on
2026-10-02 was never tagged; its content ships here). The catalogue grows from
159 to 168 canonical topics across 22 families (waves 3 and 4), with finite
information identity/coarsening, relative-support information and genuine
finite rate–distortion results, the Q7 source-pair acceptance, publication
capture with seven deadline-bounded evidence stages, render custody, and
cooperative process supervision. Citation metadata describes 168 canonical
topic bodies and separates compiled statements from semantic adequacy and
empirical validity. Lean/Mathlib stays at the v4.34.1 pair.

Streamlined release gate (2026-10-06)

A versioned release now requires exactly: one version and date across every
version-bearing file, a dated CHANGELOG.md section, a clean main equal to
origin/main, and a successful hosted ci.yml run on that exact commit
(Python, distribution matrix, Lean and render jobs). docs/release_check.py
checks all of it and extracts these notes; docs/release.md
is the procedure. Research and evidence lanes — the frozen H3 study, full
seven-stage capture and archive parity, the installed-wheel runtime matrix,
Q7 runner execution, provider (Hermes/OpenGauss) runs and the nine-criterion
audit — are tracked in TODO.md and reported in release notes, but no longer
block a release. No scientific or acceptance claim is upgraded by this change:
H3 primary seeds remain unopened and generated-runner execution unverified.

The hash-bound H2.7-R0 custody record pins the root release token to its
recorded 1.3.0 → 1.4.0 transition. Its validator and readiness tests stay
byte-identical; a test fixture validates a copy of the custody inputs with the
approved 1.5.0 token mapped back to 1.4.0, so every other bound byte must
still match and any other version or duplicate root still fails.

The CI render job fetches the pinned GNU FreeSerif archive from
ftp.gnu.org or two identical mirrors and accepts only its SHA-256, after an
unreachable ftp.gnu.org failed the render job.

Package, browser and process-supervision repairs (2026-10-06)

Browser interaction validation now uses the canonical 118 theorem pairings.
The mobile dashboard fits the unchanged two-viewport height limit on Linux as
well as macOS: tighter summary padding, a smaller mobile title and side-by-side
family/status filters take the 390×844 document from 1,705 px to 1,538 px on
macOS. Hosted Linux Chrome had measured 1,727 px against the 1,688 px limit, so
the earlier padding-only fix failed there. Summary touch targets stay at
63 px. Worker-level controls protect Chrome and Lean grouping under
pytest-xdist. Historical browser artifacts are retained until fresh capture.

Synthetic custody fixtures execute byte-checked historical validator source in
an isolated module. The current validator continues to reject fabricated
historical evidence. Nine supplemental controls passed, including concurrent
real PDF and provenance reproducibility; the installed 80-case matrix remains
unchanged. These local results establish neither native nor release acceptance.

Cooperative process supervision now holds a leased caller's result until owned
descendant cleanup settles within the original shared cleanup deadline. A
child's handled timeout no longer fails its ancestor, guard-wait faults still
publish final cleanup state, and only a command whose guard had not reported is
treated as cleanly cancelled by its ancestor, so an undelivered terminal
response remains an obligation regardless of release order. Remote requests
skip an empty frame remainder that raised EPIPE against a rejecting peer.

CUR-02 and CUR-04 leave the open-only backlog after their bounded acceptance:
the original matrix still collects 80 cases, exactly nine separate additions
pass, formatted source has fresh independent review, and reconciled guidance
passes strict links, Markdown hygiene and cross-reference audits. Historical
receipts and substantive PR44/45 remain preserved. The backlog retains 30 open
items; package, current native, hosted and scientific gates remain separate.

Publication capture now refreshes the paired manuscript projections and test
census only after strict native acceptance, renders those final values, then
runs Python acceptance against the accepted render. The seven evidence stages
retain their deadlines and independent validators. Generated cache ownership
and the citation, configured image, manuscript and font/receipt input guards
are explicit; missing projection preparation fails before tools start.

CI render custody binds the checkout to the declared template revision and
records its exact uninitialized submodule entry separately from tracked blobs.
Only the pinned path and commit in a canonical empty directory are accepted;
unknown entries, initialization, aliases, population and metadata drift are
rejected before evidence can be retained. Git does not enter submodule stores.

The canonical backlog contains open work only. The
upcoming scope defines
minor, medium and major improvements for package and formal work with explicit
dependency, acceptance and failure boundaries.

Retained Q7 source-pair acceptance (2026-10-05)

The independently reviewed current Q7 observation
records all twelve stages and four closing checks passing under the original
limits. The actual root terminal closed exit 0 in 1,199.3769 seconds, and its
separate native receipt was claim-ready for its then-current 291-input roster,
selected isolated source pair and twelve static coefficient theorem reports.
All 124 pure cases, including twelve inventory controls, passed without skips.
Five fresh CPython parser/serializer probes returned the same 56,968-byte
canonical digest with all semantic/schema controls passing. These probes are
separate from native proof and generated-runner execution. Runner execution remains
unverified; Q5/Q6 observations and all failed/historical receipts keep their
original scope and bytes. Active GNN work remains untouched. Subsequent
browser-capture and dashboard owner changes make this observation historical;
a reviewed successor must establish currency for the new source pair.

After current-guidance read-only closure, FEP-Q7-CURRENT leaves the open-only
backlog. The seven-stage package capture, two identical independently validated
archives, local installed-wheel matrix, final-SHA hosted acceptance and complete
frozen H3 chain remain open. The 1.4.0 candidate remains unreleased.

Render deadlines, process ownership and current guidance (2026-10-05)

Renderer preparation, execution, normalization, artifact capture and cleanup
now share one monotonic deadline. Finite timeout validation and bounded regular
file capture reject malformed budgets, aliases and nonregular producer output.
Cooperative subprocess leases use distinct capabilities and explicit ownership
trees; closing a nested lease revokes and terminates its descendants while
preserving its caller and sibling groups. CI custody checks bind the physical
executable mode and descriptor-relative empty gitlink entry to the Git record.
Fresh independent infrastructure review approved these existing-owner changes.

Actual supervised local validation passed all 413 cases across the release
bundle, subprocess watchdog and distribution modules without skips on CPython
3.14.4. Mypy passed. The earlier two nested-ownership failures and subsequent
lint findings remain retained as failed attempts; the tests were preserved and
the ownership defect was repaired. Ruff lint/format, strict links, Markdown
hygiene, cross-references, lock and dependency compatibility checks passed.
These focused results do not establish
whole-suite, native, provider, package-bundle or scientific acceptance.

Maintained navigation now identifies current source and historical evidence
separately, removes obsolete planning references and completed checklists, and
uses the canonical open-only backlog. The active-guidance cleanup leaves that
backlog after strict link, Markdown and cross-reference checks. Minor, medium
and major follow-on scopes retain their dependencies and measurable probes.
The successful hosted checkpoint at c99e933 remains evidence for that source
epoch; later source changes require fresh same-SHA hosted acceptance.

The local wheel matrix retains its 15 GiB prerequisite before launch and every
cell. Current Q7/native capture, production capture, two independently accepted
identical archives and the frozen H3 export/review/outcome/reproduction chain
remain open. The later incomplete Q7 capture supplies no accepted prefix.
Licensed empirical data remain governed by the recorded no-go. Candidate
1.4.0 remains unreleased and retains its 2026-10-02 authored snapshot date.

Output/configuration repairs and reopened currency (2026-10-05)

The published main checkpoint
fa3c88e
has a successful
same-SHA hosted run:
all 15 distribution cells and the Python, Lean and render jobs passed.
Nonserial Python passed 2,371 tests with 15 policy skips in 829.74 seconds and
91.28% coverage on Linux CPython 3.14.8. Local validation uses the separate
CPython 3.14.4 environment.

Four existing owners now resolve selected output-root precedence,
checks of that output location, empty validation filters, read-only canonical
Hermes credential validation, and removal of a duplicate appendix writer.
The local CPython 3.14.4 focused run passed 298 cases with one explicit
live-provider skip. After correcting the valid-area intersection test,
the whole pipeline module passed all 37 cases without skips. Mypy, Ruff lint
and formatting, strict links/hygiene/xrefs, lock and environment checks passed.
Fresh independent infrastructure review approved the final source. Failed
attempts remain retained; these focused results are not full-suite, native,
provider or release acceptance.
Native-owner changes reopen the 291-input native and Q7 source-pair gates;
FEP-Q7-CURRENT returns to the open-only backlog. The old accepted Q7 captures,
failed attempts and fa3c88e CI results retain their source epochs without
receipt rewriting.

The local installed-wheel 15 GiB prerequisite remains unmet. Production capture,
two independently accepted identical package archives, once-only frozen H3
primary execution, claim review and study reproduction remain open. The empirical
branch remains governed no-go without licensed data. Candidate v1.4.0 is
unreleased and retains its canonical authored date of 2026-10-02.

Candidate publication and custody fixture repair (2026-10-03)

Published the reviewed candidate source to main at
9aaa30e
with exact remote parity. All 15 hosted distribution cells passed. The wider
Python job's failed custody/date checks remain recorded at that source epoch.

Disposable census fixtures now rebind all predecessor maps before testing
deliberate drift. Refresh composition tests use a fixture-root-scoped adapter
that validates rebased bindings and reports native execution as not_executed;
the real H2/H3 validators and historical receipts retain their exact bytes.
Malformed, missing, misbound and incorrectly typed inputs still reject.
All 224 cases across the four affected modules pass without skips; the release
metadata regression passes separately. Citation and sidecar dates agree with
the canonical authored candidate date. Full production and H3 acceptance and
the v1.4.0 release remain open.

Fresh Q7 source-pair closure (2026-10-02)

The fresh Q7 capture
passes all 12 newly executed stages and four closing checks in 1,249.5858
seconds under the unchanged 7,200-second acceptance bound. All 124 pure controls
pass without skips; five actual CPython runtimes produce the same 56,968
canonical scaffold bytes. The new native receipt has SHA-256
5630eef0b58185dfd359f627453e386174d25157ddcf935c928930c2d8eef7f0
and is independently validated as native-claim-ready for its exact checked
statements and source pair. Runner execution remains unverified. No historical
accepted prefix or native receipt is promoted, and active GNN work is preserved.

FEP-Q7-CURRENT leaves the open-only backlog. The former Q7 README heading is
explicitly historical; its exact capture-time preimage remains retained. These
postcapture guidance changes do not alter the 191 FEP or 291 native inputs.
Catalogue-native, full package/hosted acceptance, H3 outcomes, provider evidence
and the v1.4.0 release remain separate pending gates.

Main publication and hosted follow-up (2026-10-02)

Published reviewed improvements to main at
409ee71,
with exact local/direct-remote parity. Complete guarded Python acceptance
passed 2,308 tests and 90.65% coverage at that source epoch; all 25 declared
static checks and strict real-template render preparation passed separately.

The exact-SHA hosted run
finished with five Windows matrix failures, one Python runtime-fixture failure
and two H2 custody failures. All ten Ubuntu/macOS matrix cells passed. The
failed Lean job produced no native or accepted-render artifact. Follow-up
repairs and active guidance changes reopen dependent source-currency gates;
prior accepted observations retain their recorded epochs. Historical receipts,
failed attempts, frozen scientific protocol and active GNN work are preserved.
The open-only remaining acceptance
records the full original scope and execution order.

Q7 portability acceptance (2026-10-02)

The post-guidance observation
passes live bridge status, static generation and the exact new native receipt's
read-only validation after independent document review, with stable custody.
This closes FEP-SCAFFOLD-PORTABILITY. The accepted native JSON stays unchanged;
Q5/Q6 remain historical, and static parity does not imply runner execution.

Isolated Q7 evidence and active guidance refresh (2026-10-02)

The public Q7 observation
records isolated recapture r3 accepted in 644.5096 seconds at actual receipt
SHA-256
8c7c8d89023eb9784405143c1750b4b7ddd4ff1453dbbd089e9d660633b912ad.
Seven actual accepted r2 substages were freshly revalidated; their parent
attempt remains failed. R3 actually executes stages 8–12 and passes all 124 pure
controls without skips. The independently isolated GNN Git checkout binds all
703 owners to 536d949829f6aed11dc540e5c5dec77578b25016, with a private copied
runtime and read-only shared dependency base. Active GNN work/output is preserved.

Five actual CPython 3.10.20, 3.11.15, 3.12.13, 3.13.15 and 3.14.4 runtimes
produce 56,968 identical scaffold bytes at SHA-256
b34a350a0c66bd611c19cd87e2343592c6ee7d15ed2fd6e422b1f891597febec,
with all 11 controls per runtime. Evidence validation remains CPython 3.14.
The exact new native receipt copy
has SHA-256 b0ecf640fffa06f019d67c74a2d02e22715120165a98d218d61f49f98366bc3b,
is native-claim-ready and leaves runtime execution unverified.
Postcapture closure r2 accepts seven actual stages in 1,648.1618 seconds at
actual receipt SHA-256
df7787a925bf1011a9bd09a8bcc5ae20240792ea0b4ca9bc435eb5204225ba59.
All three real positive-axiom/wrong-F/wrong-Q native controls pass; the actual
outer process exits 0 with transport accepted. Failed attempts and the original
Q5/Q6/Q7 JSON receipts remain unchanged. Q5/Q6 observations remain historical
after W2 re-pinning; this Q7 result does not refresh them. Full operator-custody
records remain local; the public summary records observations, while the
separately validated official native receipt supplies the native claim.

Fourteen maintained Markdown files now distinguish baseline hosted evidence,
current native checkpoints and remaining acceptance. Their exact prior bytes,
second-read verification and manifest are retained in
specs/comprehensive-science-improvement/evidence/active-guidance-history-20261002-r3/.
Only README and REPORT change within the Q7 artifact-document map; this is not
a claim that the repository has only two changed files. Independent review and
three fresh read-only Q7 checks after this edit remain pending, so the
portability row stays open.

Wheel r7 remains historical after two guarded Q7 inputs changed and before this
guidance refresh. Its five cells each contain one actual installed target-runtime
case and 32 CPython 3.14 harness cases, 165 passes total. Wheel r8, hosted 15-cell
acceptance, canonical Python r2, render preparation r3, production capture r2,
two accepted identical archives and frozen H3 primary execution remain unrun.
The prior 2,303-pass, 12-skip, 90.66% Python run still fails its wider guard;
the private orchestrator repair has 11 focused passes. A projected 2,318-test
total is arithmetic only. Provider, release and H3/FORM-4 gates remain open.

Programmatic test isolation and current evidence (2026-10-01)

Programmatic catalogue tests now use the same private canonical project fixture
as direct pipeline tests. Redirecting output alone still allowed generated
manuscript source files to be replaced in the live checkout. The default-root
contract remains explicit, private projections are checked, and every test
brackets the live manuscript's bytes and metadata. All 11 focused controls pass.
The preceding canonical run's 2,303 passes and 90.66% coverage remain a rejected
wider-guard attempt; identical bytes do not erase observed file replacement.

Native r3 verifies all 168 topics without warnings or sorry; strict real
template render preparation r2 and five local installed-wheel runtime probes
pass at their recorded source scopes. H3 export r2 and three independent
proof-role reviews pass, with no scientific draws executed. Production capture,
complete guarded Python acceptance, hosted evidence and H3 outcomes remain open.
Completed historical delivery checklists were removed from the active ISA;
their exact prior text, scientific records and repository history are retained.
Active GNN registry work is preserved; isolated Q7 capture requires separate
runtime and root custody.

Render metadata ownership (2026-10-01)

The actual template render produced legitimate config.yaml, preamble.md
and references.bib copies that strict publication incorrectly rejected as
unexpected. Hydration now stages the exact raw metadata roster, capture declares
those outputs, and release archives require all three payloads with their own
evidence class. Missing, altered and linked copies still reject; chapter counts,
variable substitution and unknown-member rejection retain their contracts.
The focused acceptance
records 247 passing nonserial rendering/release tests, whole-tree type/style
checks and an unchanged 293-file source/test bracket. The original production
validation failure is retained. Native, installed-wheel and production evidence
must bind the resulting source epoch before current acceptance is claimed.

Finite information and rate–distortion acceptance (2026-10-01)

FORM-1, FORM-2 and FORM-3 close with fresh independent
semantic review,
38 passing current native consumer/source tests, a warning-free whole aggregate
build and a live-validating 1,596-declaration axiom audit. The
native controls
and aggregate/audit packet
retain commands, compiler output and unchanged source brackets. The failed
review/control attempts remain history.

Deterministic identity and complete coarsening now have actual finite KL
witnesses. Relative-support finite/native KL and mutual information identities
retain shared-zero atoms, sparse observation garbling and the singular native
infinity boundary. The asymmetric informative joint has positive information;
the independent asymmetric product has exactly zero information.
Finite fixed-source rate–distortion feasibility supplies a compact attained
minimum and an infimum-derived dual lower bound, with zero multiplier, negative
budget, informative quarter-budget, nonunique half-budget and unique zero-budget
controls. The original primary theorem and structural-proxy disposition remain
unchanged; general strong duality, dual attainment, an interior closed form and
general optimizer uniqueness are excluded.

The historical released-120 body and maturity digests remain unchanged. The
reviewed-delta regression
reverses only the exact approved additions before checking both old digests;
11 tests pass, including seven mutation refusals. Whole Python, package/H3
synthetic, hosted and publication acceptance remain open.

Scientific improvement checkpoints (2026-09-30)

PKG-1 closes: static status/readiness now rejects unavailable or changed inputs,
captures no checkout stamp through Git, and never promotes missing runtime
evidence. The independently reviewed status module passes all 26 focused cases
in the current checkout, including reachable comparisons, input mutation and
process/write sentinels. Strict publication validators remain separate.

Fresh H2 acceptance binds 329 passing mandatory cases, unchanged native inputs
and three actual source-bound reviews at terminal SHA-256
2d72e6b7205b92a625106e2830c2a01200682ffb3af036dd48118eb88f9067be.
The prior terminal bytes remain immutable history. Continuous G0 passed and
H3.0 was frozen before implementation at protocol SHA-256
50b3575316fb272dc7bf209f50a21330fc5ea0764da9399e20aedeb0b2302c93,
with two independent AI reviews. New proof, synthetic and publication acceptance
remain open; the empirical branch is governed no-go without licensed data.

The automatic Gauss capability check now uses an isolated local version probe
with dotenv and updates disabled, preserving required/advisory semantics.
Independent process/account-home/real-dotenv controls pass, and the isolated
upstream CLI reports v0.2.2. The requested free OpenRouter model returned HTTP
404; the attempted full 168-topic path rejected all topics and produced no
successful report. No paid fallback was enabled.

FEP-LEGACY-TRANSPORT closes after 155 current transport/status controls pass,
with one explicitly opt-in live Gauss test skipped. The earlier failing run is
preserved. A stdlib-only isolated guard and bounded startup fixtures retain
call-relative product deadlines, real descendant readiness, raw timeout bytes
and cleanup boundaries; fresh independent review covers the final helper.
The local acceptance observation
retains the exact transcript and post-run source map. This establishes local
transport behavior; provider, native and publication evidence remain separate.

Main status and open-backlog reconciliation (2026-09-30)

The dated review checks main/remote parity at
cd4a84c, reconciles the 168-topic/22-family source with the tagged 1.3.0
cut and latest published GitHub release 1.2.0, and scopes minor, medium,
and major package/formalization improvements. Current overview documents
now distinguish the recorded H3 continuous selection and feasibility spike
from independent acceptance and H3.0 protocol freezing.

Removed FEP-H27-RESEAL from the open-only backlog: live
validate_terminal_acceptance(Path('.')) passes at receipt SHA-256
8e20abbd4f5d63e09398014a0c64bee46ee6b42697eb24d16a488ede74e65c87
(328 mandatory cases, 208 source hashes, three reviews). The 40 focused
H2/H3 custody/preregistration tests pass. Completed-history prose and dated
clearing logs were removed from TODO.md; four genuine residual rows remain,
with FEP-FULL-CURRENT replacing the stale 155-topic task name.

Same-commit CI native and axiom-audit artifacts were independently validated
against the live source (168/168 topics at v4.34.1, zero warnings or sorry,
1,411 audited declarations). Older ignored local receipts were preserved.
Local cache incompatibility and manuscript/render/bridge freshness remain
explicit residuals; no new full native sweep, provider run, or publication
is claimed by this review.

Wave-4 catalogue wiring: 168 topics / 22 families (2026-09-28)

The catalogue grows 165 → 168 topics (families unchanged at 22) with the
wave-4 formalizations. fep-166 (Uniqueness of the Law-Weighted
Reversible/Circulation Split, area ActiveInference) joins the
standalone-efe-formalizations family through a new canonical
FEP.LawWeightedSplit module whose 7 public theorems are aliased into the
body: competing-pair uniqueness of the Helmholtz–Ao split, the reversible
boundary in uniqueness form, and the t-0060 two-state oracle datum with
nonzero forward circulation 1/2 and strictly negative reverse entry -1/4.
fep-167 (Frobenius Least-Squares Projection of the Symmetrizer) and fep-168
(Quantitative Control of the Uncancelled Divergence Remainder) extend
geometric-mechanics-notation (area Thermodynamics); the
FEP.GeometricMechanics module grows to 37 qualified declarations with the
exact Frobenius Pythagoras/minimality/uniqueness layer and the squared
Cauchy–Schwarz remainder budget with its equality witness on aligned data.
Three new FEPComposed bridges pair the new topics with their nearest
endpoints (fep-166→fep-160, fep-167→fep-162, fep-168→fep-163; seven bridges
total in the two geometric-mechanics composition files), and both extended
numerical witnesses grow typed checks and theorem mirrors
(boltzmann-efe-affinity-gap 18 mirrors / 14 checks,
geometric-solenoidal-drop 17 mirrors / 16 checks; 17 witnesses total).
The release seal moves to 168 topics / 22 families / 5 areas / 17 witnesses /
146 relations / 50 capabilities, and the third-expansion manuscript shape is
sealed at {standalone-efe-formalizations: 6, geometric-mechanics-notation: 7}
with the per-family size pin replacing the uniform-size singleton assertion.
Native capture and render re-acceptance ride the coordinator's remaining
phases and are pending, not part of this wiring.

Wave-3 catalogue wiring: 165 topics / 22 families (2026-09-27)

The catalogue grows 159 → 165 topics and 21 → 22 families with the wave-3
formalizations. fep-160 (Helmholtz–Ao Decomposition of Nonequilibrium Steady
Currents, area ActiveInference) joins the standalone-efe-formalizations
family; a new geometric-mechanics-notation family (area Thermodynamics)
carries fep-161..165: skew trace/quadratic cancellation, the discrete-Hessian
symmetrization (finite Clairaut), the three-term solenoidal expansion with its
conditional drops, the graph-plane divergence decomposition, and the necessity
witness refuting the unconditional solenoidal drop. The FEP.HelmholtzAoNess
module contributes 25 public declarations (13 main theorems aliased into the
body); the FEP.GeometricMechanics module contributes 22 declarations of
which 19 public theorems are aliased (including the symmetrize/Clairaut
family — the pushed module commit message under-counted and is not the
inventory). Two new FEPComposed composition files carry 6 novelty bridges
(fep-160→fep-025, fep-161→fep-160, fep-162→fep-161, fep-163/164/165→fep-025),
a cap-geometric-mechanics-solenoidal capability node lands with 8 evidence
entries, and the geometric-solenoidal-drop numerical witness gives the new
family its typed non-proof evidence (17 witnesses total). The release seal
moves to 165 topics / 22 families / 5 areas / 17 witnesses / 143 relations /
50 capabilities. The 143/50 relation and capability planes correct a
pre-existing seal drift: a0f4ed3 added cap-standalone-efe-theorems and four
formal_pairing edges without bumping RELEASE_SEAL (live counts were 49/137
against a 48/133 seal at a64ca6c). Manuscript 04d's NESS disclaimers are
retired against the now-compiled statements (ansatz-universality, trace
quadratic and Hessian cancellations, conditional solenoidal drop, and the
2-node failing witness), keeping the state-dependent-diffusion and PDE/SDE
boundaries. The roster-v23 source-owner coordinated refresh, native capture,
render re-acceptance, and pin-cycle #32 re-seal ride the coordinator's
remaining phases and are pending, not part of this wiring.

Coordinated evidence refresh: SOURCE_OWNER_ROSTER v24 (2026-09-28)

OWNER_MANIFEST_VERSION 23 → 24 covering the wave-4 fold chore's custody
cascade: the new law_weighted_split canonical formal module
(FEP.LawWeightedSplit, 7 public theorems) joins the native roster through
the formal-modules plane with its byte-identical lean/FepSketches/ mirror,
the wave-4 wiring delta touches the reviewed owner files it reuses (the two
family body modules, the numerical-witness extensions, the release-seal and
third-expansion surfaces), and the H2.7 evidence-custody chain
(acceptance → matrix → R0 prior/successor custody receipts → 05d/05b/06a →
terminal packet → H3 spike lockstep) is re-bound over that delta with the
strip model in tests/_support/h2_r0_custody.py gaining this wave's
added-module group so the reconstruction still hashes to the sealed R0
manifest digest. Prerequisite for fep-lean verify claim-readiness at any
tip containing the wave-4 catalogue (the native capture fails the roster
pre-gate otherwise); the native capture, render re-acceptance, and pin-cycle #33
re-seal follow.

GEO-INFER notation slice growth to 17 rows (2026-09-28)

• Grew specs/geo-infer-notation-bridge/data/notation-map.yaml from 13 to 17
  notation-level rows: fep-004 (Fisher metric quadratic form,
  information_metric), fep-018 (Fisher-Rao distance between parameter
  points, geodesic_distance), fep-038 (Fisher information as log-likelihood
  curvature, fisher_information_matrix), and fep-158 (Bayes-odds evidence
  weight between models, ModelSelection), each anchored on a lane-verified
  GEO-INFER-MATH anchor — the first cross-module anchors, ratified this
  session — and validated by check_geo_notation_bridge.py --check; no proof
  or verification claim is added.

GEO-INFER notation slice growth (2026-09-28)

• Grew specs/geo-infer-notation-bridge/data/notation-map.yaml from 10 to 13
  notation-level rows: fep-025 (row-normalized transition kernel,
  MarkovDecisionProcess), fep-156 (Boltzmann control posterior,
  PolicySelector), and fep-157 (perception update as forward filter,
  MarkovDecisionProcess), each anchored on an already-verified GEO-INFER-ACT
  anchor and validated by check_geo_notation_bridge.py --check; no proof or
  verification claim is added.

Coordinated evidence refresh: SOURCE_OWNER_ROSTER v23 (2026-09-27)

OWNER_MANIFEST_VERSION 22 → 23 covering the wave-3 fold chore's custody
cascade: the new geometric_mechanics catalogue body joins the roster via
body_source_relative_paths() and the H2.7 evidence-custody chain
(acceptance → matrix → R0 prior/successor custody receipts → 05d/05b/06a →
terminal packet → H3 spike lockstep) was re-bound over the wave-3 wiring
delta (two foundation modules + two bridge compositions in
FORMAL_MODULES, their byte-identical lean/FepSketches/ mirrors, and the
expanded release seal). The strip model in tests/_support/h2_r0_custody.py
gains the wave-4 group constants so the reconstruction still hashes to the
sealed R0 manifest digest. Prerequisite for fep-lean verify
claim-readiness at any tip containing the wave-3 catalogue (the native
capture fails the roster pre-gate otherwise); the native capture and the
render re-acceptance follow before the pin-cycle #32 re-seal.