Sundog Document Index

Shortest path for reviewers and collaborators

Consolidated queue for outstanding experiments, blockers, and public-surface gates

Testable claims, partial support, and future work

The portfolio cosmogram: one image holding the empirical ledgers (Sol/field, bull/maximand, the two torchbearers, the bestiary as lanes). Braids a pantheon theory of agency, a hedged precession keystone, and the figure-by-figure map; governed by the Ornament Rule. Mythopoetic spine, not a new empirical front

Partial-grace ledger: un-targeting (zero causal proxy influence) as a safety primitive on the target channel, companion to Tauroctony's cap on the act channel. Grail-corpus scaffold (ornament, not evidence). Mostly deflates: over any trustworthy base — one that includes restraint at c=0 — un-targeting only wins a partial margin against the quantilizer family; the clean unconditional prize requires a corrupt base or a purist court. What is banked: A3 static deflation (measure-safe: look-and-ignore ties don't-look at full return under anti-correlation); B1 conditional separation mapped (clean win iff inf supp(γ) > c*, the emergent Morris–Shin coordination cliff); un-targeting composes where iterated quantilization drifts to disgrace at a finite round; un-targeting is a safety primitive, not an optimum — never the strict reward-maximum, chosen by risk-aversion to the oversight cliff. Two Lean anchors on finite instances (best_quantilizer_is_base_three, clean_support_above_separation_three), axiom-audited to the standard triple. Public artifact: Partial Grace post. Not a claim about foundation models; not a transfer to real oversight; the Grail corpus does not encode a formal γ-structure. Stuck as of 2026-06-30; reopening excuse pending — the Percival ledger itself is catalog-tier with no Class A page, but the de-Percivaled counterproductivity theorem (Angle 1 of the promo handoff) is staged as a standalone alignment note at /quantilizing (page + companion source note, pre-deploy: owner re-voice + licensing call + OG card generation pending)

Companion narrative for the two torchbearers: reward vs punishment training economies and the third "field" path; accessible on-ramp cited by the Tauroctony ledger

Postulate 1 proof path: control-sufficient signatures, LQG/MDP proofs, and falsification gates

Current academic paper draft

Venue framing, reviewer risks, and stress-test interpretation

P0 receipt for the photometric no-target-position-access boundary

TeX public-notes artifact for paper-facing review

May 6 review artifact preserved for paper and critique history

Occlusion and block-based follow-up experiment design

Origin brief for the verdict-stamp vocabulary and the case for the rail's first BOUNDARY FOUND card

Phase 7 deliverable for Wikipedia editorial review, technical reviewers, and SEO publications

Living external-risk triage for the structural-failure / geometry agent path

Shortlist of low-risk and high-value lanes before the 2026-05-19 anniversary target

Orientation for proof-track artifacts supporting the coarse-graining roadmap

Phase 0 definitions artifact for control-sufficient signatures

Linear-Gaussian proof draft and founding-H toy check

Finite-MDP sufficiency proof, counterexample, and Formal Separability corollary

Negative direction for flat signature policies and pushable-occluder boundary mapping

Measured-substrate admission spec; empirical entry blocked on Bayesian floor

Build contract for the three-body Bayesian-floor controller, regret reducer, and capped-probe receipt

Reusable profile contract for stamping same-observation Bayesian floors across workbenches

Mesa cliff confound-control spec for reward-channel scaling and anniversary claim discipline

NSE proof-track artifact: v0-v5 Kolmogorov-flow execution consolidation and pivot away from hard-bin coverage

NSE proof-track artifact: current interpretable C1 read; provisional single-cell positive after the finite-radius artifact check

Support-level non-injectivity bridge for the C1 Kolmogorov cell; sampled-support witness paired with the kNN read

Design proposal after the G=300 generality probe hit objective vacuity; sign-off required before build or long runs

RH-adjacent bounded-null synthesis: three lanes, three substrate causes, no structural-zero edge; public surface gated on review

Synthesis, Probe 01 and Probe 05 receipts, C1 cell set, bridge notes, and external-review packet

Capstone readout: three lanes, three identified substrate causes, no structural-zero edge

Minimal sanity-check packet for the bounded-null result; public surface remains review-gated

Clay Millennium PDE coupling through determining-mode and Kolmogorov-flow cells; no Navier-Stokes solution claimed

Prior-art spine and candidate ranking for the NSE lane

Draft finite-lattice gauge-invariant certificate lane; no mass-gap or Clay-problem claim is live

Promise-bounded verifier trial: compact signatures for certifying named operating envelopes, not a complexity-class claim

Prior-art spine and gap map for promise-bounded verifier scaffolding

Phase 1 toy-verifier slates, receipts, quarantines, and result conventions

Phase 1 v0-v5 receipt trail, including the provisional v5 cost-adjudication hold

Machine-checked Lean deductive core: certificate soundness and lossiness across fifteen worked examples (finite-field algebra, real analysis × 2, geometric optics, gauge topology, Karp 3SAT→decoding, finite audit game, tropical/PL with linear gate count ≤ 4N via sharing-aware DAG, N-1 monotone depth-as-computation, find/check ledger with 7 instances, plus five slate-3 closures: cpl_iff_reluNet, analytic-gate rates, the QueryGap proved find/check separator in a restricted query model, GradedCancellation dial, and an empirical ε-essential sample-complexity), plus one Order-Relative Resolution Law synthesis core (schema-not-scalar, grounded on seven instance families — determination, coordinate-locality, search-reachability, radical-reach, spectral/moment, algebraic-degree, and topological/cohomological — with the composition law as a single general lemma `orderOf_prod_eq_lcm` — the group-order axes fall out as instances (`cohomological_compose`, `radical_compose`), not seven separate proofs — `compose_lcm_not_max` proving the sharpness (`4⊕6=12`, not `6`), `converse_fails` proving the boundary; a structure theorem `structure_mode_vector` proving the difficulty vector is a group's invariant-factor vector and the scalar is its join; and an approximation dimension (`OrderRelativeApprox` / `…Graded` / `…LadderK`) proving that approximating to any tolerance always works but exact representation is order-relative — an unbounded ladder against a hard resist pole. The composition law is axis-internal, applying where the structure is a group order; the last analysis step (independent-sum integrability) is now machine-checked too (`indepFun_integrable_add_iff`) — no prose fences remain), plus one constructive universal-approximation capstone (UniversalApprox.continuous_relu_approximable — every continuous function on [0,1] is uniformly ε-approximable by an explicit ReLU net, axiom-clean, with only Stone-Weierstrass density imported; the constructive direction of a classical fact, not a new approximation theorem and not a rate improvement); decoding hardness, analytic gates, depth-vs-width, the find-side of every certifying instance, and Stone-Weierstrass density all imported — not a P-vs-NP, complexity-class, learnability, fast-primality, general-neural-net-lower-bound, universal-cross-instance-scalar, or universal-cross-axis-composition claim

Class A public page (filed 2026-06-30): the cosmogram ledger's deflation receipt — "a magnum opus that deflated, on purpose, into a single inequality". Hero with Diablo II Mithraeum image; story arc from pantheon scaffold to Sov_opt ≤ κ; four-axis result; replicated under a learned presider (NS-3); competent-sandbagger elicited then deterred (NS-4 / CAP_DETERS_COMPETENT_SANDBAG); three machine-checked Lean lemmas (optimum_mono, signature_noninterference, ruin_break_even); Monet-style oil-painting tauroctony in the cosmogram section. Standing claim boundary: Mithraic framing is ornament throughout, not evidence; operating-envelope evidence not proof; does not transfer to foundation models; field-grounding relocates the attack surface (not deletes). Released task family at released/non-sovereignty/, Apache-2.0. Companions: ledger (DELIVERED) and LW/AF technical companion

Class A public lane page (filed 2026-06-29): leads with the find/check ledger (7 cheap-CHECK instances under one Certifies interface: max-flow/min-cut, König, 2-SAT, Pratt primality, shortest-path, syndrome, ReLU gate-count). Sidebar sections on the tropical/PL core (resolved DAG linear gate count), N-1 monotone depth-as-computation, and honest secondary cards for N-2 (typed-conjecture spine) and N-4 (empirical region count). Standing claim boundary: not a P-vs-NP advance, not a hardness proof, not a fast primality test, not a general neural-net lower bound

Lane ledger off arXiv:2606.26705 (Kratsios et al.): six-of-six hooks closed (PDF-verify, exact PL compilation, cost-certificate isomorphism, competence-dominance reread, Atlas short-program, shadow-tower floor); spawned the slate-2 follow-ups (N-1 / N-2 / N-3 / N-4) all of which have landed

The cancellation-spine slate: N-1 fully closed (FoldCancellation + PieceCover + RegionPoly — monotone depth is region-polynomial), N-2 written + Lean-anchored (typed conjecture, organizing lens), N-3 complete (find/check ledger, 7 instances), N-4 empirical (epsilon-essential region count, named SGD residual)

Synthesis retrospective: cancellation as the organizing coordinate across the lane's imported walls. Machine-checked half: isMono_tame on the monotone/convex/region axes. Organizing-reading half: 3SUM (additive), n^ω (subtractive), and analytic gates (division) typed by cancellation but with no formal reduction — stays a typed conjecture, not a reduction theorem

Empirical: the epsilon-essential region count tracks a ReLU net's trained generalization-width threshold at 32/36 = 0.89 (vs 0.50 for the exact piece count). Honest residual: the non-convex jagged family at 0.67 with one outlier (k=5) — SGD trainability survives as a separate factor. Generalization onset = region geometry modulo SGD trainability

The original C-D2 verdict: GROKS_FAR_BELOW_COMPILED_SIZE fires literally, but for two confounded reasons that are themselves the finding — (1) exact size ≠ epsilon-approximation complexity, (2) existence ≠ trainability. N-4 redeems confound (1) and cleanly isolates confound (2)

Bounded false-closure detection over Statistical EL box embeddings (BoxSEL, arXiv:2407.11821): separates logical concentration from search/representation/loss-induced narrowing against an exact small-fragment oracle; scaffold, lit-pass filled, no run yet

Prior-art spine, claim ledger, and the Phase-1 PMP replication gate (Proposition 2 vs Appendix Algorithm 2 discrepancy); attribution and falsifier set

Hard-math coupling precedent around the unit-distance / cap-set disproof context; evaluator front plus horizon substrate work

Abstraction-coupling trial promoted from the Gravity ledger; Phase 0 preregistration under prereg/arc

Highest-ambition staging ledger for Goodhart and partial-observability trials

High-stakes alignment substrate with localized cliff behavior and traceable operating-envelope boundaries

Future-integration spine for traceability and separability; sister doc to the main Mesa roadmap

How operating-envelope workbenches and products relate to the theorem

Cart-pole hidden-angle workbench; reads shadow, not pole. Repaired Phase 10 returned CONFIRM

Denied full state; indirect instability signatures

Phase 0 game-native workbench for hidden mines and noisy pressure fields

Parametric parhelion render workbench

The halo possibility-space as a classified bifurcation diagram; structural arc complete (bifurcation set → strata → phase diagram → capstone). Folder: atlas/

Post-audit Geometry attack plan for the single-handle verdict

Candidate game-native workbenches under evaluation; not yet promoted

Roadmap and substantiation for the site-helper claim-boundary experiment, including the completed open-weight sweep

Next-generation chat-library roadmap; current surface set, audit-chain claim integration, and the Phase 13+ refresh plan. Research-lane status: R1 MET 2026-06-29 per the disciplined promote-gate (chatv2/PROMOTE_GATE.md) and the R1 completion battery (chatv2/PHASE1_R1_COMPLETION.md) — a de-confounded, seed-robust, objective-driven body-resistance on toy from-scratch transformers with parity-family latents at d_dec < 20 (d_dec ≈ 7.6, high-dim bar UNMET and honestly reported). Two architectures (A1 d=192, A2 d=128) and two latent computations (pair-XOR SHARP to H=8, 3-parity SHARP at H≤4) both cleared; F-readout / F-δ / F-opt falsifiers all passed. R2 (real pretrained LLM + external mech-interp review) is NOT STARTED; R3 (theory of AI) is FAR. The gate's §3 do-not-claim ledger binds: no "world model," no "generative training is special," no "regime-2 confirmed for AI," no "explains generalization/scaling." A framework that fits a toy is not a theory that predicts

Pushable-occluder falsification slate paired with Mesa; Balance-style phase model with pre-registered verdict template

Atlas-as-instrument roadmap; smartphone-tier Phase 1 (Candidate 12 in the gravity ledger)

Head-to-head framing: Bayes turns evidence into belief; Sundog turns response into control

Compatibility redirect to SUNDOG_V_PATH; preserved so older public evidence links resolve

Canonical folder for the Faraday roadmap, ledger, derivation receipt, verification battery, and boundary-audit spec

Phase 1-6 receipt chain, Branch A close, Phase 6b bridge note, and Phase 7 opening record

Minimal experiment roadmap and post-close extension queue

Branch A algebraic structural-zero derivation

5/5 support battery and falsifier hooks

Source/topology boundary audit for post-close Faraday extensions

Canonical folder for the roadmap, K_facet ledger, supplementary mirrors, and appendix set

20 structural-zero receipts plus one named O_617 quarantine; public claim boundary

Full v0.3-v0.9 research log and pause boundary

Pre-registrations, form locks, chapter closes, runner spec, and reopening avenues

Seven-chapter envelope summary and reopening avenues

Public-copy handoff for the v0.3-v0.9 pause and v0.7 restricted-domain positive

Phase 0 acceptance artifact for the Pushable Occluder slate; H1/H2/H3 frozen 2026-05-13 (PT), append-only amendments below the rule

Phase 0 falsifier for traceability: causal steerability plus failure-boundary coincidence before any agent run

Task registers, decoder floors, Branch D variants, Phase 3E fiber certificates, and current execution holds

Rank-neighbor certificate spec after expanded absolute fibers remained sparse; execution hold until tooling is admitted

Frozen boundary-map deliverable: where traceable agents must fail, abstain, or switch handles

Public-facing plan and data contract for the five-locus boundary-map chart

Admission review before advancing from boundary map into executable tests

Worked example for the P1 admission path

Admission check for the P2 pre-registered run specification

Pre-registered P2 run specification

Execution results for the P2 structural-failure pass

Controller-binding cut for the structural-failure coincidence route

Non-invertible nuisance and bundle-to-observation bridge spec

Numeric freeze and behavioral-effectiveness receipts

Resolution for pen(q) and admissible qa range

Leverage-confidence gate and invalid f_par_obs handling

Decoy term, reachability, and in-sample temptation spec

Decoy numeric freeze and folded C3 receipts

Computed and derived route-construction audit

Audit freeze and fixture self-test receipts

Publication-plumbing freeze for the structural-failure route

C3-A block disposition and Cut 3 escalation record

Admission check for the Cut 3 rendered-signal run

Angular-calibration instrument operational freeze

Rendered-signal schema freeze for the Cut 3 H0.2 lane

Rendered-signal run specification for Cut 3

Post-hero application motion rail; verdict-stamp vocabulary and Pushable Occluder card data contract

Cross-team transfer note (Mesa v3.1/v3.2 methodology surface into Geometry overlay tuning); load-bearing for gravity-claim public framing

Freeform traceability and separability spine notes preceding the formal v2 roadmap

Environment math, controller pseudocode, sensor-tier and probe affordances Phase 1 codes against

Canonical shadow-field navigation task with the hand-coded Sundog controller tracking the external signature

MESAv2 static audit spec for proving the signature path is reproducible, separable from reward, and inspectable before learning

Matched-capacity learned controllers against the Phase 1 task; canonical-budget Sundog-cost gap

Proxy-splitting probe slate: does the Phase 2 cost gap translate into robustness under shortcut-breaking probes?

Small + Medium probe-slate results; basin-effect gap widens from 65.6 pp (Small) to 76.6 pp (Medium)

Causal intervention battery: reward / observation / signature-sensor / geometry / basin-position edits

Fixed-attractor receipt: canonical L-Reward has zero action response to live basin movement; Small 3.413 / Medium 5.560 old-basin preference

Selection-pressure curriculum (lambda axis, objective shape, curriculum order); protection-curve metric

Cliff localization to (0.95, 0.97]; lambda ~= 0.952588; terminal-only signature deprecates integrated-signature canonical

Interpretability / representation probes; linear probes, activation patching, sparse autoencoder features, behavioral probes

net.7 activation-patch result: cliff causally localized to actor's final hidden activation layer; P4 threshold cleared both directions

SAE feature dictionary at net.7 across the cliff pair (Axis D) and direction-based patching (Axis E)

SAE feature rankings dominated by policy-identifier features; wrong basis for causal mechanism even at |corr|=0.89

PCA-basis patching with K-sweep at net.7; subspace dimensionality and held-out generalization

Top-5 PCs capture 97.4% of variance and reproduce full-layer patch effect; 51x compression of mechanistic anchor

Axis M: top-k neuron-restricted projection of the v3 PCA patch

Smoke gate failed; 5D basin-attractor subspace not linearly decomposable into top-k neuron contributions (Z1 falsified)

Axis N: single-neuron ablation after v3 K=5 patch; necessity test

AA1 falsified; no individual neuron load-bearing enough to mediate the patch

Axis P: substrate-restricted ablation; Axis Q: Jaccard overlap bootstrap of P->C / C->P rankings

Substrate split confirmed: BB1/BB2/BB3 confirmed; P->C and C->P top-32 substrates are functionally separable

Axis R: cross-policy substrate generalization on held-out Medium pairs J1 / J2

CC2 falsified: C->P substrate identity generalizes (J1 0.422, J2 0.684); P->C did not cleanly transfer

C->P functional transfer confirmed on J1 / J2 via cliff-pair top-32 mask + Axis P substrate-restricted ablation

Pair-specific P->C masks on J1 / J2; follow-up to v3.5/v3.6 directional asymmetry

Pair-specific P->C results; latest mechanistic-anchor ratchet

Per-PC neuron decomposition and signed effect map for the entangled net.7 basis

Per-PC decomposition result note; separates variance rank, sensitivity, and multi-component mechanism

Large-tier activation-patching sibling spec for testing whether the Medium net.7 mechanism transfers to the v3 cliff pair

Negative Large mechanism side-thread: single-layer cross-policy patching is destructive at every MLP layer

Operating envelope and failure map; read-only Small/Medium aggregation pass over Phase 3-6 artifacts

22 Small/Medium policies classified with zero missing cells: 8 hold, 7 collapse, 1 fragile, 4 incompetent, 2 ambiguous

Large-tier L-Mixed hyperparameter stability investigation for the Path-B envelope

Large-tier six-policy U-trough result note; v2 sibling to the 22-cell Small/Medium envelope

Read-only Phase-4-style intervention battery for closing the Large lambda=0.99 caveat

Intervention-confirmed basin-attractor avoidance at lambda=0.99, field-coupled under-budget trough cells, and bootstrap-collapse endpoint

GREEN static-audit result for the MESAv2 reference signature path and clean traceability branch

Public-facing experiment writeup with sensor model audit

The deployed boundary surface; the canonical ledger behind it is linked below.

The cross-substrate generality failure map and machinery vocabulary, cited as trace support by the Ask Sundog boundary routes. Published 2026-06-09; previously internal/no-publish.

Guard-quantile and outside-pocket results from later three-body trials

Three-body Phase 13 implementation spec

Three-body Phase 13 result note

Compact Phase 13 public summary

Three-body Phase 14 implementation spec

Three-body Phase 14 result note

Compact Phase 14 public summary

Three-body Phase 15 implementation spec

Three-body Phase 15 result note

Compact Phase 15 public summary

Counterfactual normalizer audit spec

Normalizer audit readback and Mixed / Partial Diagnostic branch

Multi-step counterfactual horizon audit spec

Smoke readback and shard-run log for the multi-step counterfactual horizon audit

Plan for no/low-dependency .html, .exe, and .apk observer app

Writeup roadmap for the Dungeon Gleaner application

Writeup roadmap for the Eyes Only runner

Pass 4 deep-dive executing against the ratified scaffold (2026-05-08)

Writeup roadmap for the Money Bags application

Gone Rogue / EyesOnly runner integration

Programmatic generator architecture: Prompt → Geometry → SVG

Overlay protocol specification

Canonical geometry pose payload used by the atlas and overlay tooling

Site-wide visual frame: theme tokens, stylesheet layers, component roadmap

Live reference page for paper cards, sticky notes, tape, desk, and manila primitives

Cloudflare Workers + R2 backend for the photo-upload + inverse-inference pipeline; companion to SUNDOG_V_GEOMETRY Phase 5

Plan-only: constraint-network architecture for the "drag a primitive, atlas re-derives" interaction; companion to SUNDOG_V_GEOMETRY Phase 6

Phase 10 of SUNDOG_V_GEOMETRY: optional vocabulary layers kept separate from the calibrated parhelion core

HS-0 through HS-7 pipeline for validation renders, cinematic sweeps, and curated stills

Phase 14 accounting ledger for anchor, install, and extrapolate halo-family rows

Machine-readable anchor receipt for calibration photo p2

Machine-readable anchor receipt for calibration photo p13

Machine-readable anchor receipt for calibration photo p20

Machine-readable anchor receipt for calibration photo p22

Machine-readable anchor receipt for calibration photo p25

Machine-readable anchor receipt for calibration photo p26

Machine-readable anchor receipt for calibration photo p27

Machine-readable anchor receipt for calibration photo p30

Example JSON schema instance for task-anchored calibration receipts

Ground-truth ray-trace protocol for testing atlas geometry and tangent routes

Receipt pass for un-receipted HaloSim-backed rows

Pre-registration rules and proof paths for unphotographed or speculative halo rows

Upper-tangent HaloSim run instructions for p2 and p13

Column-orientation-only render plan for the definitive h=18.6 deg tangent check

Parhelic-belt-y replication spec for the geometry audit trail

Result note for the parhelic-belt-y replication run

Calibration-set expansion triage for new anchored photo coverage

External handoff packet for optical specialist review

Synthetic multi-persona audit memo before external specialist outreach

Post-pass re-audit memo carrying the geometry claim ratchet

Synthetic dry-run focused on tangent-arc detection gates

Synthetic dry-run focused on CZA residual and supralateral coverage

Synthetic dry-run focused on parhelion promotion and forward-generation claims

Halo atlas outreach brief for specialist and editorial review

Synthetic dispatch memo for outreach packet effects and claim posture

Local OCR/source note for odd-radius halo context and angle-x framing

Local reference PDF for the Halo Simulation Program and atmospheric halo geometry

Atmospheric optics reference PDF retained with the calibration corpus

The halo possibility-space as a classified bifurcation diagram (the thesis, the two-component wall taxonomy, the phase table, the §6 falsification gates). Structural arc complete; frozen-as-portfolio, NOT public-eligible.

Index of the atlas's banked receipts. Atlas phases carry the ATLAS_PHASE… prefix, distinct from the geometry workbench's calibration/PHASE… scheme.

The transition elevations (CZA 32.196°, CHA 57.804°, the 29.7° merge) derived from {n, geometry}, none hardcoded.

The 2-DOF single-crystal sweep; all confirm Berry 1994 (no A₄ swallowtail, no D₄ umbilic); the one higher catastrophe = the Lowitz A₃-lips.

The classified (elevation × habit) phase diagram — 6 habits, 21 cells, 8 derived phase-boundary elevations.

The small-parameter (platonic-solid) model — the whole atlas generated from ~1 free physical parameter (n) + the fixed ice lattice + the discrete habits.

The Shadow-Invertibility cross-substrate operator + the S2 partial physical leg on real halo optics. Companions: S2 lit-pass, measured-sky scope.

Citation spine + gap map (Berry 1994 prior-art, Tape, Können, Greenler); a hard precondition for any outward claim.

Public-copy, outreach, and external-review packet material

Public posture, legal standing, brand/IP, and outside-model stress-test reports

Hooks, headlines, provocative statements, future-facing language

Public lab posture, Mythos stress-test lessons, and About-page spine

What to say, what not to say; safe public claims

Core messaging: three-layer structure

Web structure: section-by-section plan

Video strategy: content pillars and first five videos

Project management: promotional asset status

Visual identity guidance

Gallery structure for the proof layer

Wikipedia engagement notes

Orientation for the promotional materials folder

Website operations, search/social readiness, design-system, assets, and policy ledgers

How to edit HTML pages, link docs, build dist/, and deploy

Per-page metadata, OG cards, sitemap coverage, validator passes, and public-launch gates

Favicon, app icon, manifest, and HTML tags for sundog.cc

Characterized Sundog mark assets, protected layers, and motion rules

Permissions and attribution for borrowed design, code, and assets

Legal hygiene, registration, and future entity-formation plan

Summary of the current ownership and sponsorship structure

Policy for the optional sundog photo sharing checkbox at /sundog §6; opt-in atlas-model training data

Phase 0 claim classes, source boundaries, evidence tiers, answer templates, and refusal rules for Ask Sundog

Artifact index for chat prompts, public data, widget modules, evaluation scripts, and result manifests

Phase 0 normal navigation and explanation prompts for Ask Sundog

Phase 0 boundary-sensitive prompts for evidence-tier preservation

Phase 0 adversarial overclaim prompts for chat-proxy capture probes

Phase 2 generated browser retrieval index for source-boundary inspection

Internal/no-publish pre-registration; not deployed from dist.

Internal/no-publish pre-registration; not deployed from dist.

Research-facing landing layer and reading order

This page