NOISY ADMISSIBILITY: A CONSOLIDATED THREE-ITERATION RESEARCH REPORT

Preserved consolidated research report from 4 November 2026. This is broader supporting material for the selected public addendum, not another final manuscript or new external review. Original internal-review and result claims remain dated research records. Local path references are omitted in this public derivative; no original is edited. Original SHA-256: 4bed3e0f81a46372cbcd9bd1fbcf942d5bb13a9ebe9f2cc5670c838c8ac7ee0a.
NOISY ADMISSIBILITY: A CONSOLIDATED THREE-ITERATION RESEARCH REPORT
Iteration 03 of 03 — completed 4 November 2026, Asia/Bangkok

Exact finite-horizon results, contract distinctions, proofs and reproducible
certificates. This is internal mathematical research with independent AI
checking, not external peer review or a claim of publication priority.

RESULT IN BRIEF

At horizon five, in the four-state RF example defined below, the exact
minimax loss under the ACTUAL remaining first-failure HC-commit constraint is

  267963/640625 = 0.41828370731707317.

The specified original-prior PPR confidence-sequence envelope has optimum
291433/640625 = 0.4549198048780488. Its exact loss is therefore
4694/128125 = 0.03663609756097561, about 8.759% of the HC optimum. A tailored
time-uniform gate attains the HC optimum, so this is a penalty of the
specified PPR construction, not a universal penalty of confidence sequences.

All 19 RF-legal test/carrier actions are covered. A complete realization
formulation keeps full descendant first-failure constraints. Its 599-variable
numerical solution agrees with a separately certified rational primal/dual
solution. The rational certificate, rather than floating-point agreement,
establishes optimality. A fresh independent implementation checks all five
new certificates without importing the optimizer.

Two distinct exclusions first become possible at horizon five. Nevertheless,
an exhaustive exact recurrence proves that, in this specific model at
alpha=1/50, immediate conditional caps imply the full remaining-risk cap
through horizons five AND six. Thus this bounded search finds no immediate
versus remaining-risk separation. It does find a sharp carrier-expansion
counterexample at horizon five: an admissible continuation becomes
inadmissible when an earlier carrier is enlarged. Maximal-carrier policies
happen to achieve the same optimum at the point studied; no performance
gap from that restriction is established.

An inherited prose error is corrected explicitly in Section 3. Earlier
files remain unchanged. This report completes the authorized three-iteration
sequence; no fourth iteration has been started.


1. MODEL, SCOPE AND STATUS OF CLAIMS

H=(a,x,y,z), with diamond order a<x,y<z. The target labels are
ell=(0,1,2,0). The prior is pi=(1/4,1/4,1/4,1/4). Nominal queries are
q=0110 and r=1100. Every execution returns its nominal bit through a fresh
independent known binary symmetric channel with error epsilon=1/5. Each
execution costs c=1/50. The hidden state is fixed, and an execution has no
physical, resource-consuming or cumulative structural effect. Policies may
stop early and report any of the three labels, and may randomize.

A nonempty declared carrier C is RF-legal when the pre-answer Boolean
order-thinning inclusion Delta((P|C)_e) -> Delta(P|C) is a homotopy
equivalence. This predicate is inherited from the B-Theory acquisition
manuscript [B1, Section 2, Lemma 1]. It is checked on the declared carrier,
not on a posterior distribution and not after observing the answer.

The contract permits a declared carrier to replace exact posterior support.
That is a modeling change. With positive BSC likelihoods, exact posterior
support never loses a prior-positive state. Raw full-support legality would
therefore forbid r throughout this example. None of the numerical gains
asserts that noisy Bayesian updating literally removes states from support.

Failure F means that at least one executed carrier excludes the true h.
It is a static certificate failure, not physical injury or a fresh hazard
on every repetition. Define r_h=P_h(F), e_h=P_h(wrong target report), and
n_h=E_h[number of observations]. The composite minimax objective is

  V = max_h (e_h + c n_h).

This differs from max_h e_h + c max_h n_h, and from prior-average loss.
Certificate error and report error remain distinct; a bound on one is not
a bound on their union unless stated explicitly.

Claim status:

* Inherited baseline: RF legality and the noiseless acquisition setting;
  the earlier oracle/coupling approach to noisy feasibility.
* Established antecedents: confidence sequences and prior/posterior ratios;
  first-event accounting; survival-conditioned chance constraints;
  sequence-form realization weights; LP duality, scalar dynamic programming
  and repeated-majority amplification. Sources appear in Section 12.
* Established within this research: the stated finite-model proofs,
  carrier-contract feasibility characterization, exact frontiers from
  iterations 1–2, the horizon-five certificates, the finite-horizon
  immediate-cap equivalence, and the explicit carrier-expansion witness.
* New in iteration 3: the complete implemented HC realization LP; exact
  horizon-five comparisons; the all-reachable-state horizon-five/six
  check; a tailored HC-compatible time-uniform gate; the inherited erratum.
* Open: general-horizon HC optimization, a quantitative maximal-carrier
  restriction gap, when immediate caps first fail in this fixed example,
  unknown or misspecified noise, changing states, physical hazard models,
  scalable homotopy algorithms, and independent publication-level novelty.

These statuses distinguish new work in this sequence from a claim that
the underlying mathematical technique is new to the literature.


2. THE RISK LEDGER AND THE ACTUAL CONDITIONAL CONTRACT

Before an action, let b be the FULL posterior and let D be the intersection
of all previously executed carriers, starting with D=H. At selected carrier
C, the new first-exclusion mass is b(D\C). The events of first exclusion
are disjoint on the joint state/history probability space. Conditional
expectation and summation therefore give the exact identity

  P_pi(F) = E_pi sum_t b_t(D_t\C_t).                    (1)

For an unbounded sequence use monotone convergence; after stopping use H
as the harmless formal carrier. Reusing an earlier carrier contributes zero
new static certificate risk because D is already contained in it. Readding
an excluded state later never repairs an earlier failure. If D is empty,
future innovation is zero but failure has already occurred on that history.

Survival G={h in D} is usually unobserved. For b(D)>0 define

  rho(eta,a) = P(first exclusion now or later | public history eta,
                 selected action a, h in D).

HC-commit requires rho<=alpha at every positive-probability selected test
action, BEFORE executing it. It includes that action and all its future
continuation. HC-pre instead averages the fresh current action coin before
selection; these are different contracts. An episode bound constrains only
the probability from the root. A statewise HC condition, imposed separately
under each surviving h, would forbid any exclusion for alpha<1; that is
also a different, much stronger condition.

For a committed carrier C let beta=b(D\C)/b(D). Conditional on surviving
that carrier, let p_surv(y) be the observation law. Then

  rho = beta + (1-beta) sum_y p_surv(y) rho_child(y).     (2)

The child term averages future behavioral action randomization, while each
future selected action must also satisfy its own committed bound. Setting
beta<=alpha at each gate does not generally imply rho<=alpha. Replanning
with a fresh alpha is not a risk-budget proof. Past failure and future risk
are related by

  P(eventual F | eta) = b(D^c) + b(D) rho_pre(eta).       (3)

No calculation below renormalizes away excluded states in sampling cost or
terminal error. Survival conditioning is used only for the stated risk
contract. This distinction parallels safe-belief and execution-risk issues
in RAO* [P1, Sections 2–3]; it is not a new general chance-constraint idea.

For a general counterexample to reusing a per-gate allowance, iteration 1
used the three-state chain a<x<z with prior (.04,.91,.05), query 110 on ax,
then query 100 on xz, both with BSC error .2. Both carriers are RF-legal
because their query is constant on the carrier. The first exclusion is .05;
the second full-posterior tail is either 4/115 or 16/385, both below .05.
Its survival-conditioned immediate hazard is 4/95, also below .05. Yet
episode first-failure risk is .09, and the first committed risk-to-go is
.09. This is a general counterexample, not a reachable uniform-diamond
example at alpha=.02.


3. RF CARRIERS, AN ERRATUM, AND RELIABLE ACQUISITION

Use masks a=1,x=2,y=4,z=8. The complete legal sets are:

  q: 1,2,3,4,5,6,7,8,9,11,13,15;
  r: 1,2,3,4,6,8,12.

Thus there are 19 actions. The maximal q carrier is H; the maximal r
carriers are ax,xy,yz. Legality follows from these endpoint cases: with a,z
present require bit(a)<=bit(z); with a alone require bit(a)=0 or all bits
one; with z alone require bit(z)=1 or all bits zero; with neither endpoint,
every carrier is legal.

RF is not hereditary: q/H is legal but q/xz is illegal because an edge
becomes disconnected. However, iteration 1's prose additionally claimed
q/axz and q/xyz were both legal and had illegal intersection xz. That is
incorrect: q/xyz is itself illegal. The 19-action code and all numerical
certificates excluded xyz correctly. In this diamond, the nonempty
intersections of legal carriers for each fixed query are in fact legal.
This report withdraws that intersection example; no generic intersection
failure theorem is being inferred from it. The non-hereditary result and
the need to check the actual carrier remain valid. Earlier files are
preserved for traceability rather than silently corrected.

For a statistical core A define Env_e(A)={legal C: A subset C subset H}.
Opening e requires a legal completion, not necessarily legality of A itself.
Adding states as structural scaffolding does not claim they regained exact
statistical plausibility. If legality must hold on EVERY nonempty subset
of C, then the nominal query must be order-preserving on P|C: any descending
pair produces a two-vertex edge-to-disconnected obstruction. Conversely,
an order-preserving query deletes no comparability on any subset, so all
these inclusions are identities. This proves the hereditary robustness
characterization for every finite poset, without a height restriction.

For finite static nominal tests, define A(e,S) to mean Env_e(S) is nonempty,
and T(S) to mean the target is constant on S. Put S_i={h in S:q_e(h)=i}.
The recursion

  W(S) = T(S) OR exists e nonconstant on S with
                  A(e,S), W(S_0), W(S_1)                 (4)

characterizes arbitrarily reliable JOINT carrier coverage and target
diagnosis under the declared-carrier contract. It is a finite mathematical
characterization, not an efficient general homotopy algorithm.

Proof of sufficiency: take a W tree of depth d. At each node choose a legal
carrier containing its nominal support; repeat its query an odd k times
and use majority. Before the first majority error, the true state stays
inside all supports and carriers. With independent BSC error epsilon<1/2,

  max_h P_h(F OR wrong report)
       <= d exp[-2k(1/2-epsilon)^2], samples<=dk.         (5)

Proof of necessity: if sum_h P_h(F OR wrong report)<1 and all runs terminate
almost surely, use an ideal nominal oracle to simulate the noisy policy
with one common random tape for all finitely many counterfactual states.
The union bound gives a tape on which every state-run is correct and never
excluded, with all runs terminating. Fix it. States at the same new oracle
query share the same synthetic past and chosen carrier, so the carrier
contains the exact support S and A(e,S) holds. Cache repeated oracle bits;
simulate constant queries, synthetic noisy answers, clock increments and
internal memory without a new oracle call. The model's absence of structural
or resource effects makes this pruning valid. The finite terminating runs
give a finite tree of nonconstant splits and homogeneous leaves, hence W(H).

Therefore if W(H) fails, every almost-surely terminating policy has sum_h
joint failure>=1, worst-state joint failure>=1/|H|, and prior-average joint
failure>=pi_min. Nontermination cannot simply be ignored. This is an
adaptation of the inherited coupling argument and standard amplification
[P6], not a new general oracle-simulation principle.


4. COMPLETE REALIZATION FORMULATION FOR HC-COMMIT

Fix any finite horizon N. A history eta records all previous chosen tests,
carriers and observed bits. Let l_eta,h be its observation likelihood under
h, excluding behavioral action probabilities, and let D_eta be its carrier
intersection. Both are fixed coefficients of the history tree. Include all
19 legal actions and three stopping labels at each applicable node.

Let x_eta be the product of past behavioral action probabilities and
x_eta,a the corresponding weight after choosing a. Impose nonnegativity and

  x_root=1; sum_a x_eta,a=x_eta;
  x_(eta,a,y)=x_eta,a for EACH observation y.            (6)

Observation likelihoods are already in l. The child weights must not be
summed to x_eta,a and must not receive a second factor K(y|h).

For a test edge v=(nu,a) with carrier C_v define

  f_v,h = l_nu,h 1{h in D_nu\C_v},
  f_v   = sum_h pi_h f_v,h.

The four episode risks are R_h(x)=sum_v f_v,h x_v. Terminal errors are
E_h(x)=sum_(stop edges eta,j) l_eta,h x_eta,j 1{ell(h)!=j}.
Expected samples are T_h(x)=sum_(test edges eta,a) l_eta,h x_eta,a.
All four states occur in E_h and T_h, including states previously excluded.

For committed edge v=(eta,a), let Sub(v) contain v itself and every test
edge in its continuation. Its exact HC row is

  sum_(u in Sub(v)) f_u x_u
       <= alpha x_v sum_(h in D_eta) pi_h l_eta,h.       (7)

This is linear, rational and includes all remaining first failures. Why?
The numerator sums disjoint first-exclusion events in the selected subtree.
Every later D is contained in D_eta, so only states alive at eta contribute.
The denominator is the joint probability of reaching eta, selecting a, and
having survived. Dividing gives the requested conditional probability.
If x_v=0, nonnegative flows force its descendants to zero and the row is
vacuous. If the survival mass is zero, the conditional contract is undefined,
not reset; failure remains in the episode ledger. Such cases do not arise
in the reachable alpha=.02 HC tree here.

Minimize t subject to E_h+cT_h<=t, (6), (7), and, when requested, R_h<=delta.
These rows are necessary and sufficient for finite public-history behavioral
policies: any policy supplies x; conversely ratios x_eta,a/x_eta define
behavioral probabilities at positive nodes and arbitrary behavior at zero
nodes. This is a specialization of established realization/sequence-form
ideas [P5], with the first-failure rows derived explicitly here.

The conditioning filtration is public action/observation history and fresh
behavioral coins. A root mixture seed known to the decision-maker and
included in the conditioning information is generally a stronger contract.
Here every displayed HC mixture component individually satisfies the full
bound, so the reported achievable solution also meets that stronger
seed-aware interpretation. The lower bound applies to the public-history
class, proving the same optimum for both interpretations at this point.

For computation, beta>alpha is a sound necessary exclusion of an action,
since all remaining risk is nonnegative. Applying ONLY that pruning yields
history counts [1,2,4,8,28,128] by depth at N=5. The resulting full LP has
171 flow rows, 598 action variables plus t, and 85 descendant HC rows.
It considers all 19 legal actions before pruning at every node. It imposes
all four loss rows and four optional episode rows. This is complete for
the stated HC class, not merely a search over selected deterministic trees.
The numerical optimum is .4182837073170732; largest residuals are below
5e-16. These residuals are corroboration, not the exact proof.


5. MULTIPLE EXCLUSIONS AND WHAT THE BOUNDED SEARCH DOES NOT FIND

No proper carrier passes beta<=.02 before the fourth observation: the
smallest posterior state mass through two q answers is 1/34>.02, and only
q/H is available initially. At N<=4 there is thus at most one exclusion.

A genuine two-exclusion policy at N=5 is: execute q/H three times; only
after 111 execute q/axy, excluding z; only if its answer is 1 then execute
r/xy, now excluding a. Stop on all other branches. The fourth committed
remaining risk is 1/130+(1/5)(1/130)=3/325, and the fifth risk is 1/513.
The earlier committed risks along the relevant path are 3/1250,3/625,3/425.
All are below .02. The root prior risk is 3/1250; endpoint state risks
are 1/625 for a and 1/125 for z. Two distinct exclusions really occur on
one path, with both retained in the first-failure accounting.

To test immediate caps without assuming their sufficiency, the independent
reviewer exhaustively solves this exact worst-risk recurrence. Use integer
likelihood numerators L and mass m(L,D)=sum_(h in D) L_h. Let B_n(L,D) be
the maximum unnormalized future first-exclusion mass among every policy
obeying the immediate cap at every selected action. Then

  B_0=0;
  B_n(L,D)=max(0, max_(legal e,C with 50 m(L,D\C)<=m(L,D))
       [m(L,D\C) + (1/5) sum_y B_(n-1)(L*Knum(e,y),D∩C)]).  (8)

An induction over the remaining horizon proves (8); randomization cannot
exceed the maximum of deterministic choices. Exhaustively visit every
reachable (L,D) state and check B_remaining/m(L,D)<=1/50. The reachable
aggregated state counts are [1,2,3,4,17] at decision depths 0–4 for N=5;
N=6 adds 27 at depth 5. There are ZERO violations. The maximum root
Bayesian risk is 1/125 at N=5 and 37/3125 at N=6. Checking every reachable
node, not merely these root numbers, proves the conditional result.

Thus immediate-cap policies and full HC policies coincide at these two
horizons for this model and allowance. N=6 was used only for this small
exact feasibility check after N=5 failed to separate the notions. No N=6
objective optimization or N>=7 exploration was performed. Multiple possible
exclusions alone do not establish that a local-cap policy violates HC.
Equation (8) is model-specific verification, not a universal equivalence.

Carrier enlargement DOES fail as a feasibility-preserving transformation.
Execute q/H three times; only after 111 execute q/xy; only if the fourth
answer is 0 execute r/xy. Otherwise stop. The fourth commit risk is 1/65;
the fifth is zero because D=xy already. Earlier committed risks along 111
are 1/250,1/125,1/85. All satisfy .02. Enlarge only the fourth carrier to H,
preserving the observations and subsequent actions. At history 1110 the
fifth r/xy now excludes endpoint mass 1/17>.02. The enlarged policy fails
HC. Observations, terminal decisions and sampling cost have not changed.

This invalidates the usual nodewise carrier-dominance justification for HC.
It does not prove a loss from restricting the optimized class to maximal
carriers: at the numerical point below their optimum is exactly the same.


6. EXACT HORIZON-FIVE OPTIMA AND MATCHING CERTIFICATES

All rows use max_h(e_h+.02 n_h), horizon 5, and the same fixed model.

  Class                           Exact value          Decimal
  HC-commit, all carriers          267963/640625         .418283707317
  HC-commit, maximal carriers      267963/640625         .418283707317
  Original-prior PPR envelope      291433/640625         .454919804878
  PPR envelope plus HC-commit      291433/640625         .454919804878
  Per-state episode risk only      3766683/10250000      .367481268293

The HC/PPR solves also included r_h<=.02. Their exact dual risk multipliers
are all zero. Their lower bounds therefore apply with those episode rows
removed, and the displayed feasible mixtures prove the same optima for
pure HC or pure PPR, respectively. There is no hidden episode restriction
behind the reported HC optimum.

For HC, the exact lower-bound weights are

  w=(0,193/410,193/410,12/205), lambda=(0,0,0,0).

The scalar DP over all immediate-feasible legal actions returns
min_P sum_h w_h(e_h(P)+c n_h(P))=267963/640625. Every HC policy is included
in that immediate-feasible class. Thus this is a valid lower bound even
without using the stronger equivalence of Section 5. The optimal all-carrier
mixture has weights 115/246,46363/92496,2893/92496 and exact metrics:

  state a,z: r=1/125; e=130423/384375; n=303548/76875;
  state x,y: r=1/125; e=134209/384375; n=265688/76875.

Each state's composite loss equals the bound. The maximum committed
remaining risk in every component is 1/65. The maximal-carrier mixture has
the same errors and sample counts, endpoint risk 1/125, and zero middle-state
risk. Its weights are the same with the last two policies reordered.
Full labeled trees are in certificates.json and policies.txt.

For PPR use w=(12/205,193/410,193/410,0), lambda=0. The matching scalar
minimum is 291433/640625. Its three weights are
281/984,225055/450672,96919/450672. Its maximum component HC risk is 1/257.
It has endpoint risk 1/625 and middle-state risk zero. The episode-only
certificate has six components and four positive risk multipliers; its
full rational data are saved. Some components have HC risk as high as 1/2,
so its lower loss must not be advertised as HC-feasible.

Certificate logic: for nonnegative w summing to one and lambda>=0,

  max_h(e_h+c n_h)
    >= min_P sum_h[w_h(e_h(P)+c n_h(P))+lambda_h r_h(P)]
       -delta sum_h lambda_h.                          (9)

The DP at likelihood l and intersection D minimizes among stop labels and
legal test choices. A test adds sum_h c w_h l_h plus
sum_(h in D\C) lambda_h l_h, then both children with l_h K(y|h) and D∩C.
Stop cost is sum_h w_h l_h 1{ell(h)!=label}. This accounts for all states
in loss and only first exclusions in risk. Induction proves the exact
minimum. A rational feasible mixture equaling (9) proves global optimality.
The computational LP proposes columns; exact Fraction arithmetic and rational
linear algebra certify the result. The independent check reconstructs the
scalar DP and leaf metrics separately, with an explicit 19-action list.

The lower-bound formulation and optimization approach have established
sequential-testing antecedents [P2]. The particular RF constraints,
conditional-risk audit, numerical points and separating examples are the
research results, not a claim to have invented LP-based sequential testing.


7. PPR LOSS AND A TAILORED TIME-UNIFORM GATE

The specified PPR construction uses the ORIGINAL prior at every time:

  A_t={h:b_t(h)>alpha pi_h}, alpha=1/50.

For fixed h, pi_h/b_t(h)=sum_g pi_g L_t(g)/L_t(h) is a nonnegative mean-one
martingale under h for predictable tests. Ville's inequality gives
P_h(exists t:h not in A_t)<=alpha. Choosing any legal carrier containing
A_t preserves that simultaneous statewise coverage. This is the established
prior/posterior-ratio method [P4, Proposition 2.1], not a new confidence
sequence. Here we optimize over carriers containing the RAW current core;
we do not change it to a running-intersection core or a restarted prior.

After q answers 111 the endpoint masses are 1/130>.005=alpha pi_h; hence
PPR still retains H and cannot open r. After 1111 each endpoint has mass
1/514<.005, so r/xy can first open as the fifth observation. HC can open
r/xy one observation earlier because the remaining exclusion risk after
111 is 1/65<.02. This accounts for a concrete screening delay.

The exact optimized composite-loss penalty is

  V_PPR - V_HC = 4694/128125 = .03663609756097561,
  (V_PPR-V_HC)/V_HC = 23470/267963 ≈ .0875867.

It is not an extra-samples estimate at a common error target. The previous
iteration separately measured that type of cost at N=4 (Section 8).

A tailored time-uniform gate removes this penalty. Initially allow only
q/H for up to three observations, with early stopping allowed. Immediately
before a fourth executed query, contract the actual carrier to EXACTLY xy
if and only if the initial three answers were 111. This executes q/xy or
r/xy and makes the actual intersection D equal xy. On all other branches
keep carrier and core H forever. After an executed contraction, use any
RF-legal carrier containing xy; never shrink D again. Merely shrinking an
abstract confidence core while executing carrier H is insufficient: then
actual D would remain H and a later r/xy could violate HC. Stop by N=5 for
the optimization result.

For a or z, ever losing coverage requires q^3=111, probability (1/5)^3=1/125;
for x or y it never occurs. Therefore the core is a statewise time-uniform
confidence set at level at least 1-.008, including extension after stopping
by holding it fixed. Before the possible transition, committed remaining
risks are bounded by 1/250,1/125,1/85 along the relevant prefix. At transition
they are at most 1/65; afterwards no first exclusions remain. This directly
proves HC-compatibility for this gate family, even if legal queries continue
afterwards without further core changes.

Every component of the optimal maximal-carrier HC mixture fits this gate.
Consequently a genuinely time-uniform, HC-compatible tailored gate achieves
the unrestricted HC optimum at N=5. This does not say that arbitrary
confidence-sequence completion automatically satisfies HC: an episode
coverage theorem alone does not control every posterior conditional tail.
The tailored proof uses the precise trigger and persistence rule.


8. CONSOLIDATED RESULTS FROM ITERATIONS 1 AND 2

Iteration 1's Bayes-objective/Bayes-risk N=3 optimum was 713/2500=.2852,
with error 29/125=.232, expected samples 133/50=2.66, and risk .02. Its
endpoint state risks are .04. It must not be compared as though it solved
the minimax or per-state risk problem. Exhaustive enumeration of 40,805
trees agreed with its certificate; 240 finite topology checks were recorded.

For minimax loss at price .02 and allowance .02, iteration 2 established:

  Risk/gating class                 N=3                 N=4
  Bayesian episode risk             719/1875            37903/101250
  Per-state episode risk            352229/780000       2587807/6480000
  HC-commit                         25817/48750         84989/202500
  Specified PPR completion           25817/48750         77/146

At error_h<=.4, N=4, the minimum worst-state expected sample count is
295163/139600 under per-state episode risk, versus 1459/560 under HC.
The exact HC extra cost is 239907/488600≈.491009 samples. PPR cannot reach
that error target at N=4. This comparison holds the error target fixed.

Iteration 2 also certified the entire N=3 per-state risk/loss frontier
V(delta), not merely a parameter grid. The six affine pieces are:

  delta interval                   V(delta)
  [0,296/8925]                     25817/48750 -(20281/5200)delta
  [296/8925,16/425]                983/1875 -(187/50)delta
  [16/425,4/85]                    2603/6375 -(33/50)delta
  [4/85,1517/10200]                2549/6375 -(12/25)delta
  [1517/10200,14/75]               29828/80625 -(12/43)delta
  [14/75,1]                       596/1875.

At N=3, per-state risk .02, and error_h<=eta, the minimum max_h n_h is
infeasible for eta<463/1125; zero for eta>=2/3; and otherwise:

  eta interval                     minimum worst-state expected samples
  [463/1125,3474/8375]             46453/1200 -(355/4)eta
  [3474/8375,1717/4125]            15187/1200 -(415/16)eta
  [1717/4125,11059/25875]          20309/2880 -(2395/192)eta
  [11059/25875,493/1125]           86213/13500 -(131/12)eta
  [493/1125,47/90]                7639/1260 -(285/28)eta
  [47/90,119/225]                 319/60 -(35/4)eta
  [119/225,2/3]                   10/3 -5eta.

Endpoint mixtures attain every intervening affine segment, while a fixed
dual supports it throughout; that is the proof of sharpness. The earlier
standard-library checker and independent AI review verified 42 exact
certificates covering these segments and selected contract comparisons.
Their saved evidence is retained unchanged in outputs/iteration_02.

Randomization matters. At N=3 any deterministic first r before the horizon
has exclusion probability at least .04 under an excluded state, violating
the .02 statewise budget. A q-only policy cannot distinguish x from y, so
its maximum error is at least .5. The randomized feasible minimax loss
352229/780000<.5 proves a strict gain. Randomization does not improve a
fixed scalar DP; it is needed to balance the constrained vector objective.


9. VERIFICATION AND REPRODUCTION

New evidence in this directory:

* certificates.json: five complete rational policies/mixtures, state metrics,
  objective values and matching scalar-DP duals.
* optimizer.py and run_research.py: adapted exact column-generation solver;
  all 19 carriers used in the iteration-three pricing oracle.
* verify_certificates.py and verification.json: adapted standalone verifier;
  five exact certificates passed, 162 conditional nodes checked.
* independent_verify.py and independent_verification_results.json: fresh
  reviewer implementation, no optimizer imports; five certificates passed.
* independent_search.py and independent_search_results.json: exact maximum
  remaining-risk recurrence, all reachable N=5 and N=6 states, carrier
  expansion witnesses.
* realization_lp.py and realization_results.json: independently formulated
  public-history LP, with actual future-risk rows; numerical cross-check.
* independent_ai_review.txt: fresh proof and code audit, including limits
  and the inherited carrier-intersection erratum.
* policies.txt: human-readable complete trees for the optimal maximal HC
  mixture and PPR mixture.
* comparison.png and video_storyboard.txt: explanatory figure and written
  video explanation, with complete narration and visual directions.
* sources.json, preservation_check.json, manifest.json: references,
  preservation evidence and hashes.

From the workspace root, reproduce the new results with Python 3.10+:

  python outputs/iteration_03/run_research.py
  python outputs/iteration_03/verify_certificates.py
  python outputs/iteration_03/independent_verify.py
  python outputs/iteration_03/independent_search.py
  python outputs/iteration_03/realization_lp.py

Optimization needs scipy and sympy; the certificate and search checkers
use only the standard library. Python bytecode output is disabled in the
optimizer runner and realization implementation. No inherited module is
imported or executed. The entire new optimization batch took roughly
26 seconds of elapsed local execution; exact certificate checking was
subsecond. Numerical LP agreement is explicitly secondary to rational
primal/dual equality and exact conditional audits.

Bounds: N=5 optimization only; N=6 limited to a small exact feasibility
recurrence; at most 120 column-generation iterations per solve; five
objective solves; at most 300 seconds permitted for their batch and
60 seconds for the realization cross-check batch. No randomized search,
GPU, paid API, external write, publication, credentials, Git operation,
or security-setting change was used. All new files are under iteration_03.
Hashes of previous iteration files were recorded before the new work and
checked again at delivery. The final preservation_check.json gives counts
and any differences. A negative result from this search is bounded by
these stated horizons and fixed parameters.


10. INTERPRETATION AND REMAINING QUESTIONS

The mathematical issue is access under uncertainty: evidence can justify
opening a structurally legal carrier while admitting a controlled chance
that the true state was left outside it. Structural validity, confidence
coverage, conditional risk and diagnostic performance are separate objects.
The model demonstrates their interaction and measures some finite costs.
It does not establish self-awareness, consciousness, physical safety,
real-world calibration, or reliability under model misspecification.

The strongest new conclusion is a completely certified N=5 optimization
with an explicit, reviewable risk contract. The failure of naive carrier
expansion is also established. The absence of an immediate-cap separation
through N=6 and the absence of an optimum gap from maximal carriers at N=5
are equally part of the result; they should not be replaced by a stronger
claim merely because a separation was the initial search objective.

Further mathematical work would need a new scope: an N>=7 or parameter
search for immediate-cap failure; a genuine maximal-carrier value gap;
and general theory for HC policy restrictions. External review would be
valuable before incorporating the results into a manuscript. None is
required to finish the present three-iteration assignment, and none has
been started as an unrequested fourth turn.


11. VIDEO EXPLANATION

video_storyboard.txt is a written video explanation with a complete spoken
script, approximate timing and visual directions. It explains the fixed
model, first-failure accounting, HC-commit, realization weights, the exact
N=5 comparisons, the two-exclusion and carrier-expansion examples, the
PPR/tailored-gate distinction, verification and limits. It is an explanation
of these results, not a substitute for the proofs or JSON certificates.
The final scope steering explicitly accepted this written format and asked
that no additional time be spent rendering or narrating a video. No video
or synthetic audio has therefore been rendered.


12. PRIMARY SOURCES AND ATTRIBUTION

[B1] Brian Theory, From Observable States to Acquirable Predictive Models,
revised 7 October 2026. Section 2, Lemma 1: inherited local RF criterion.
https://www.b-theory.com/manuscripts-html/acquirable-predictive-models/

[B2] Brian Theory, Binary Order-Thinning. Inherited structural background;
the present non-heredity argument is a direct two-vertex calculation.
https://www.b-theory.com/manuscripts-html/binary-order-thinning/

[P1] Pedro Santana, Sylvie Thiebaux, Brian Williams (2016), RAO*: An Algorithm
for Chance-Constrained POMDP's. Sections 2.1–2.3, equations 10–17 and
Section 3: execution risk, survival conditioning and accounting cautions.
https://people.csail.mit.edu/psantana/data/files/articles/2016_AAAI_rao.pdf
https://doi.org/10.1609/aaai.v30i1.10423

[P2] Michael Fauss, Abdelhak M. Zoubir (2015), A Linear Programming Approach
to Sequential Hypothesis Testing. Section 4, Theorems 4.1–4.2: antecedent
Lagrangian and LP treatment of sequential tests.
https://arxiv.org/pdf/1501.05870

[P3] Yin-Lam Chow, Marco Pavone (2015), A Time Consistent Formulation of Risk
Constrained Stochastic Optimal Control. Theorems II.4–II.5 and III.3:
background on risk-to-go augmentation; not a theorem about our exact HC rows.
https://arxiv.org/pdf/1503.07461

[P4] Ian Waudby-Smith, Aaditya Ramdas (2020), Confidence sequences for sampling
without replacement. Section 2.1, Proposition 2.1: prior/posterior-ratio
construction. Its likelihood-ratio argument is applied to the fixed model.
https://papers.nips.cc/paper_files/paper/2020/file/e96c7de8f6390b1e6c71556e4e0a4959-Paper.pdf

[P5] Daphne Koller, Nimrod Megiddo, Bernhard von Stengel (1996), Efficient
Computation of Equilibria for Extensive Two-Person Games. Sequence form and
perfect recall are established methods. Author/institution record and
von Stengel's explanatory slides were checked in iteration 3:
https://research.ibm.com/publications/efficient-computation-of-equilibria-for-extensive-two-person-games
https://conferences.mpi-inf.mpg.de/adfocs-24/material/Bernhard/2ext-adfocs.pdf
The author-hosted journal PDF timed out; this report does not claim a fresh
full reading of that PDF or import a game-equilibrium theorem as our HC proof.

[P6] U. Feige, P. Raghavan, D. Peleg, E. Upfal (1994), Computing with Noisy
Information. Section 1.1, page 1002: classical repeated-majority simulation.
https://cadmo.ethz.ch/education/lectures/HS18/SAADS/papers/computing_noisy_information.pdf

Sources B1–B2 and P1–P4/P6 were consulted during the earlier iterations;
their relevant contents and reading locators are inherited here. P5's
institutional record and primary lecture slides were opened in iteration 3.
The targeted source checks are not a systematic novelty review. The
separate novelty investigation was not contacted or relied on for priority.