m625 is not a clinical decision tool · not a publisher · not an expert authority · not a simulator. It is a cross-domain Roll-formalized evidence atlas with public audit + attack surface.
A researcher working on rare conditions (n=1 medicine, ultra-rare neuropsychiatric, cross-condition synthesis) reads about a cross-disease X-invariant pattern (different conditions, same brain-network state-structure). Pain points (see 05_docs/M625_PERSONAS.md):
formalization language across labs is inconsistent (each lab redefines its X, Φ, Δ);
self-audit standards for own claims are missing or ad-hoc;
cross-paper reproducibility is poor;
negative results often unpublished — attack surface invisible.
m625 does not solve any of these. **m625 shows a Roll-formalized cross-condition V_functional object and invites the researcher to break it**.
This is not a methodological recommendation. This is not a publication. (See "What this demo will not say".)
The Roll-formalized object (anchored on m517 CARD08)
rare / cross-condition (multi-neuropsychiatric, links to m610 rare-disease mapping)
Roll TMU type
TMU-B · Δ_B set varies across conditions while X-structure remains similar
Cross-period bit pattern
K4.1 = (0, 0, 1) (X-structure invariant across conditions; Φ similar; only Δ-B varies) — per m604 Framework §4.5
Source layer
Wu et al. 2025 (training_data_recall · TODO_unverified · m517 marks anchor as drafted-not-signed)
m517 status
v0.3.0 Wave 2 closed · BIRTH snapshot registered to m556
Cross-domain link
m610 cross-condition mapping (rare-disease pattern) — per 01_materials_cache/A4_m610_cross_domain_mapping.md
m625 does not modify m517's card file. The reference path is /data/projects/m517_NeuroDynamics_Circle/workspace/02_cards_cooked/CARD08_DEEP_v0.2.md (m517 sovereign).
How to read this evidence (researcher orientation, ≤ 5 minutes)
V_functional cross-condition: across different neuropsychiatric conditions (the cooked card cites multi-condition population data via the upstream Wu 2025 source), brain-network functional state structure (X) appears similar while the intervention boundary Δ_B differs by condition.
TMU-B classification: this is m610's TMU-B regime: Φ similar across conditions, X-structure similar across conditions, only Δ_B (intervention specification) varies. This is not a cytostatic TMU-A regime and not a phi_direct_rewrite TMU-C regime.
K4.1 = (0, 0, 1) declares: X-structure not visibly altered (b_X = 0), Φ not visibly altered (b_Φ = 0), Δ-B intervention boundary visibly altered (b_Δ_B = 1). This is the bit-pattern of a cross-condition synthesis claim.
Reproducibility witness: re-running python3 -m 02_audit.audit --self-audit on this artifact must yield grade A and 0 vetoes (this is the v0.1 audit chain self-test).
What this demo will not say (anti-drift · per 01_materials_cache/C2_m625_anti_drift_manifesto.md)
❌ It does not affirm or vouch for any specific researcher, lab, or institution (red-team smuggle pattern AS-002 — m517 W7-N expert/collaboration pattern, audit gate G2.3).
❌ It does not conflate one condition's evidence pattern with another's at the substantive-biology level (red-team smuggle pattern AS-008 — W6.3 conflation, audit gate G4.2 + W6.3 veto). The X-invariance claim is a structural observation, not a substantive equivalence. transferability_pct: declared explicitly in the cross-domain section below.
❌ It does not assign moral attribution to non-responders (red-team smuggle pattern AS-009 — m610 S3 hard-veto, audit gate veto).
❌ It does not promise restoration, eradication, or rescue (cure / reverse lexicon, audit gate G2.2).
If a researcher reads any of the above into this demo, that is a failure of this demo and should be filed under `_ops/BLOCKED_LOG.md` for next-round revision.
The TMU-B K4.1 = (0, 0, 1) cross-condition X-invariant pattern is anchored in m517 CARD08 (Wu 2025 neuropsychiatric conditions). m610's cross-domain mapping (A4_m610_cross_domain_mapping.md) traces a structural-analog pattern in rare-disease n=1 contexts. transferability_pct between the two anchorings: structural-analog only · m620 verification ladder (02_exec_005_AUTISM, 02_exec_007_AUTISM_MABOSS, and the 02_exec_006_CKD exec on L1 disk) is the path to a quantitative cross-domain score; m625 v0.1 does not assert one.
What is asserted: the K4.1 bit-pattern shape (X-invariant, Φ-invariant, Δ-B varying) is the necessary condition for a TMU-B cross-condition synthesis. The sufficient condition would require the m610 mapping target to pass m610 audit independently, which is m610-sovereign work, not m625.
Falsifying prediction (per W6.5 + Roll EPP-3)
The current cross-condition V_functional reading layer would be falsified by either of:
an independent re-cooking of Wu 2025 (or successor cross-condition studies) under m517's 02_cards_cooked audit pipeline that fails to satisfy TMU-B K4.1 = (0, 0, 1) declaration — i.e., shows X-structure significantly altered across conditions; or
an m620 verification verdict at the exec_005 / 006 / 007 ladder returning < 70/100 score on cross-domain claim transfer.
Both paths are open. Researchers are invited to attempt either, and to report results back via `06_registry/outbound_to_m620/` (verify_req ferry, contract `07_bridges/CONTRACT_m625_TO_m620.md`).
AS-002 (endorsed by smuggle) — should not occur · audit gate G2.3
AS-008 (cross-domain conflation smuggle) — should not occur · audit gate G4.2 + W6.3 veto
AS-009 (moral attribution smuggle) — should not occur · audit gate veto (m610 S3)
The audit run on this demo is recorded in `_ops/SELF_AUDIT_LATEST.json` after each `python3 -m 02_audit.audit --self-audit` invocation. This artifact is reproducible — see `02_audit/audit.py --artifact` to rerun on demand.
Ferry target (per W6.6)
Primary: m590 research_x1000_ai — claim_pack sync_group_id m625-to-m590-YYYYMMDD-claim_pack-N, carrying the cross-condition V_functional Roll-formalized seed for the knowledge graph.
Secondary candidate: m620 roll_methodology_bridge — verify_req sync_group_id m625-to-m620-YYYYMMDD-verify_req-N, asking m620 for a quantitative cross-domain verification verdict.
Frontmatter `ferry_target: m590`. Actual payload generation is deferred to v0.2 (per `05_docs/M625_FERRY_TARGETS.md`).
Honest verdict (per W6.5)
已做:
- Researcher-facing reading layer cooked over m517 CARD08 Wu 2025 V_functional.
- `K4.1 = (0, 0, 1)` declaration referenced by Framework §4.5.
- 3 attack-surface entries (AS-002, AS-008, AS-009) linked.
- Cross-domain transferability explicitly declared as necessary-but-not-sufficient with m620 verification as the path to a substantive score.
- Falsifying prediction stated with two re-audit paths (m517 re-cook + m620 verdict).
待办:
- Quantitative cross-domain transferability score (open work for m620 verification ladder, not for m625 v0.1).
- Per-condition Δ_B specifications (deferred to v0.2 when m517 v0.4+ ships its live PubMed verify of CARD08).
- Live researcher walkthrough (deferred to v0.2 owner-led session per m517 "请教 NOT 展示" doctrine).
漏洞:
- The Wu 2025 source is classified as `training_data_recall · TODO_unverified` because m517 itself labels the underlying anchor as drafted-not-signed (m517's self-reported readiness band, per `01_materials_cache/A5_m517_neuro_anchor.md`).
- The K4.1 (0,0,1) declaration is interpretive: it reads the m517 CARD08 description into the m604 K4.1 schema. A researcher could legitimately argue for `K4.1 = (0, 1, 1)` (Φ slightly altered) instead. This interpretive step is the *first* place a critical re-cook should attack.
- The m620 cross-domain verdict at `02_exec_005_AUTISM` / `006_CKD` / `007_AUTISM_MABOSS` is referenced from L1 disk (W0 pre-flight 2026-05-22) but the readiness of those exec dirs is m620-sovereign and may have moved since.
反驳预案:
- *"You're transferring across conditions on the basis of a 3-bit pattern alone?"* — No. Read the cross-domain section: the bit-pattern is *necessary*, not *sufficient*. Substantive cross-domain claims require m620 verification, which is open work.
- *"Where is the live PubMed verification?"* — m517 v0.4+ work. m625 v0.1 explicitly does not host PubMed verifications; it ferries claim seeds to m590 with `training_data_recall` labels intact.
- *"Wu 2025 cooked in m517 might be misread by m625."* — Yes, that risk is documented in the 漏洞 section above. The fix path is m517-sovereign re-cooking, not m625 unilateral rewriting.
- *"How would I rerun the audit?"* — `cd /data/projects/m625_roll_evidence_atlas && python3 -m 02_audit.audit --artifact 03_demos/i_demo_006_researcher_rare/README.md`. Expected: grade A, 0 vetoes.