Skip to content

IS_AND_IS_NOT.md

Latest commit

 

History

History
421 lines (377 loc) · 20.5 KB

IS_AND_IS_NOT.md

File metadata and controls

421 lines (377 loc) · 20.5 KB

DM3 — IS / IS NOT Ledger

Last updated: 2026-05-02 (Session 7 closeout + Session 8 Phase A/A5/B3/A6 + Phase G v2 partial chain + reconstruction Tier-2)

This ledger captures scoped positive and negative statements about what DM3 is and is not, as evidenced by receipted experiments across Sessions 1–8 (live Phase G v2 chain in flight; backwards reconstruction lane at static Tier-2).

DM3 IS

Object

  • A precompiled Rust binary, dm3_runner, SHA-256 daaaa84a052b60523bf9d63152f1154225abf119c279aa4b3aabf14487279672, executable on Red Magic 10 Pro.
  • A 380-vertex, C3-symmetric graph constructed via exact-rational 2D Sri Yantra plus a 3D toroidal-twist lift.
  • A 72,960-dimensional (380 vertices × 192 features) relaxation dynamical system.

Dynamics

  • The harmonic task is bistable with HIGH basin (E≈88, Coh≈0.77) and LOW basin (E≈75, Coh≈0.88) at default parameters.
  • The holography task is monostable in a RETRY basin (E≈15, Coh≈0.72 at asym=0).
  • Basin selection at default harmonic parameters is IID Bernoulli with p(HIGH) ≈ 0.34.
  • Across the Session 7 dynamics battery, harmonic p(HIGH) remains statistically stable across six measured arms and 665 total episodes: S2H baseline (N=250), S7 cold (N=100), S7 hot (N=100), S8 battery (N=30), S8 bypass (N=70), and S5 basin volume (N=23). All arm-level Wilson 95% confidence intervals overlap the Session 5 baseline [25.5%, 43.7%].

Capability surface

  • The binary exposes a self-evaluating six-gate campaign via --task exp_r1_r4_campaign.
  • At defaults, three gates PASS (EPSILON_CRIT, R4, WAKE_SLEEP_ALIGN) and three gates FAIL (R1, R2, R3).
  • The campaign is deterministic up to run_sec.
  • The Session 7 smoke receipt fixes a default --steps 5 reproducibility fingerprint at canonical SHA 9006df4ec02c8872b2037ce49ba9f2e9f27cfb7b92f62dfea5e7982d6be7d912.
  • R1 flips when --adj RandomAdj_v1.bin is used.
  • R2 flips when --tags RegionTags_v2.bin is used, and claim_level advances CL-0 → CL-1.
  • The 2×2 cross-control table is closed: R1 responds to --adj, R2 responds to --tags, and the axes remain separable at default --steps 1.
  • The combined-axis configuration RandomAdj_v1.bin + RegionTags_v2.bin + --steps 50 flips R1 and R2 together and raises r4.transfer_ratio from 1.369 to 2.678.
  • Session 7 final reporting closes a bit-level gate-layer substrate-null battery across the tested smoke conditions. The local checkout keeps the smoke fingerprint and harness tree visible, and keeps the standalone directory-name mismatch explicit in REPO_AGENT_FINDINGS.md.
  • R3 remains unreachable from the exposed CLI on the SY-default surface; requested --steps > 20 does not raise operational_steps above 10.

Hidden task inventory

Beyond harmonic and holography, the binary accepts 10 additional task names:

  1. interference
  2. holographic_memory
  3. exp_r1_r4_campaign
  4. exp_i1
  5. exp_i2
  6. exp_h1_h2
  7. exp_k2_scars
  8. exp_k3_truth_sensor
  9. resonance_r3
  10. resonance_v2

Session 7 / Session 8 learning line

  • exp_k2_scars is the first receipted positive learning line in Sessions 3–7. best_uplift rises 0.010 → 0.075 → 0.273 → 1.324 across --steps 1 → 5 → 10 → 20, then overfits at --steps 50.
  • Session 8 Phase A/A5/A6 refines that line: --steps 20 is a shoulder, not the peak. The one-step map across the 28→50 region is a trimodal sawtooth with local maxima at --steps 33 (best_uplift = 1.873756), 41 (1.708374), and 49 (1.819397), sharp drops at 33→34 and 41→43, and a final 49→50 cliff to 0.000000.
  • At the μ baseline config --steps 20 / SriYantraAdj_v1.bin / RegionTags_v1.bin, exp_k2_scars is invariant across xnor_train, xnor_mini, and xnor_test in the mirrored Phase A logs.
  • Both adjacency topology and region-tag partition affect exp_k2_scars: the Phase A cross-graph table gives SY+v1 = 1.324074, SY+v2 = 0.806000, RA+v1 = 1.341583, and RA+v2 = 0.000000 at --steps 20.
  • exp_k3_truth_sensor does real internal work. Session 7 showed the exposed --sensor-strength and --sensor-threshold flags do not parameterize its KPI triple on the tested surface. Session 8 B3 shows --steps does change baseline and gap values (75.0% reduction at steps=1, 79.4% at steps=20), while selected sensor_error stays near 22.3.
  • resonance_r3 does real work inside its task loop, but the exposed --steps flag does not parameterize the emitted output.
  • Session 8 B3 identifies three --steps-decorative task surfaces at steps 1 vs 20: resonance_r3, resonance_v2, and exp_i2. This is scoped to those step values and default inputs.

Phase G v2 chain — closed cells through G.2

  • Path-independent at fixed config. The Phase G cell G.6 path_dependence reports 4/4 same-config canonical SHAs match across path A and B. DM3 exposes no observable cross-invocation state on the tested surface. This closes claim χ.
  • σ″ trimodal sawtooth is geometry-independent in shape. Phase G G.2 trimodal_portability reports verdict = PASS, summary = "trimodal shape preserved in 3/3 configs". All three preserved cross-controls (cfg-A RandomAdj_v1, cfg-B RegionTags_v2, cfg-C xnor_mini) reproduce the σ″ peaks at s33 / s41 / s49 and the cliff at s50 = 0.000000. σ″ promotes to CONFIRMED for shape; magnitudes are config-dependent and explicitly not promoted as portable.
  • Cliff at s50 is universal. All three G.2 configs return p50 = 0.000000 exactly. The cliff is the strongest universal feature on the tested surface; its fine structure is G.7's next deeper test (in flight at this update).
  • The s50 zero is a clipped best_uplift, not absence of per-condition learning dynamics. The 2026-04-30 08:30 UTC G.7 in-flight logs show cfg-A complete at s50 and cfg-B partial at s50: both return best_uplift = 0.000000 by max-over-conditions clipping while lesson=3 per-condition uplifts go negative. cfg-A reproduces -0.201691 at noise=0.100 and -0.219131 at noise=0.200 across 3/3 receipts; the first completed cfg-B s50 receipt gives -1.803669 and -1.860878. Direction is preserved across these two configs; magnitude remains config-bound. This is a scope note on the σ″ cliff, not an independent G.7 claim.
  • Cycle past the cliff. Phase G G.1 cycle_probe reports s56 best_uplift = 1.970840 on the σ″ baseline, higher than the σ″ primary peak s33 = 1.873756. The 49 → 50 cliff is therefore not a permanent collapse. This is claim ψ (CANDIDATE).
  • Cycle-7 mechanism (PARTIAL). G.1.5 cycle_disambiguator reports the multiplicity-7 sawtooth fits tighter than multiplicity-6 or -8 on the σ″ baseline. Recorded as suggestive; not promoted.
  • σ″ baseline determinism extension. G.0.5 determinism_recheck reports 10/10 bit-exact best_uplift reproductions at both s30 and s33. Folded into the ξ evidence base.

Reconstruction (static, Tier-2)

The independent dm3-runner-reconstruction-2026-04-27/ lane closes seven of eight pre-registered reconstruction hypotheses at static Tier-2 evidence level. The eighth, R8, requires Tier-3 Android execution tracing and remains the sovereign open gate. See RECONSTRUCTION_TIER2_NOTE.md for the full evidence table and CLAIMS.md "Reconstruction (static, Tier-2)" for the formal claim block. The relevant IS statements for this surface are:

  • The loaded SriYantraAdj_v1.bin fixture graph is exactly P_95 ☐ K_4 at parsed-fixture Tier-2 (380 vertices, 946 edges, degree histogram {4: 8, 5: 372}).
  • The default internal dm3_runner skeleton (build_helix_meru + build_dual_meru) is exactly P_95 ☐ K_3 at static-disassembly Tier-2 (285 vertices, 567 edges, degree histogram {3: 6, 4: 279}).
  • The two surfaces share the path base P_95 = 2 · (12 · 4) − 1 and are related by complete-graph fiber promotion K_3 → K_4. The loaded fixture's first Betti number is 946 − 380 + 1 = 567, equal to the internal P_95 ☐ K_3 edge count. This is an exact arithmetic identity, not a runtime coupling proof.
  • RegionTags_v2.json is exactly the degree-4-root BFS shelling of the loaded fixture, with RING_k = min(distance // 5, 8) + 1. JSON mismatch count is zero against the host-side reconstruction.
  • The Laplacian spectrum of the loaded fixture is the closed-form λ_k(P_95) + μ(K_4) with μ(K_4) ∈ {0, 4, 4, 4}. Fiedler value 0.001093485318147902, λ_max 7.9989065146818525.
  • The full automorphism group of the loaded fixture graph is Aut(P_95 ☐ K_4) = C_2 × S_4, of order 48. The previously claimed C_3 symmetry remains a sub-action of the natural S_4 action on the four K_4 fibers; it is preserved, but it is not the full automorphism group.
  • Static disassembly identifies a file-loaded L-branch surface in generate_tags_v2 and run_spectral_analysis, distinct from the internal Dm3State::initialize / build_dual_meru default path. No static call from the L-branch callees to Dm3State, build_dual_meru, or flatten_dual_vertices was found. Whether any live task-run actually consumes RegionTags_v2.bin (legacy TAG2) versus only the JSON / current TAG3 writer path is the open R8 question.

Reproducibility

  • exp_r1_r4_campaign outputs are deterministic up to run_sec, with multiple canonical SHA equivalence classes catalogued across Sessions 6 and 7.
  • exp_k2_scars promoted KPI outputs are deterministic at fixed configurations across the mirrored Session 8 Phase A equivalence classes.
  • A5/A6 add cross-cell exact matches for exp_k2_scars at --steps 30, 35, 40, and 45.
  • τ confirms ARM64 cross-platform determinism for exp_k2_scars on the tested baseline values: RM10 native Android and Apple M1 Android ARM64 emulator emit bit-exact best_uplift and max_scar_weight at --steps {20,30,40,45,50}.
  • The harmonic task is stochastic in selection but stable in distribution at the tested granularity.

DM3 IS NOT

Computational claims it is NOT

  • Not an AI system. No external agency, no general intelligence, no task-neutral cognition.
  • Not a transformer-created bistability. H2 was killed; the graph geometry is sovereign.
  • Not a tunable resonance computer. --freq remains a null line on the tested harmonic surface.
  • Not a C3-asymmetric coupling. Claim γ remains retracted.
  • Not a deterministic basin selector. Basin selection is IID at the promoted default surface.
  • Not a system with multiple independently confirmed dynamical control parameters. Confirmed exposed axes are narrow: --asymmetry for basin position, --adj for R1, --tags for R2, and --steps for exp_k2_scars only.

Scope claims it is NOT

  • Not a single E-continuum across tasks. Harmonic and holography are distinct regimes.
  • Not a universal basin classifier. The Session 4 locked classifier is valid only on asymmetry within [−2, +2].
  • Not a system whose harmonic basin selection shows a receipted coupling to the tested thermal or power-path interventions.
  • Not a system where requesting --steps > 20 on the SY-default gate surface increases operational_steps above 10.
  • Not a system where --sensor-strength or --sensor-threshold parameterize exp_k3_truth_sensor.
  • Not a system where --steps parameterizes resonance_r3.
  • Not a system where --steps 1 versus --steps 20 parameterizes resonance_v2 or exp_i2 on the tested default surface.
  • Not a system where every callable task responds to --steps.
  • Not a system where parallelizing dm3_runner yields throughput gain. That line was tested and discarded.

Interpretive claims it is NOT

  • Not a cognitive, spiritual, or mystical architecture. Sri Yantra, Bhupura, Lotus, Bindu, Om, Aum, and Meru are source vocabulary, not evidence.
  • Not a broad “DM3 learns” claim. The promoted positive learning result is narrow and task-specific to exp_k2_scars in the tested step window.
  • Not a system where RandomAdj_v1.bin + RegionTags_v2.bin + --steps 20 learns on exp_k2_scars. The mirrored Phase A logs show best_uplift = 0.000000 in all three repeats. This is scoped to --steps 20, not a ceiling claim for all RA+v2 settings.
  • Not a system whose peak exp_k2_scars uplift is at --steps 20, 30, or 40 on the mapped baseline surface. A5/A6 move the candidate curve shape to a trimodal sawtooth with local maxima at 33, 41, and 49.
  • Not a bimodal exp_k2_scars learning curve on the mapped baseline 28→50 surface; old σ and σ′ wordings are rejected-before-promoted.
  • Not a system whose σ″ magnitudes are configuration-invariant. Phase G G.2 shows the trimodal sawtooth shape is preserved across all three cross-controls, but cfg-B (RegionTags_v2) returns substantially weaker peaks (p33 = 0.995 vs SY baseline p33 = 1.874) and zeros at p34 and p43. The σ″ shape is geometry-independent; the σ″ values are not.
  • Not a system where fixed-config exp_k2_scars KPI output is stochastic on the mirrored Phase A equivalence classes.
  • Not a generally tunable CLI capability surface. Session 7 adds narrow sensor-flag decorativeness for exp_k3_truth_sensor and --steps decorativeness for resonance_r3; Session 8 B3 narrows that by showing exp_k3_truth_sensor is --steps-responsive in absolute values and adds scoped --steps-decorative candidates for resonance_v2 and exp_i2.
  • Not a folder for commercial framing. repo_stage/ remains the scientific and receipted language lane.

Reconstruction-side IS_NOT (static, Tier-2)

  • Not a complete reconstruction. R1–R7 are static Tier-2 evidence: parsed fixtures, host-recomputed invariants, and static-disassembly call-graph identity. R8 (Android argv / file-open / output trace under live execution) remains OPEN_TIER3_BLOCKED. Until R8 closes, the reconstruction is not authority-complete.
  • Not "the binary's runtime graph is P_95 ☐ K_4." The parsed loaded fixture is P_95 ☐ K_4. Whether the live task actually consumes that fixture, generates an isomorphic graph, or combines fixture loading with runtime mutation is the R8 question. Static evidence favours independent file-load for the L-branch tag/spectral callees, but Tier-3 trace is still required.
  • Not a "Sri Yantra mystical geometry" claim at graph level. The graph is the Cartesian product of a 95-vertex path and a 4-clique. The geometric/Sri-Yantra vocabulary is source-history, preserved at the construction layer (claim α) and not re-promoted at the reconstruction layer.
  • Not "transformer-static and graph-static both reconstructed." The 192-wide MicroTransformer surface is identified statically as a separate component but is not the subject of the R1–R7 graph reconstruction. Its interaction with the relaxation dynamics is not reconstructed at this surface.
  • Not "TAG2 is a live consumed file." The legacy RegionTags_v2.bin is TAG2/version 2; the current binary writer string is TAG3/version 3. Whether any live task consumes the legacy TAG2 file is the open R8 question; static evidence says only that the current writer is TAG3.

Session 7 closeout notes

  • The Session 7 closeout claim set is now θ, ι, κ, λ, μ, ν, with δ.3 visibly weakened.
  • The strongest public headline is the combination of: first receipted positive learning (μ), dynamics-layer substrate null (ι), and closed 2×2 gate cross-control (λ).
  • The gate-layer substrate-null closeout is carried with an explicit repo seam: the engineer handover names standalone S2_pinned / S4_airplane / S6_core directories that are not exposed verbatim in this checkout. That mismatch stays visible in REPO_AGENT_FINDINGS.md.

Session 8 Phase A / A5-B3-A6 close note

  • Phase A is closed under a pure-scientific-learning frame in docs/restart/DM3_SESSION8_PHASE_A_FINAL_REPORT_20260424.md.
  • The local mirror includes the A.1/A.2/A.3/A.4 receipt trees and device_snapshot/bin/run_cell.sh.
  • The follow-on final report docs/restart/DM3_SESSION8_PHASE_A5_B3_A6_FINAL_REPORT_20260425.md closes 57 additional A5/B3/A6 local per-run receipts.
  • The promoted Phase A/A5/B3/A6 lines are ξ fixed-config determinism, ο sharp overfit cliff localized to 49→50, and τ ARM64 cross-platform determinism. Candidate lines are π scoped dataset invariance, ρ RA+v2+steps=20 zero learning, σ″ trimodal sawtooth curve shape, and φ three scoped --steps-decorative tasks.
  • Old σ and σ′ curve-shape wordings are rejected-before-promoted and remain visible as process wins, not active claims.
  • Count seam: the final report and summaries use 55 total-run accounting, while the local mirror has 52 per-run receipt/log pairs. The promoted numeric statements use the 52 local per-run logs and the matching summary hashes; the later A5/B3/A6 chain adds 57 directly mirrored per-run receipts. See REPO_AGENT_FINDINGS.md.
  • The final report documents an airplane-mode deviation: 40 reported Phase A receipts were captured with airplane mode OFF. No receipt is quarantined, and the deviation remains visible.
  • Engineering-side admissible finding: the old all-sensor thermal gate on RM10 is not a trustworthy execution guard because the PMIC sensor can stay hot long after CPU zones cool. The mirrored harness snapshot now contains the CPU/GPU-only patched run_cell.sh and the pre-patch backup.

Phase G v2 chain note (G.2 closed; later cells unpromoted; updated 2026-05-02 18:38 UTC)

  • Phase G v2 launched 2026-04-25 18:25:54 UTC. Master log on device at /data/local/tmp/dm3_harness/phase_g_chain.log. PRD at docs/restart/DM3_PHASE_G_AUGMENTED_PRD_v2_REORDERED_20260425.md.
  • Closed cells with promoted ledger entries:
    • G.0 PASS (pre-launch gates)
    • G.0.5 PASS (folded into ξ; 10/10 bit-exact at s30/s33)
    • G.6 PASS (claim χ, path-independence)
    • G.1 PASS (claim ψ, s56 = 1.970840 cycle past cliff, CANDIDATE)
    • G.1.5 PARTIAL (multiplicity-7 mechanism candidate, not promoted)
    • G.2 PASS (σ″ trimodal sawtooth shape geometry-independent across cfg-A / cfg-B / cfg-C cross-controls; promotes σ″ to CONFIRMED for shape; magnitudes config-dependent, not portable)
  • The latest GitHub-resident G.7 pull remains the 2026-04-30 partial snapshot: 22 immutable in-flight G.7 receipts, cfg-A complete at 15/15, cfg-B partial at 7/15, cfg-C pending at 0/15, and no outcome.json.
  • The 2026-05-02 audit note reports that G.7 later closed on-device and G.3 is in flight. Those statuses are evidence handles only here: no G.7 verdict, G.3 class assignment, G.4, G.5, or G.5+ outcome is promoted without the chain-close handover and host-side receipt pull.
  • Closed-cell evidence for G.0.5, G.1, and G.1.5 remains mirrored under artifacts/phase_S8_PG_followup_20260429T023308Z/. The 13:01 UTC partial pull under artifacts/phase_S8_PG_followup_20260429T130215Z/ supersedes G.2 and G.6 evidence paths. The 08:30 UTC partial pull under artifacts/phase_S8_PG_followup_20260430T082723Z/ is the latest G.7 GitHub-resident snapshot. The May 2 audit note is mirrored at docs/restart/REPO_AGENT_NOTE_PHASE_G_G7_CLOSE_AND_G3_IN_FLIGHT_20260502T1838.md. The next ledger update for Phase G lands when the chain closes and the remaining cell receipts are pulled to host.

Reconstruction (static, Tier-2) note (2026-04-28)

  • Independent backwards-reconstruction lane: dm3-runner-reconstruction-2026-04-27/.
  • Eight pre-registered reconstruction hypotheses R1..R8 at phase_05_falsification_harness/H1_H8_STATIC_PREFLIGHT.md.
  • R1–R7: PASS_STATIC_TIER2 (or PASS_STATIC_TIER2_DYNAMIC_OPEN for R6).
  • R8: OPEN_TIER3_BLOCKED. Tier-3 closure requires Android argv/openat/write trace under live dm3_runner execution and is parked until the live Phase G v2 chain closes.
  • See RECONSTRUCTION_TIER2_NOTE.md for the full identity card, closed-form spectrum, RegionTags shelling formula, and runtime surface flow.
  • Until R8 closes, no claim of complete reconstruction is promoted on this surface.

How this ledger is maintained

Each session appends or revises lines with an explicit justification and receipted basis. Removals require a written retraction or weakening that remains visible in CLAIMS.md.