Skip to content

Latest commit

 

History

History
320 lines (237 loc) · 13.7 KB

File metadata and controls

320 lines (237 loc) · 13.7 KB

Repo Agent Findings

Last updated: 2026-05-02

This file records seams that matter to a hostile reviewer and should stay visible until cleared.

F1 — Harmonic helper summaries are generic, not the scientific verdict

Paths:

  • artifacts/phase_S7_P0_receipt_harness_20260418T151800Z/S2H_final/S2H_stat_summary.json
  • artifacts/phase_S7_P0_receipt_harness_20260418T151800Z/S7_thermal_final/S7_thermal_summary.json
  • artifacts/phase_S7_P0_receipt_harness_20260418T151800Z/S8_final/S8_battery_summary.json
  • artifacts/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.

F2 — Session 7 gate-layer directory names in the docs do not match the local checkout

Paths:

  • docs/restart/DM3_STATUS_SNAPSHOT_20260422.md
  • repo_stage/HANDOVER_TO_REPO_AGENT_20260422.md
  • artifacts/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.

F3 — exp_k3_truth_sensor writes its scientific content to .log, not .bin

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.

F4 — resonance_v2 task-side investigation narrowed by B3

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.

F5 — Session 8 Phase A total-run count differs from local per-run anchors

Paths:

  • docs/restart/DM3_SESSION8_PHASE_A_FINAL_REPORT_20260424.md
  • repo_stage/HANDOVER_TO_REPO_AGENT_PHASE_A_20260424.md
  • artifacts/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.

F6 — Session 8 thermal-gate patch is mirrored with pre-patch backup

Paths:

  • docs/restart/DM3_SESSION8_PHASE_A_FINAL_REPORT_20260424.md
  • artifacts/phase_S8_P0_learning_20260422T213500Z/device_snapshot/bin/run_cell.sh
  • artifacts/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.

F7 — Repo-agent Phase A promotion decisions

Paths:

  • repo_stage/HANDOVER_TO_REPO_AGENT_PHASE_A_20260424.md
  • docs/restart/DM3_SESSION8_PHASE_A_FINAL_REPORT_20260424.md
  • repo_stage/CLAIMS.md
  • repo_stage/IS_AND_IS_NOT.md

Finding:

The engineer-agent requested scrutiny on three Phase A publication-surface questions. Current repo-agent decision:

  • Claim σ was recorded as CANDIDATE and coarse, not CONFIRMED, during the Phase A integration. F8 supersedes this: A5/A6 killed σ and σ′ before promotion and replaced them with candidate σ″.
  • Claim ρ is recorded as CANDIDATE and scoped to RA+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.

F8 — A5/B3/A6 supersedes the old σ curve wording before promotion

Paths:

  • docs/restart/DM3_SESSION8_PHASE_A5_B3_A6_FINAL_REPORT_20260425.md
  • repo_stage/HANDOVER_TO_REPO_AGENT_A5_B3_A6_FINAL_20260425.md
  • artifacts/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.

F9 — B3 summary FAIL is a generic-script limitation, not a science fail

Paths:

  • artifacts/phase_S8_P0_learning_20260422T213500Z/cells/B3_cli_audit/B3_cli_audit_summary.json
  • docs/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.

F10 — κ handover wording needed numeric tightening

Paths:

  • repo_stage/HANDOVER_TO_REPO_AGENT_A5_B3_A6_FINAL_20260425.md
  • docs/restart/DM3_SESSION8_PHASE_A5_B3_A6_FINAL_REPORT_20260425.md
  • artifacts/phase_S8_P0_learning_20260422T213500Z/cells/B3_cli_audit/B3_cli_audit_exp_k3_truth_sensor_steps1.log
  • artifacts/phase_S8_P0_learning_20260422T213500Z/cells/B3_cli_audit/B3_cli_audit_exp_k3_truth_sensor_steps20.log
  • repo_stage/CLAIMS.md
  • repo_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.

F11 — τ is promoted from a separate cross-platform lane

Paths:

  • repo_stage/CLAIM_TAU_CONFIRMED_20260424.md
  • repo_stage/HANDOVER_TO_REPO_AGENT_ORCHESTRATOR_M1_TAU_20260424.md
  • docs/restart/DM3_AGDH1_M1_SCIENCE_TEAM_NOTE_20260424.md
  • artifacts/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.

F12 — 2026-04-25 playbook alignment keeps the DM3 license exception

Paths:

  • README.md
  • REPRODUCIBILITY.md
  • .zenodo.json
  • scripts/compliance_audit.sh
  • pyproject.toml
  • validation/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.

F13 — First Phase G partial host pull is evidence, not chain close

Paths:

  • artifacts/phase_S8_PG_followup_20260429T023308Z/
  • artifacts/phase_S8_PG_followup_20260429T023308Z/phase_g_chain.log
  • artifacts/phase_S8_PG_followup_20260429T023308Z/cells/G2_trimodal_portability/outcome.json
  • artifacts/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.

F14 — Second Phase G partial pull sharpens s50 without promoting G.7

Paths:

  • artifacts/phase_S8_PG_followup_20260429T130215Z/
  • artifacts/phase_S8_PG_followup_20260429T130215Z/PULLED_AT.txt
  • artifacts/phase_S8_PG_followup_20260429T130215Z/cells/G2_trimodal_portability/outcome.json
  • artifacts/phase_S8_PG_followup_20260429T130215Z/cells/G6_path_dependence/outcome.json
  • artifacts/phase_S8_PG_followup_20260429T130215Z/cells/G7_cliff_class_characterization/G7_cliff_class_characterization_ca_s50_r1.log
  • artifacts/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.

F15 — Third Phase G partial pull is G.7 progress, not G.7 closure

Paths:

  • artifacts/phase_S8_PG_followup_20260430T082723Z/
  • artifacts/phase_S8_PG_followup_20260430T082723Z/PULLED_AT.txt
  • artifacts/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}.log
  • artifacts/phase_S8_PG_followup_20260430T082723Z/cells/G7_cliff_class_characterization/G7_cliff_class_characterization_cb_s50_r1.log
  • artifacts/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.

F16 — May 2 G.7 close note is an audit boundary, not a promotion

Paths:

  • docs/restart/REPO_AGENT_NOTE_PHASE_G_G7_CLOSE_AND_G3_IN_FLIGHT_20260502T1838.md
  • artifacts/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.