Last updated: 2026-05-02
This file records seams that matter to a hostile reviewer and should stay visible until cleared.
Paths:
artifacts/phase_S7_P0_receipt_harness_20260418T151800Z/S2H_final/S2H_stat_summary.jsonartifacts/phase_S7_P0_receipt_harness_20260418T151800Z/S7_thermal_final/S7_thermal_summary.jsonartifacts/phase_S7_P0_receipt_harness_20260418T151800Z/S8_final/S8_battery_summary.jsonartifacts/phase_S7_P0_receipt_harness_20260418T151800Z/S5_final/S5_basin_volume_summary.json
Finding:
These helper summaries are produced by a deterministic-cell
summarize_cell.sh path and therefore mark harmonic cells as FAIL
when they see more than one canonical SHA. That is appropriate for a
deterministic gate surface and wrong for stochastic harmonic receipts.
The promoted Session 7 dynamics claim therefore uses Wilson-CI overlap
on the underlying .bin receipts, not the helper verdict strings.
Paths:
docs/restart/DM3_STATUS_SNAPSHOT_20260422.mdrepo_stage/HANDOVER_TO_REPO_AGENT_20260422.mdartifacts/phase_S7_P0_receipt_harness_20260418T151800Z/
Finding:
The status snapshot and handover name standalone gate-layer directories
S2_pinned / S4_airplane / S6_core. The local checkout does not expose
those directory names verbatim under the Session 7 harness tree. It
does expose the smoke receipt, the later S2H_* / S7_* / S8_* / S5_* / S10_* / S11_* / T* trees, and the final Session 7 report that closes
the gate-layer substrate-null line. The repo therefore keeps the
gate-layer closeout claim and this naming mismatch visible together.
Paths:
artifacts/phase_S7_P0_receipt_harness_20260418T151800Z/S10_final/
Finding:
The S10 .bin files are empty by design. The task writes KPI content to
stdout, so the scientific payload lives in the nine .log files plus
the summary JSON. The empty-file canonical SHA
e3b0c44298fc... is not evidence loss here.
Path:
artifacts/phase_S7_P0_receipt_harness_20260418T151800Z/T1_cartography/artifacts/phase_S8_P0_learning_20260422T213500Z/cells/B3_cli_audit/
Finding:
The older Session 7 cartography note found resonance_v2 unhelpful on
that tested surface and treated it as a task-side investigation item.
B3 narrows the present repo-facing statement: at --steps 1 versus
20, resonance_v2 is --steps-decorative on primary output. This
does not prove every resonance_v2 entrypoint is null; it only records
the tested CLI axis.
Paths:
docs/restart/DM3_SESSION8_PHASE_A_FINAL_REPORT_20260424.mdrepo_stage/HANDOVER_TO_REPO_AGENT_PHASE_A_20260424.mdartifacts/phase_S8_P0_learning_20260422T213500Z/
Finding:
The Session 8 Phase A final report says Phase A closed with 55 receipts
across A.1/A.2/A.3/A.4. The local mirror now contains the A-cell trees,
but the directly countable per-run receipt/log pairs are 52:
A.1=10, A.2=21, A.3=12, A.4=9. The summary files report
total_runs = 22/13/10 for A.2/A.3/A.4 while also listing
unique_receipt_sha256 = 21/12/9. The summary SHAs match the
handover. The repo promotes only statements supported by the local
per-run logs and keeps this accounting seam visible.
Paths:
docs/restart/DM3_SESSION8_PHASE_A_FINAL_REPORT_20260424.mdartifacts/phase_S8_P0_learning_20260422T213500Z/device_snapshot/bin/run_cell.shartifacts/phase_S8_P0_learning_20260422T213500Z/device_snapshot/bin/run_cell.sh.pre_cpugate_bak
Finding:
The final report records a real engineering-side finding: the old
all-sensor thermal gate on RM10 was being held open by the PMIC sensor
long after CPU zones cooled, so the live harness rule was patched to
CPU/GPU-only thermal gating. The local Phase A mirror now includes both
the patched run_cell.sh and run_cell.sh.pre_cpugate_bak. Any mixed
old-gate / new-gate split inside A.2 remains visible in the progress
logs and final report.
Paths:
repo_stage/HANDOVER_TO_REPO_AGENT_PHASE_A_20260424.mddocs/restart/DM3_SESSION8_PHASE_A_FINAL_REPORT_20260424.mdrepo_stage/CLAIMS.mdrepo_stage/IS_AND_IS_NOT.md
Finding:
The engineer-agent requested scrutiny on three Phase A publication-surface questions. Current repo-agent decision:
- Claim
σwas recorded as CANDIDATE and coarse, not CONFIRMED, during the Phase A integration. F8 supersedes this: A5/A6 killedσandσ′before promotion and replaced them with candidateσ″. - Claim
ρis recorded as CANDIDATE and scoped toRA+v2+steps=20, not as a broad RA+v2 capability ceiling. - The airplane-mode deviation is surfaced in the ledger and findings. It is not sanitized; the affected receipt JSONs are mirrored locally.
Paths:
docs/restart/DM3_SESSION8_PHASE_A5_B3_A6_FINAL_REPORT_20260425.mdrepo_stage/HANDOVER_TO_REPO_AGENT_A5_B3_A6_FINAL_20260425.mdartifacts/phase_S8_P0_learning_20260422T213500Z/cells/A5_peak_finder/artifacts/phase_S8_P0_learning_20260422T213500Z/cells/A6a_peak_fill/artifacts/phase_S8_P0_learning_20260422T213500Z/cells/A6b_post_peak/repo_stage/CLAIMS.md
Finding:
The earlier Phase A σ wording described a coarse 30/40 peak and the
A5 interim σ′ wording described a bimodal curve. The A6 fills kill
both before promotion. The active candidate is now σ″: a trimodal
sawtooth on the scoped baseline exp_k2_scars surface with local
maxima at s33, s41, and s49. The rejected-before-promoted history
stays visible because it shows the kill criterion firing rather than a
smooth story being retrofitted.
Paths:
artifacts/phase_S8_P0_learning_20260422T213500Z/cells/B3_cli_audit/B3_cli_audit_summary.jsondocs/restart/DM3_SESSION8_PHASE_A5_B3_A6_FINAL_REPORT_20260425.md
Finding:
The B3 cell intentionally mixes all twelve callable tasks at
--steps 1 and --steps 20. Its summary JSON reports verdict=FAIL
because the generic deterministic-cell summarizer sees seven canonical
output classes across heterogeneous tasks. The scientific finding comes
from per-task comparisons: nine tasks respond to --steps, three are
decorative at 1 versus 20, and exp_k3_truth_sensor is partial.
Paths:
repo_stage/HANDOVER_TO_REPO_AGENT_A5_B3_A6_FINAL_20260425.mddocs/restart/DM3_SESSION8_PHASE_A5_B3_A6_FINAL_REPORT_20260425.mdartifacts/phase_S8_P0_learning_20260422T213500Z/cells/B3_cli_audit/B3_cli_audit_exp_k3_truth_sensor_steps1.logartifacts/phase_S8_P0_learning_20260422T213500Z/cells/B3_cli_audit/B3_cli_audit_exp_k3_truth_sensor_steps20.logrepo_stage/CLAIMS.mdrepo_stage/IS_AND_IS_NOT.md
Finding:
The handover asks to refine κ without weakening or retracting it, and
also says the error-reduction ratio is fixed near 79.4% across
--steps. The local B3 logs show a more precise result: at
steps=1, baseline_error=89.255325, sensor_error=22.317333, and
the reduction ratio is about 75.0%; at steps=20,
baseline_error=108.164818, sensor_error=22.292860, and the ratio is
about 79.4%. The public repo surface therefore preserves Session 7
sensor-strength / threshold invariance, but does not promote a
universal --steps-invariant ratio story.
Paths:
repo_stage/CLAIM_TAU_CONFIRMED_20260424.mdrepo_stage/HANDOVER_TO_REPO_AGENT_ORCHESTRATOR_M1_TAU_20260424.mddocs/restart/DM3_AGDH1_M1_SCIENCE_TEAM_NOTE_20260424.mdartifacts/phase_S8_AGDH1_external_M1_20260424/
Finding:
τ is independent of the A5/B3/A6 phone chain. It confirms bit-exact
exp_k2_scars KPI matches between RM10 native Android and an Apple M1
Android ARM64 emulator at --steps {20,30,40,45,50}. The public
surface promotes τ as cross-platform ARM64 determinism only for the
tested binary and inputs; it does not claim x86/source-rebuild parity.
Paths:
README.mdREPRODUCIBILITY.md.zenodo.jsonscripts/compliance_audit.shpyproject.tomlvalidation/results/repo_surface_preflight.json
Finding:
The 2026-04-25 repo playbook and live-project ethos add a parser-facing
## Upcoming Workstreams section, stronger front-door category /
blocker prose, mandatory compliance audit, and packaging upper-bound
discipline. DM3 adopts those repo-shape rules while keeping the explicit
ethos exception that DM3 is separately licensed under
LicenseRef-Zer0pa-DM3-RRL-1.0, not SAL v7.0. The public surface now
keeps the strongest τ metric, source-provenance blocker, missing
dm3-verify wrapper, and Mode A non-claim in the front-door posture
instead of burying them in downstream notes.
Paths:
artifacts/phase_S8_PG_followup_20260429T023308Z/artifacts/phase_S8_PG_followup_20260429T023308Z/phase_g_chain.logartifacts/phase_S8_PG_followup_20260429T023308Z/cells/G2_trimodal_portability/outcome.jsonartifacts/phase_S8_PG_followup_20260429T023308Z/cells/G7_cliff_class_characterization/
Finding:
The 2026-04-29 02:35 UTC partial pull mirrors closed-cell receipts and
outcome.json files for G.0.5, G.1, G.1.5, G.2, and G.6,
plus six immutable in-flight G.7 receipts. It is not a Phase G chain
close. The raw phase_g_chain.log includes an earlier pre-resume
G.2 verdict=KILL line before the later resumed G.2 verdict=PASS;
the promoted σ″ shape claim is anchored to the final
cells/G2_trimodal_portability/outcome.json PASS packet and keeps the
raw chain log visible. G.7 has no promoted outcome here, and
MANIFEST.tsv remains deferred to chain close.
Paths:
artifacts/phase_S8_PG_followup_20260429T130215Z/artifacts/phase_S8_PG_followup_20260429T130215Z/PULLED_AT.txtartifacts/phase_S8_PG_followup_20260429T130215Z/cells/G2_trimodal_portability/outcome.jsonartifacts/phase_S8_PG_followup_20260429T130215Z/cells/G6_path_dependence/outcome.jsonartifacts/phase_S8_PG_followup_20260429T130215Z/cells/G7_cliff_class_characterization/G7_cliff_class_characterization_ca_s50_r1.logartifacts/phase_S8_PG_followup_20260429T130215Z/cells/G7_cliff_class_characterization/G7_cliff_class_characterization_ca_s50_r2.log
Finding:
The 2026-04-29 13:01 UTC pull supersedes the public G.2/G.6 evidence
paths with an identical later immutable tree and adds eight in-flight
G.7 receipts. It is still not a chain close. The G.7 cfg-A s50
replicates show that best_uplift = 0.000000 is a clipped
max-over-conditions summary: lesson=3 per-condition uplifts are
negative and bit-identical across ca_s50_r1 and ca_s50_r2
(-0.201691 at noise=0.100, -0.219131 at noise=0.200). The
public surface records this as a scope note on σ″ and as supporting
evidence for ξ fixed-config determinism, not as an independent G.7
claim. MANIFEST.tsv remains deferred to chain close.
Paths:
artifacts/phase_S8_PG_followup_20260430T082723Z/artifacts/phase_S8_PG_followup_20260430T082723Z/PULLED_AT.txtartifacts/phase_S8_PG_followup_20260430T082723Z/cells/G7_cliff_class_characterization/artifacts/phase_S8_PG_followup_20260430T082723Z/cells/G7_cliff_class_characterization/G7_cliff_class_characterization_ca_s50_r{1,2,3}.logartifacts/phase_S8_PG_followup_20260430T082723Z/cells/G7_cliff_class_characterization/G7_cliff_class_characterization_cb_s50_r1.logartifacts/phase_S8_PG_followup_20260430T082723Z/cells/G7_cliff_class_characterization/G7_cliff_class_characterization_cb_s50_r2.log
Finding:
The 2026-04-30 08:30 UTC pull is a G.7-only progress snapshot: 22
completed receipts, cfg-A complete at 15/15, cfg-B partial at 7/15,
cfg-C pending at 0/15, and no G.7 outcome.json. It is not a Phase G
chain close. The new science observation is limited to cliff fine
structure: cfg-A s50 negative uplift is now exact across 3/3
receipts, and the first completed cfg-B s50 receipt shows the same
negative direction with larger magnitude (-1.803669, -1.860878).
The cb_s50_r2.log file was pulled while in flight and does not carry
a complete receipt set, so cfg-B replicate determinism is not promoted
from this snapshot. Battery-health and process-health details from the
repo-agent note remain operational context, not public claim language.
MANIFEST.tsv remains deferred to chain close.
Paths:
docs/restart/REPO_AGENT_NOTE_PHASE_G_G7_CLOSE_AND_G3_IN_FLIGHT_20260502T1838.mdartifacts/phase_S8_PG_followup_20260430T082723Z/
Finding:
The May 2 engineer note records later on-device state beyond the latest GitHub-resident receipt pull: G.7 closed on-device with a dual verdict shape (harness summary PASS, cell outcome KILL), and G.3 is in flight. Those details are not promoted on this public surface because the chain-close handover and host-side receipt pull have not landed here. The repo therefore keeps the note as audit trail, keeps the README on the System Mechanics front-door shape, and keeps G.7/G.3/G.4/G.5/G.5+ out of claims, manifests, and public proof anchors until the real chain close boundary is available.