← All the evidencetheory-not-formalized-2026-10-29.md
SHA-256 30b1a96f20caab6b9b5c661d72485724887c2aa06a83c442fd710b52df810c2a · download the original
# Claims not formalized, 29 October 2026
Every claim the four 29 October coverage files record as not proved, per note, with the reason. Source: `docs/coverage-2026-10-29/*.json`.
## Evidence.lean
### How-Sentinel-Chooses-a-Safe-Test-When-Explanations-Agree.md
- Zero extra probes selected per situation; decision threshold P(H=1|E)>20/34 from utilities. Not proved because: simulation result from Python; the threshold depends on utilities the note chooses, only the general EVSI inequality is proved
- argmax over permitted probes is optimal. Not proved because: Propose/permitted-set operations unspecified; proved only EVSI nonnegativity and the fixed-decision case
### How-Sentinel-Discovers-That-Its-Own-Senses-Are-Wrong.md
- Residual correlation matrix C_ij identifies faulty sensors; mixture model with outlier probability p_g. Not proved because: correlation estimation and robustness claims are statistical, no observation model given; proved only bias-shift non-identifiability
### How-Sentinel-Finds-Independent-Evidence-When-All-Senses-May-Be-Wrong.md
- Delay-aligned witness scores S(j,l) find independent evidence; reported recovery rates. Not proved because: empirical Python simulation; delays and witness model not specified mathematically
### How-Sentinel-Invents-a-New-Source-of-Evidence.md
- Candidate subtract/multiply compositions find a new evidence source; I(W;X)=0 when F=g(X) and H independent. Not proved because: the zero information case is covered by existing independent_noise_zero_information; the search over compositions is a procedure with empirical results
### How-Sentinel-Invents-an-Observation-Protocol.md
- Primitive grammar WAIT/READ/DELAY/SUBTRACT/MULTIPLY discovers a protocol; selection by argmax value. Not proved because: placeholder operations and empirical results; only the XOR ambiguity that motivates it is proved
### How-Sentinel-Knows-When-Evidence-Is-Redundant-and-Finds-a-Better-Source.md
- Conditional information gain G_t(s)=I(H;O_s|E,C) and redundancy D_t; XOR example I(H;X1)=0, I(H;X1|X2)=1. Not proved because: XOR example is mapped to existing xor_single_information_zero and xor_joint_information_exact; reported recall gain of 8.89 points is empirical
### How-Sentinel-Learns-Which-Senses-to-Trust-When-Groups-Disagree.md
- Estimate P(V1,V2,V3|H) as a smoothed table and detect drift Delta_t. Not proved because: table normalization proved; the statistical estimation quality and drift detection rates are empirical
### How-Sentinel-Makes-Useful-Discoveries-When-Past-and-Present-Are-Uncertain.md
- Hypothesis H=(m,delta,z,v,g) posterior, Net_t(d)=E[U]-C costs, selective abstention. Not proved because: hypothesis space and utilities unspecified; only the chain rule factorization is proved
### Sentinel-Natural-Calibration-FRESH-2026-10-29/How-Sentinel-Discovers-Its-Own-Calibration-Opportunities.md
- Dependence statistics G(p), V(p), CalValue_t(p); simulated calibration with five cue channels. Not proved because: empirical simulation and unspecified dependence measure; only CUSUM recursion and the shared-fault algebra are proved, and the CUSUM change-detection delay is not proved
## Timing.lean
### How-Sentinel-Discovers-Cross-Sense-Connections-Without-Paired-Examples.md
- Dependence-maximizing search over transforms, lags and streams finds true links. Not proved because: empirical and depends on an unspecified dependence measure Dep
- Context-residual correlation recovers delayed nonlinear links in the simulation. Not proved because: Python experimental result
- Predictive versus causal bridge probabilities q_t. Not proved because: probabilistic model not constructed in the note
### How-Sentinel-Discovers-Event-Correspondence-and-Time-Alignment-Together.md
- Joint argmin objective over transforms and clock, and posterior q_t(phi,psi,omega,pi). Not proved because: note says the model is not built or validated
- Recovery under 52 and 57 percent dropout and the offset-only ablations. Not proved because: Python simulation percentages
- Event detectors discard evidence a later clock estimator needs. Not proved because: qualitative; only the period-shift ambiguity was proved
### How-Sentinel-Discovers-a-Common-Timeline-When-Senses-Disagree.md
- Smooth match score J(alpha,b) recovers the clock from unpaired events. Not proved because: only affine algebra and non-identifiability proved; recovery is a Python result
- Hidden alpha in [0.990,1.010], beta in [-24,24] estimator accuracy. Not proved because: Python simulation results
### How-Sentinel-Maintains-Continuous-Thought-Without-Constant-Computation.md
- Jump update h(t+)=h(t-)+K(x,h), learned K. Not proved because: K is unspecified and unresolved; only decay between updates proved, scalar/diagonal componentwise
- Wake rule scheduling in practice. Not proved because: only the algebra of the wake time proved; no claim about usefulness
### How-Sentinel-Learns-When-Thinking-Is-Worth-Its-Cost.md
- Value of computation Q=pv-c optimal knapsack allocation. Not proved because: only the nonnegativity of VOC and the sign rule were proved; optimal selection and the empirical gains are not
- Beta estimate with 0.994 decay beats alternatives. Not proved because: empirical tuning result
### How-Sentinel-Suspends-and-Resumes-Unfinished-Thoughts.md
- Bellman value of continue, suspend, retire and optimal resumption. Not proved because: needs transition and reward model the note does not specify; only discount monotonicity proved
- Sequential log-odds learner with p(z|Y)=0.77/0.23 and step estimate n-hat. Not proved because: simulator assumption and Python results
### How-Sentinel-Compresses-Unfinished-Thoughts-Without-Losing-Their-Future-Usefulness.md
- Robust compression objective and reindex rule E[retrievals]*v > c. Not proved because: policy claims without a specified recovery model; only the counting impossibility proved
- Compressed unfinished thoughts retain future usefulness. Not proved because: empirical claim from Python experiments
### How-a-New-Experience-Reconstructs-an-Old-Thought.md
- Ridge translation T-hat and top-K retrieval recover old experiences. Not proved because: Python result; pairing supplied by the simulator
- Action choice argmax of future value minus costs. Not proved because: unspecified expectation terms; only effective sample size proved
### How-Sentinels-Attention-Caps-Change-as-Reality-Changes.md
- Interest curve counterexample at tau=0 rejects it as a ranking rule. Not proved because: interest ceiling and monotonicity map to existing interest_ceiling and interest_increases_with_exposure; operationalization of tau is empirical
- Slow trace dl/dt=(f-l)/tau and the exponentially weighted integral. Not proved because: maps to existing trace_matches_clipped_update, trace_time_subdivision and blend_bounds; time-varying f integral form not proved
- Caps change as reality changes with budget B_t. Not proved because: only fixed-budget arithmetic proved; how B_t varies is empirical
## Values.lean
### How-Sentinel-Learns-What-Is-Worth-Interrupting-Me-About.md
- Offline bounded learner improves interruption decisions; flawed silent-labels learner becomes nearly silent. Not proved because: Python simulation result; the silence-as-negative-label failure is empirical (the non-identifiability behind it is proved).
- Potential outcome Y_t(1) causal identification and randomization not fixing nonrandom missingness in general. Not proved because: Only the binary-outcome selective-feedback inequality was proved; general causal identification needs an observation model the note does not give.
- Budget 8 notifications per 100 candidates. Not proved because: Parameter of the simulation; only the generic cap A_t<=P_t, sum A_t<=B was proved.
### How-Sentinel-Preserves-a-Discovery-Before-an-Opportunity-Disappears.md
- Deadline-aware model false alerts 56.55 and audited 28.81 per 1000; 35% opportunity base rate. Not proved because: Python simulation percentages.
- Best action argmax over silent/tentative/verify is optimal (U_V uses U_verified, p_{t+Delta,z}). Not proved because: U_verified and the posterior expiry model are unspecified; proved only the nonnegative gross value of verification with affine utility and the silent floor.
- No algorithm gets both perfect timely recall and low false-alarm under identical observation distributions. Not proved because: Proved only through the decoder non-identifiability form (existing identical_observations_cannot_decode_distinct_targets); the rate-equality statement with probability measures P(A=1|Y=1)=P(A=1|Y=0) was not stated over a measure space.
### Sentinel-Behavioural-Meaning-Drift-Research/How-Sentinel-Knows-When-Its-Behavioural-Interpretation-Has-Changed.md
- Standardized residual Z_g(t) detector and D_t disagreement monitor; false-alarm 6/64, delays 40.5 and 107 steps. Not proved because: Python results; the note itself says the statistic is not calibrated.
- P(O|W1)=P(O|W2) for the paired worlds as a probability statement. Not proved because: Proved the deterministic observation-equality form (flip_nonidentifiable); the likelihood-equality form over sequences was not formalized.
- Value change versus observation-mechanism change is not decidable from the observation stream. Not proved because: only the narrower logical fact above is proved; no observation or listener model is formalized
### Sentinel-Experiential-Continuity-Research/How-Sentinel-Discovers-Which-Experiences-Still-Mean-the-Same-Thing.md
- Joint alignment objective S(pi1,pi2) accuracy numbers, consensus abstention, bootstrap interval 5.60 to 12.89, relabelling invariance of tie-breaking. Not proved because: Python results; only the count of 24 bijections was proved. Relabelling-invariance of the consensus rule was not proved.
- Role rotation versus readout rotation cannot be told apart from the observations. Not proved because: only the narrower logical fact above is proved; no observation or listener model is formalized
### Sentinel-Personal-Meaning-Consequence-Research/How-Sentinel-Learns-Which-Differences-in-Personal-Meaning-Matter.md
- delta = E[Y|G=1]-E[Y|G=0] is not causal; tau(E_t) and argmax_q over permitted questions plus empty. Not proved because: Not identified from passive data by the note itself; the silent-option floor was proved in the generic argmax form.
- AUC +0.11634, precision 33.18%, net value +49.13 per 1000, five clarifications per 100 cap, adaptive decay estimator. Not proved because: Python simulation results; exponentially decaying weights beyond shrinkage bounds not proved.
### Sentinel-Personal-Value-from-Weak-Signals-Research/How-Sentinel-Learns-Personal-Value-from-Weak-Signals.md
- Learning personal value from 352 weak labels, 7.5% corrupted. Not proved because: Python simulation; depends on the sparse anchor.
- p_hat = q*p_recent+(1-q)*p_long with q untested. Not proved because: Covered by existing blend_bounds (bounds in [0,1]); q itself has no definition to prove anything further.
- P(S observed|V,E) != P(S observed|V) selection. Not proved because: Stated as a possibility, not a theorem; the binary-outcome bias inequality is proved under interruption.
### Sentinel-Values-vs-Circumstances-Research/How-Sentinel-Distinguishes-Changing-Values-from-Changing-Circumstances.md
- Eight-state filter accuracy, guarded discrepancy rule, y_hat_context threshold. Not proved because: Python results with hand-chosen thresholds.
- Y = V and A: Y=0 has two explanations. Not proved because: Covered by existing action_does_not_identify_preference; the posterior consequence is in blocked_action_value_survives.
### Sentinel-Constructive-Meaning-and-Philosophy-Research/How-Sentinel-Learns-When-Different-Meanings-Help.md
- D>0 does not imply H=1 or C=1; two people with identical evidence have identical output distribution. Not proved because: Trivial non-implication; same-evidence-same-output follows from existing identical_observations_cannot_decode_distinct_targets for deterministic algorithms, the stochastic version is not separately stated.
- Quote interpretation Q_i, bias +0.084, 3.72 interruptions per 1000. Not proved because: Empirical or design statements, no proof content; only the permission-gate of the softmax policy was proved.
### How-Personal-Significance-Learns-and-Changes-Through-Experience.md
- Total probability identity P(A|V)=sum_{g,c} P(A|V,g,c)P(g,c|V). Not proved because: Covered by existing mixture_bounds and markov_mixture_normalized (weights summing to 1, outputs in [0,1]); the conditional-probability identity itself was not restated.
- Held-out accuracy, 90% confidence wrong 2.50%, 153 shifts, shuffled-constraint collapse. Not proved because: Python results.
- Impact functional S_t(e) and correction as operational constraint. Not proved because: Impact is explicitly unspecified; the operational permission gate is proved only as policy_permission_gate.
### Sentinel-Constructive-Meaning-and-Philosophy-Research/Philosophical-Quote-Interpretation-Template.md
- Template. Not proved because: No mathematical claims.
## Language.lean
### Sentinel-Conversational-Experiential-Semantics-FRESH-2026-10-29/sentinel_conversational_clarification/How-Sentinel-Discovers-the-Right-Clarification-as-Meaning-Evolves.md
- Results on 12 identities, 10 word ids, 30,000 interactions (28.53 percent same referent different association), recent-calibrated model comparison. Not proved because: Python simulator output and empirical percentages.
- The do(q) causal effect of a clarification and that asking improves later exchanges. Not proved because: Stipulated by the simulator; an unprovided observation model.
- E[U | do(q)] - c(q) argmax over permitted questions. Not proved because: Proved only in the narrowed finite form: best value with a silence option is monotone in the option set and silence wins when all values are lower; the conditional expectation itself is unspecified.
- Recommending a clarification does not give permission to message others. Not proved because: Policy statement, not mathematics.
### Sentinel-Intent-Discovery-Through-Dialogue-Research/How-Sentinel-Tests-Intent-Discovery-Through-Dialogue.md
- Held-out utilities, goal accuracy, question rates, bootstrap intervals, eight environment tables. Not proved because: Python results.
- Joint question and expression selection is not better than two-stage; elaborate hypothesis not earned. Not proved because: Empirical comparison.
- Value of information differs from entropy reduction. Not proved because: Proved only that value of information is non-negative and zero when the best action is constant; no entropy comparison was formalised.
### Sentinel-Interpersonal-Experiential-Semantics-Research/How-Sentinel-Discovers-When-People-Mean-Different-Things.md
- ROC AUC and Brier values for four methods, 0.745 vs 0.551, prevalence 62.8 to 39.1 percent. Not proved because: Python results.
- Accommodation parameter fitted by maximum likelihood; Brier improved in 36/36 worlds while AUC fell. Not proved because: Empirical result under a known noisy-action model not given in closed form.
- Emotional significance and private feeling are separate and unobservable. Not proved because: Philosophical and privacy claim; the observational-equivalence core is covered by same_history_one_world_wrong and identical_inputs_brier_floor.
- Social field H_t = P(S,L,A_p,A_q,G,R | E_0:t). Not proved because: Stated as a research target, not an implemented distribution.
### Sentinel-New-Expression-and-Shared-Meaning-Research/How-Sentinel-Tests-Whether-a-New-Expression-Becomes-Shared-Meaning.md
- Echo(u) does not imply Understands(u). Not proved because: Proved only as an information limit (same_history_one_world_wrong); Understands has no formal definition in the note.
- Policy tables for replay, echo, transfer and joint transfer, 25 percent reuse cap compliance. Not proved because: Python results and simulator settings.
- Selection rule max over |U_t| <= B_t of sum (E[Delta U] - cost) subject to permission. Not proved because: Proved only that nothing beats the empty selection when every net gain is non-positive; E[Delta U] is unspecified.
- Copied expression, transferred reference and subjective understanding are distinct; no threshold makes dependent evidence independent. Not proved because: Conceptual and provenance claim; independence of evidence is an empirical property.
- Copying behaviour cannot tell a compositional listener from an imitator when the history is identical. Not proved because: only the narrower logical fact above is proved; no observation or listener model is formalized
### Sentinel-Personal-Expression-Research/How-Sentinel-Constructs-Language-from-Personal-Experience.md
- Held-out semantic fidelity, novelty and abstention results; semantic reversal drops frozen accuracy to 50.46 percent. Not proved because: Python results.
- u* = argmax E[U(u; G, C, H, S)] and P(u | ...). Not proved because: Research targets with unspecified expectation; only the finite silence-inclusive argmax property is proved.
- Authenticity and social appropriateness of generated phrases. Not proved because: Requires consented real interactions; not mathematical.
### Sentinel-Personal-Language-Meaning-Drift-FRESH-2026-10-29/How-Sentinel-Learns-When-Words-Change-Meaning.md
- Spectral initialisation with discrete local refinement recovers the personal codebook; held-out and streaming results. Not proved because: Algorithm behaviour and Python results; only the global-flip invariance is proved.
- Delta_i = Pr(O != Z_hat | w_i, recent) - Pr(O != Z_hat | w_i, history) distinguishes a local change from a broken witness. Not proved because: Detection power is empirical; the statistic is a difference of probabilities, no theorem beyond the identifiability limit was claimed.
- 8 percent unobserved pragmatic variation noise model. Not proved because: Simulator setting, no derived property.
### Sentinel-Shared-Experience-and-New-Expressions-Research/How-Sentinel-Learns-New-Expressions-from-Shared-Experience.md
- Fusion z = norm(u + 0.85 v) when <u,v> >= 0.40, else u, and withholding on weak or ambiguous signatures. Not proved because: Proved only that the fused vector is nonzero; whether the rule improves meaning recovery is empirical.
- Comfortable continuation threshold 0.55 as a proxy for comfort or consent. Not proved because: Smoothed estimate bounds are mapped to an existing theorem; the proxy's meaning is an unprovided observation model.
- Ordinary, difficult and adaptation results; shared-fault failure. Not proved because: Python results.