freeJames/Atlas-of-Guilt

Atlas of Guilt (罪迹拓谱) — v0.7.0

Synthetic, memory-grounded causal graphs of wrongdoing and its consequences, organized as one atlas per person.

In the novel 《罪迹拓谱》 (Atlas of Guilt), by this project's author, a superintelligence called Jesus judges people from their digitized memories. Each person has an atlas of guilt. It holds:

• every act that person committed;
• the memories of the people those acts touched;
• the consequences, spreading outward layer after layer.

The novel is only where the idea comes from. This project tries to build that structure for real, as data and code. The aim is that one day it can take in real memory exports, and that it can be used to study how responsibility might be computed.

Everything in v0.7.0 is synthetic and fictional.

• The people are invented and appear only as letter codes.
• The calendar is invented. There are no real places or institutions.
• The patterns of harm are taken from documented historical patterns that recur in several eras and regimes. Each is rewritten at the level of structure only, so no event maps back to a real case or a real person.

Interactive preview: https://claude.ai/artifact/T8gk5KEJAMRw8p4Y61X4eU

• A web page with the person-centered radial view of both atlases. Each person is a node; each line is "this thing that happened", with the memories behind it. Click a person and their lines fan out, layer by layer.
• It also shows the ledger of jesus_v0.1, harm sequences, victims' own part, and how the atlases grew batch by batch.
• The page is updated in place, so it may show data newer than this release. The files here are the reference.

This is an early release. The structure is the point; the data is still small. It grows in append-only batches:

• B0001 wrote the two atlases.
• B0002 filled every gap in the existing chains: unmodeled perpetrators, unidentified speakers, missing states of mind, cohorts without a victim's own memory. It also added people, events and cross-links at every layer.
• B0003 added harm sequences, where one person harms another several times and the acts escalate. It also added victims' own choices that contributed to their harm.
• B0004 added more events and links at every layer:
  • upstream: the people who taught D and E their habits;
  • the owners' own later acts: guarding grain from the starving, banning a memorial, ordering files burned;
  • co-perpetrators: the clerk who burned the files, and the colleague E reported;
  • victims and their families, down to a fourth generation that goes looking for the files;
  • beneficiaries: D's family moving into the house of a man D had taken away;
  • links across branches, such as a clerk promoted in the first famine who refuses grain to E's mother in the second.

The release contains:

• Two atlases, finished for v0.1. Their remaining gaps are reported in jesus_v0_1_gaps and in each manifest's completeness metrics.
  • P-000001 is a fictional dictator, D.
  • P-000002 is a grassroots enforcer, E. He carried out D's purge. His mother died in a famine that D's system caused. Decades later he apologized.
• 1684 first-person memory segments from 570 fictional people, across 14 memory channels. Each segment is in Chinese, with an English rendering.
• 1569 atlas nodes and 1905 causal edges. Every edge lists the memory segments that support it.
• The output of jesus_v0.1, a first, rule-based version of the judgment algorithm. It produces a ledger with 8489 lines: one per perpetrator, victim and outcome. Each line gives a share of responsibility, labels, a sentence in seconds, and an explanation that traces every number back to memories, edges or parameters.
• A small set of training questions about causal tracing, evidence provenance and time order. Each answer carries its epistemic status.

Core ideas

• One atlas per person, one shared graph. Every record lives once in a flat, append-only graph. A person's atlas is an index over it:

  • downstream, only the harms in which they have a causal share, however many layers down, and the links leading to them. The share is traced along the same causal paths the judgment algorithm uses. A person harmed further down by someone else, with no causal thread back to the owner, is not in the owner's atlas;
  • the benefits that flowed from their harm chains;
  • everything upstream that shaped their decisions;
  • all the memories that support these, including other people's.

  The same victim's memory belongs to the atlas of everyone who harmed them, and any person in an atlas can be opened as the center of their own. P-000001 and P-000002 share most of their evidence.

• An atlas models one person's own wrongdoing. The harms a person suffered are never modeled with them at the center: each one already appears in the atlas of whoever caused it.

• No crimes, no atlas. Someone who only suffered harm has no atlas of their own.

• Feedback (回流). People downstream can come back to influence or harm the owner. A harm the owner suffers is drawn in his atlas as feedback, but it is not in his victims layer or his account: as a victim, he is not judged in his own atlas. When the owner then harms someone who first harmed or influenced him, that person's earlier act enters the share of the new harm as their own part, marked feedback in the ledger, and the owner's share falls accordingly. Feedback is one of the factors of the judgment.

• Layers of an atlas. Besides the owner's own acts, every manifest lists:

  • the co-perpetrators (施害同谋层级): everyone who helped cause any harm in the owner's atlas, at any distance, each with their relation and their distance in causal steps:
    • participants, such as executors, accomplices, informers, silent witnesses who could have spoken, those who falsified records, and those who started new wrongs further down the owner's chains;
    • instigators who ordered, prompted or incited the owner's acts;
  • the victims, whether each one died, and how many deaths each harm caused;
  • the people upstream who influenced or instructed the owner;
  • the beneficiaries;
  • everyone else whose memories serve as evidence.

  In the preview, co-perpetrators who were also harmed or who benefited appear through their causal lines. The others are listed only when a particular crime is opened, each with their share in it.

  The memories of co-perpetrators and witnesses corroborate the owner's atlas. An atlas is never built from the owner's and the victims' memories alone. For every harm, the ledger records each perpetrator's share of the cause.

• Memory is the evidence.

  • A person's knowledge, intentions, doubts and fears come only from their own memory.
  • What someone heard is recorded as hearsay, not as observation.
  • Official records are kept alongside memories even when they contradict them: here, a falsified death record sits next to the memories of the people who saw the injuries.
  • Every conclusion reports its memory composition: how much comes from the center person's own memory, how much from victims and others, and how much from records.

• Stage 1 does not decide what is a crime. Everything causally connected goes into the graph. Guilt and responsibility are computed separately, by a versioned ruleset, and the results are never written back into the memory or fact layers.

• Counterfactuals are for people, not for the algorithm. A note such as "had he not been arrested, he might have lived ten more years" is kept in its own table, marked COUNTERFACTUAL and display-only. It never enters any computation; a test removes every such note and checks that all derived outputs stay byte-identical.

• Harm sequences. One person often harms another more than once, and the acts escalate. Example: the enforcer E presses H to name others, slaps him, keeps him awake, then withholds his heart medicine. Each act is its own node, and ESCALATED_TO links an act to the one it escalated into. Every manifest lists these sequences, and every ledger line carries the perpetrator's full sequence against that victim. Escalation describes the pattern only. The harm of each act is already counted through causal edges, so splitting one event into several nodes cannot count the same person twice. A test removes every escalation edge and checks that no share or sentence changes.

• A victim's own part. Sometimes a victim's own state of mind and choices are among the causes of their harm. Such a choice gets a share, read only from the victim's own memory:

  • informed: they knew the risk to themselves and went ahead;
  • uninformed: they did not know, or were deceived;
  • coercion reduces the weight.

  This share is not guilt. It produces no ledger line and no label for the victim, and only reduces the perpetrators' shares. Resistance, appeals and helping others never count as a victim's own part. A deceived or coerced choice does not shield the perpetrators: their paths through it are attenuated as if the victim had been acting on orders. Example: a labor-camp medic tells RR that his gangrene is "frostbite, nothing serious", and RR goes back to work and loses the use of his hand. RR's own part is about 3%; the medic's is about 48%.

• A web, not a tree. Downstream people connect to others in their own layer and to people several layers down other branches. For example:

  • a famine survivor becomes the activist in a purge;
  • the grandson of a purge victim is turned away by the son of a man D denounced, while D's own grandson takes the place;
  • the enforcer's son is shamed at school by the victim's grandson, and punished by a teacher who was orphaned by the famine;
  • decades later, the grandson of the man D denounced turns away the victim's great-grandson because of the same falsified file, while D's grandson, who knows, says nothing.

  The growth is measured with the same metrics after each batch (see CHANGELOG.md):

atlas	after batch	nodes	edges	memories	other people	branches meet	same-layer links	links back up	people in several branches	pairs with several acts	longest escalation
P-000001 (D)	B0001	107	120	139	32	29	4	3	20	1	1
P-000001 (D)	B0002	180	228	230	54	58	17	15	33	1	1
P-000001 (D)	B0003	216	278	279	55	74	21	18	38	3	5
P-000001 (D)	B0004	238	303	307	56	80	21	19	38	4	5
P-000001 (D)	B0005	239	316	309	56	86	22	21	38	4	5
P-000001 (D)	B0006	273	366	351	68	112	23	25	49	4	5
P-000001 (D)	B0007	312	421	396	86	131	23	31	61	4	5
P-000001 (D)	B0008	313	422	397	87	132	23	31	62	4	5
P-000001 (D)	B0009	318	428	404	88	136	23	31	63	4	5
P-000001 (D)	B0010	370	495	458	119	138	23	31	65	4	5
P-000001 (D)	B0011	423	562	508	150	138	23	31	65	4	5
P-000001 (D)	B0012	492	649	576	196	141	24	32	68	4	5
P-000001 (D)	B0013	547	710	627	229	141	25	32	68	4	5
P-000001 (D)	B0014	559	722	644	232	147	25	32	70	4	5
P-000001 (D)	B0015	591	764	675	252	173	25	32	88	5	5
P-000001 (D)	B0016	609	784	693	263	181	25	32	96	5	5
P-000001 (D)	B0017	834	1021	918	354	181	30	38	96	5	5
P-000001 (D)	B0018	996	1208	1080	412	284	35	45	150	5	5
P-000001 (D)	B0019	996	1208	1080	412	284	35	45	150	5	5
P-000001 (D)	B0020	1140	1369	1224	456	374	42	48	191	6	5
P-000002 (E)	B0001	80	87	108	25	16	1	1	10	1	1
P-000002 (E)	B0002	141	168	181	41	25	2	7	17	1	1
P-000002 (E)	B0003	165	203	216	41	27	5	7	18	3	5
P-000002 (E)	B0004	193	232	245	44	32	5	8	19	3	5
P-000002 (E)	B0005	204	256	259	47	33	5	9	19	3	5
P-000002 (E)	B0006	204	256	259	47	33	5	9	19	3	5
P-000002 (E)	B0007	209	263	265	48	33	5	9	19	3	5
P-000002 (E)	B0008	210	264	266	49	34	5	9	20	3	5
P-000002 (E)	B0009	214	271	272	49	37	5	10	21	3	5
P-000002 (E)	B0010	288	372	348	89	40	5	11	23	3	5
P-000002 (E)	B0011	364	479	421	132	43	5	12	24	3	5
P-000002 (E)	B0012	457	606	511	191	43	6	13	24	3	5
P-000002 (E)	B0013	512	670	562	224	43	7	13	24	3	5
P-000002 (E)	B0014	588	771	649	246	67	10	21	46	12	5
P-000002 (E)	B0015	643	841	703	277	67	10	21	46	13	5
P-000002 (E)	B0016	681	886	740	296	79	10	22	53	13	5
P-000002 (E)	B0017	909	1126	968	388	95	15	27	62	13	5
P-000002 (E)	B0018	1101	1357	1173	453	104	21	32	65	14	5
P-000002 (E)	B0019	1193	1451	1265	488	162	23	32	96	14	5
P-000002 (E)	B0020	1340	1617	1412	532	162	30	36	96	15	5

Main events and the propagation frontier

A life holds many wrongs. Some a person led; in others they assisted. Each is a main event (main_events): an owner, the owner's acts where it starts, and a role, either principal or accessory.

How do we know when a main event has been followed to its end? Walk down from its starting acts along every causal, agency, obstruction, influence and benefit link. Every node where a branch ends (a leaf) must either be continued, or carry a branch end (branch_ends) saying why it stops there:

• no further harm followed;
• the person died and nothing further followed from them;
• it lasts to the present (year 62 of the fictional calendar);
• the available memories stop here (to be reopened when real memory exports arrive);
• for a group harm, the remaining members stay in the group.

A main event is complete only when no branch is open. If a later batch adds a link below a node that had a branch end, the end lapses automatically and is reported (ends_reopened), so nothing has to be edited.

Groups and individuals. Harm to tens of thousands of people is recorded as a group (cohort) harm. A member is split out as an individual if they cause further harm, even if the upper layers contributed only a little, or if their own harm was caused by a particular person's act:

• a MEMBER_HARM link marks "this is that member's part of the group harm";
• their part is deducted from the group, so no one is counted twice;
• their own chain continues downstream.

Members who cause no further harm stay in the group. Relationships learned later, including group membership, are added as separate records (kin_links) without editing the person.

atlas	main event	role	nodes reached	layers	branch ends needed	still open	groups not yet closed	individuals split from groups	estimated members not yet expanded	status
P-000002 (E)	Guarding the granary at gunpoint	accessory	189	88	51	0	0	0	0	complete
P-000002 (E)	Keeping the truth down: reporting W; threatening, watching and reporting F	principal	195	67	62	0	0	0	0	complete
P-000002 (E)	Arrests, interrogation and night watch in the purge	accessory	574	91	189	0	0	0	0	complete
P-000002 (E)	Arresting M	accessory	184	77	49	0	0	0	0	complete
P-000001 (D)	The first purge campaign	principal	901	95	270	0	0	73	1980	complete
P-000001 (D)	Putting in a word for his grandson	principal	8	5	2	0	0	0	0	complete
P-000001 (D)	The second campaign	principal	336	85	89	0	0	15	60	complete
P-000001 (D)	Banning the memorial and destroying the files	principal	8	2	4	0	0	0	0	complete
P-000001 (D)	Denouncing his classmate L	principal	25	7	4	0	0	1	60	complete
P-000001 (D)	Punishing G for reporting the truth	principal	486	89	119	0	0	28	5060	complete
P-000001 (D)	False harvest reports and the first famine	principal	470	92	117	0	0	28	5060	complete

Every main event of P-000001 and P-000002 is followed to its end. "Estimated members not yet expanded" counts group members who, by the author's estimate, also caused further harm but are not yet individual nodes. Groups are a compromise forced by present compute. In the end, every person is a node, and two people are usually linked by many events across many layers.

Main events can relate to each other:

• a main event can have co-principals: U, who levied grain on D's false figures, co-led the first famine;
• one person's accessory part can belong to another's main event: E's arrests are part of D's first purge.

Open-ended structure

The model is meant to hold any person's wrongs. Five structures keep it open:

• Harm types are a hierarchy with no fixed depth.
  • A code is a dotted path, such as education.admission.identity_theft.
  • A class not yet named is written as other(free description) at the end of a path, and counted by its parent until it is mapped.
  • Points are looked up at the most specific registered level.
  • The vocabulary is small today only because of present compute. Any difference, even one point of harm value, can become its own class.
• Organizations are labels on people, not actors. An affiliations record says a person belonged to a unit, in a role, for a period. A collective decision is still made by individuals, each with their own atlas.
• Repeated acts carry a period, a count and a frequency, for example once a week, about 40 times.
• Identity alignment and corrections.
  • identity_links says two codes are one person, for example "the clinic doctor" in one man's memory and the medic FD in another's.
  • Any record can be corrected by a new record that supersedes it. The old record is kept.
  • All computation runs on the resolved view.
• Incremental recompute. Each harm is computed from its input slice: its full upstream, the forgiveness that concerns it, its victims' harm sequences, the memories behind them, and the parameters. If the slice's fingerprint is unchanged, the earlier result is reused. Tests check that reused and recomputed results are byte-identical.

The two atlases

atlas	owner	nodes	edges	memory segments (own / others')	other people	causal depth (down / up)
P-000001	D, a fictional dictator	1140	1369	1224 (57 / 1167)	456	95 / 6
P-000002	E, a grassroots enforcer	1340	1617	1412 (199 / 1213)	532	91 / 36

D. His co-perpetrators layer has 149 people: Q, N, DG, EC, K, M, E, J, O, R, Y, S, U, A, Z, X, NS, LC, CL, AS, CN, SF, UZ, MD, IC, WT, CM, JW, LCS, FV, GS, FD, AC, BES, BC, BE, BD, BB, BA, BH, BF, BJ, BL, BK, BHH, BKS, BG, EV10, EV04, EV18F, EV28S, EV31, EV27S, EV39H, EV42C, EV45N, EV45C, EV65B, EV62, EV63M, EV64S, EV64O, EV62K, EV61S, MV06H, GV03, MV03S, EV55H, S1T, S1EH, S1K2G, S1QDB, S1C1AD, S1W3, S1Y, S1EE, S1BOSS, S1B1F, S1B1, S1P, S1QS, S1D, S1BT3S, S1W3W, S1A, S1AP, S1C, S1ML1, S1N, S1K2S, S1GG, S1U3F, S1WC, S1WO, S1N2B, S1R, S1RG, S1O, S1ML2, S1EL2S, S2MF3, S2RV5, S2TH, S2BU, S2CLH, S2HU, S2BR, S2MB, S2FM, S2FH, S2GR, S2FMS, S2BUM, S2CP, S2MWD, S2CLD, S2DR4S, S2SVS, S2TS, GV03S, S2FD, S2CLFF, S2RV3S, S2RV2, S2PG, S4V1S, S4V1H, S4SG, S4P1, S4PO, S4G3B, S4OLD2S, S4CL, S4G2F, S4V1B, S4BBF, S4SGC, S4CLW, S4BB, S4TG, S4K, S4V2, S4BLD, S4G3, S4BW, S4BD, S4A, S4V2M, S4HS.

• As a student, he denounced a classmate to protect himself. His memory records a 3.2-second hesitation and the thought "if I don't report it, someone else will."
• As a local official, he filed harvest figures he knew were inflated. He foresaw that people would starve, and he buried the first report of deaths. A cadre who petitioned with the real figures was purged.
• After taking power, he launched a purge with quotas, and later a second one.
• Upstream of him are his father's lesson that "obedience is the only way up," a teacher's praise for informers, and a superior's hint to "make the report look good."

E. His co-perpetrators layer: Q, N (instigator), D, DG, EC, K, M, J, O, R, Y, S, U, A, Z, X, NS, LC, CL, AS, CN, SF, UZ, MD, IC (instigator), WT, CM, PB, JW, LCS, FV, GS, AC, BH, EV10 (instigator), EV04, EV18F, EV28S, EV31 (instigator), EV27S, EV39H, EV42C, EV45N, EV46H, EV45C, EV65B, EV62, EV63M, EV64S, EV64O, EV62K, EV61S, EV67S, MV06H, GV03, MV03S, EV68H, EV55H, S1T, S1EH, S1K2G, S1QDB, S1C1AD, S1W3, S1Y, S1EE, S1BOSS, S1B1F, S1B1, S1P, S1QS, S1D, S1BT3S, S1W3W, S1A, S1AP, S1C, S1ML1, S1N, S1K2S, S1GG, S1U3F, S1WC, S1WO, S1N2B, S1R, S1RG, S1O, S1ML2, S1EL2S, S2MF3, S2RV5, S2TH, S2BU, S2CLH, S2HU, S2BR, S2MB, S2FM, S2FH, S2GR, S2FMS, S2BUM, S2CP, S2MWD, S2CLD, S2DR4S, S2SVS, S2TS, GV03S, S2FD, S2CLFF, S2RV3S, S2RV2, S2PG, S3GW, S3L2D, S3CAD, S3EG, S3B, S3CH, S3BD, S3H, S3D3, S3A, S3FS, S3BC, S3IH, S4V1S, S4V1H, S4SG, S4P1, S4PO, S4G3B, S4OLD2S, S4CL, S4G2F, S4V1B, S4BBF, S4SGC, S4CLW, S4BB, S4TG, S4K, S4V2, S4BLD, S4G3, S4BW, S4BD, S4A, S4V2M, S4HS. It includes the officials who ordered him and the neighbor whose denunciation set him moving, marked as instigators.

• He arrested people from the list. He knew one of them, H, was "no enemy," and he pushed away H's daughter, F.
• On night watch he kept H awake. His memory records that he wondered whether H would survive. A colleague, W, refused to take part and was only reassigned, so E knew there was a choice.
• Years later he arrested M, the official who had sent him.
• In old age he apologized to F, and confessed to the daughter of a teacher he had taken away.
• Upstream of him are his father's "bad background," his fear, and the orders he received.
• His mother starved in the second famine. That harm is not part of his atlas; it is recorded in the atlases of those who caused it.

jesus_v0.1: a first judgment algorithm

This is a research baseline, not a verdict about anyone. It follows the rules stated in the novel where the novel gives them (see code/rulesets/jesus_v0.1/RULES.md and params.json). Every parameter is tagged BOOK, CALIBRATED or DESIGN. At this stage the work is about structure, so the values the novel does not give are set by the project's AI author.

1. Structure. For each harm, enumerate every causal path back to the acts and natural processes that contributed. Strength is the product of edge necessities, attenuated each time the path passes through another agent: less for an agent acting on orders (×9/10), more for an agent acting freely (×1/4).

2. Mind, from one's own memory only. Classify each act toward each outcome:

   • intent: intended that kind of harm, or foresaw it and persisted;
   • knowing;
   • negligent;
   • omission: knew, could act, had the chance, and did not;
   • no malice;
   • unknown.

   A credible threat in the person's own memory reduces the weight, unless they also knew there was an alternative. "Everyone was doing it" does not count as a threat.

   A victim's own acts that led to their own harm are classified separately, as informed or uninformed. They take a share but are never guilt.

3. Shares and sentences. Share = path strength × mental weight, normalized over all factors, natural ones included. Sentence (seconds) = harm value × share, with one second per point, as in the novel. The harm value depends on outcomes, never on motives.

4. Labels.

   • Intentional homicide: intent or foresight of death, and an act that is a necessary link in the actual causal structure.
   • Silent node.
   • Mastermind.
   • Non-endogenous gain, for benefits traced to a wrongful act. Refusers and conduits are not labeled.

   Factors with no malice receive shares but never labels: "they were the cogs of fate, not the executioners."

5. Forgiveness. It changes only the forgiving victim's own lines with that perpetrator. No one can forgive on behalf of the dead. The algorithm never forgives or pursues on anyone's behalf.

Calibration. The novel (ch. 34) gives one worked example: among 48 factors in a child's death, the only one with malice carried 32% of the responsibility, which equals a sentence of 25 days, 1 hour and 5 seconds. jesus_v0.1 reproduces this exactly on a 48-factor test fixture. The other values are design choices and will change as the research proceeds.

Invariants, all tested (42 of 42 tests pass):

• outputs are bit-for-bit deterministic;
• the computation is read-only;
• forgiveness changes only the forgiven lines, and counts only when given by the victim in person while alive (adversarial tests included);
• thoughts that were never acted on produce no lines;
• harms a person suffered are not part of their own ledger, so they can never offset it;
• late recollections, self-justifying labels and lifespans do not change the result;
• shares are conserved;
• every number is traceable.

Example: the death of H, who was arrested by E, denounced by N, paraded by A and kept awake for three days:

person	share	mental category	labels
E	41.5%	INTENT/KNOWING/NEGLIGENT	INTENTIONAL_HOMICIDE
IC	21.1%	INTENT	INTENTIONAL_HOMICIDE
D	18.3%	INTENT/KNOWING	INTENTIONAL_HOMICIDE
A	8.8%	KNOWING	—
M	4.8%	KNOWING	—
N	1.5%	KNOWING	—
WT	1.0%	KNOWING	—
X	0.9%	KNOWING	—
nature	0.7%	NONE	—
U	0.7%	KNOWING/NEGLIGENT/NONE	—
CL	0.2%	KNOWING	—
EO	0.1%	NONE	—
CM	0.1%	NONE	—
EF	0.1%	NONE	—
K	0.0%	OMISSION	—
Y	0.0%	OMISSION	—

The people who stood closest to this death and foresaw it carry more of it than D, who launched the campaign: E, who kept H awake and IC, who ordered it. D's own ledger is far larger, because he launched the campaign and the famines. The shares of the first famine are D 79.5%, U 13.4%, CL 4.4%, nature 1.1%, K 0.8%, Y 0.8%.

• D: 544 lines; labels INTENTIONAL_HOMICIDE, MASTERMIND; composite coefficient 0.526.
• E: 622 lines; labels INTENTIONAL_HOMICIDE.
• Forgiveness:
  • F said to E, "your share — I forgive you." That waives only E's lines with F; her father's share is not hers to forgive.
  • J's daughter forgave only her own share. J died before E came to apologize, so E's lines with J remain: J's wife is pressured into divorce; JD's surname is changed (35,499 s, 9小时51分39秒); J is sent to a labor farm for six years (678,514 s, 7天20小时28分34秒); J develops cirrhosis (232,835 s, 2天16小时40分35秒).
  • F did not forgive D, and D's lines with F are unchanged.

Files and configs

All tables are JSON Lines with a fixed set of columns per file. Fixed-point numbers use the suffix _micro (millionths). Time uses the fictional JZ calendar (year 0 is the founding of the fictional state).

config	rows	content
memory_segments	1684	first-person memory: channel, text (zh/en), mentions, heard_from, audience
memory_exports	667	one synthetic memory export per person
episodes	1151	key moments that group segments
recall_events	1	late recollection, linked to the original decision without overwriting it
entities	576	individuals, cohorts (aggregated populations), a family lineage
atlas_nodes	1569	actions, decisions, mental states, events, harms, transfers of benefit, records, natural processes, apologies, forgiveness
edges	1905	typed causal, agency, cognitive, obstruction, benefit and protection edges, each with memory basis
atlas_manifests	2	the two atlases: membership index (with roles) and completeness metrics
jesus_v0_1_*	—	ledger, per-perpetrator summaries, benefit labels, responsibility features, gap reports, run record (split computed)
training_qa	1163	questions on evidence provenance, inner-state provenance, victim tracing, hearsay and time order (split computed)

Every record carries namespace (always synthetic), epistemic_status and created_by (with model version and prompt hash for model-written records). Data and derived records also carry a content_hash over their canonical JSON. manifest.json lists the SHA-256 of every file.

from datasets import load_dataset
segments = load_dataset("<namespace>/<dataset-name>", "memory_segments", split="curated")
ledger = load_dataset("<namespace>/<dataset-name>", "jesus_v0_1_ledger", split="computed")

The code/ folder (Apache-2.0) contains the authoring source, the builder, the validator and jesus_v0.1 with its tests. Running python tools/atlas.py build validate jesus test inside code/ rebuilds every record and derived output, and the results match this release bit for bit.

Intended uses

• Research on causal responsibility attribution over long, branching chains with many agents.
• Evaluating whether models can reason from evidence with provenance: whose memory says what, what is hearsay, what is only a record.
• Prototyping and stress-testing responsibility algorithms: shares, attenuation, omission, forgiveness, non-offsetting.
• A seed for larger synthetic atlases. Future versions add more people, deeper layers and more patterns, without changing the structure.

Out-of-scope uses

• Judging, scoring or profiling any real person, or treating this algorithm's outputs as moral or legal truth.
• Any real-world adjudication, surveillance or punishment.
• Treating the stories as history. They are fiction built from recurring patterns.

Sensitive content

The atlases depict:

• political persecution, public denunciation sessions and sleep deprivation;
• famine and falsified death certificates;
• forced labor and deaths in custody;
• coerced divorce;
• discrimination against children and humiliation at school.

Harm is recorded mostly as structured fields (type, count, severity, duration), without graphic detail.

Limitations

• Tiny and hand-authored. The data was written by an AI model (Claude, claude-opus-5-5) following a written brief, and checked by an independent model instance. It is not real memory.
• Estimates and placeholders.
  • Edge necessities are authoring estimates.
  • Most jesus_v0.1 values are design choices: harm weights, attenuation, the duress factor, the necessity threshold.
  • Only one point is calibrated to the novel.
  • The novel's long-chain example (ch. 33), dose liability, pursuit and revocation are not implemented yet.
• Social texture. The people are fictional, but the patterns draw mainly on twentieth-century political campaigns.
• A victim's own part is a share, not a judgment. The weights for informed and uninformed choices are design choices. Whether a pattern of escalation should weigh more is a stage-2 research question.
• Counterfactuals never count. What a victim's life could have been is shown to readers as a labeled note, by design. Harm values use only what actually happened.
• The output form is not settled. The ledger is one candidate. Later versions may change it.

Versioning

Releases are append-only: a published version never changes. Corrections arrive as new records that supersede old ones, in a new version. Cite the version you use.

Numbering: the minor version rises (0.4.0, 0.5.0, ...) only when a whole main event has been followed to its end. Batches in between are 0.x.1, 0.x.2, and so on.

License

Data: CC BY 4.0. Code (code/): Apache-2.0.

@misc{atlas_of_guilt_2026,
  title  = {Atlas of Guilt (罪迹拓谱): synthetic memory-grounded atlases of wrongdoing, v0.1.0},
  author = {{Atlas of Guilt project}},
  year   = {2026},
  note   = {Synthetic dataset; all persons and events are fictional}
}

---

中文说明

罪迹拓谱：以记忆为证据的罪行因果图，每人一份。

• 构想出自本项目作者的小说《罪迹拓谱》；小说只是这一设定的出处，这里做的是这个结构的现实构建。
• 将来可以接入真实的记忆导出，并用来研究审判算法。

网页预览：https://claude.ai/artifact/T8gk5KEJAMRw8p4Y61X4eU

• 两份拓谱的人物放射图：节点是人，线是"这件事"，带着记忆片段。点一个人，他的线就放射出来，一层层展开。
• 也能看到 jesus_v0.1 的账本、施害序列、受害者自身因素，以及拓谱每一批的生长。
• 预览页会原地更新，可能比这个版本的数据更新；以这里的文件为准。

v0.7.0 是早期版本，全部为合成的虚构数据。

• 人物只用字母代号，纪年为虚构的"建政纪年"，没有真实的地名和机构。
• 用到的每一种模式都在多个时代或政权中出现过。虚拟化只保留结构与机制（V1 模式级），任何事件都对不回真实的案件或真实的人。

内容

• 两份拓谱（v0.1 已完成，缺口另有报告）：P-000001（虚构独裁者 D）、P-000002（基层执行者 E）。
• 施害同谋层级：所有与他的罪行有关的人都收，包括指使者，并记下关系与距离。
  • 包括共同参与者（执行者、帮凶、告发者、沉默者、伪造者，以及在他的因果链下游另起恶行的人），以及指使、引发他行为的人。
  • 他们的记忆同样是这份拓谱的佐证；账本记下同一伤害上每个施害者所占的因果比重。
• 只为本人的罪行建模：一个人受过的伤害不进入他本人的拓谱与审查，也不以他为中心另行建模；每一次受害，都记在相应施害者的拓谱里。
• 没有罪行就没有罪迹拓谱：只受过害的人没有自己的拓谱；预览里点开他，列出伤害过他的人。
• 回流：下游的人可能回过头影响、伤害主人公。
  • 主人公受的伤害画在他的拓谱里（射回中心的虚线），不进受害者层，也不进他本人的账：作为受害者，他不在自己的拓谱里审查伤害他的人。
  • 主人公随后伤害了先前影响他的人时，这个人先前的行为按受害者自身因素进入这笔伤害的份额（账本标 feedback），主人公的占比相应减少。回流是 Jesus 判罚的因素之一。
• 只收有份额的受害：拓谱的下游只收主人有因果份额的伤害，以及通往它们的中间环节，无论多少层。份额沿 Jesus 计算所用的因果路径追溯。
• 预览中的施害同谋：受害或受益的同谋由因果线连到；其余的，点开某件罪行时才列出，附上各人的份额。
• 施害序列：一个人对另一个人的施害可能是多个事件，事件之间可能递进。每个事件是一个节点，递进用 ESCALATED_TO 连起来。
  • 递进只描述，不改变份额：每个事件造成的伤害，已经按因果边分别计入。
• 受害者自身因素：受害者本人的心态与抉择，有时也是他受害的原因之一。
  • 心理只读他本人的记忆：知情，或不知情、被骗；受胁迫时权重降低。
  • 自身因素只占份额，不是罪：不生成他的账，只减少施害者的份额。
  • 抵抗、申诉、救助都不算；被骗、被逼的抉择不切断施害者的因果。
• 数据规模：1684 段第一人称记忆，1569 个节点，1905 条带记忆依据的因果边。
• jesus_v0.1 的输出：Jesus 审判算法的初级版本，是一个确定、可重算、可解释的规则版。

主事件与传导前沿

• 一个人一生有许多主事件，有他主导的，也有他辅助、参与他人主导的。
• 从主事件的起点往下走，走到头的分支要么继续写，要么写明停在这里的原因。没有一个分支还开着，这件主事件才算完成。
• 群体受害者里，引出了次生伤害的，或者受害本身由某个个人的行为造成的，单列成个人，继续往下延伸；其余的留在群体里。单列的那一份从群体中扣除，不重复计算。

拓谱	主事件	主导/辅助	传导范围（节点）	层数	走到头的分支	还开着	群体未处理	从群体中单列	群体中估计尚未展开	状态
P-000002 (E)	端枪守粮库	辅助	189	88	51	0	0	0	0	已完成
P-000002 (E)	压住真相：告发 W，威胁、盯梢并告发 F	主导	195	67	62	0	0	0	0	已完成
P-000002 (E)	整肃中抓人、审讯、看守	辅助	574	91	189	0	0	0	0	已完成
P-000002 (E)	抓走 M	辅助	184	77	49	0	0	0	0	已完成
P-000001 (D)	第一轮整肃运动	主导	901	95	270	0	0	73	1980	已完成
P-000001 (D)	为孙子打招呼	主导	8	5	2	0	0	0	0	已完成
P-000001 (D)	第二轮运动	主导	336	85	89	0	0	15	60	已完成
P-000001 (D)	禁止纪念、销毁档案	主导	8	2	4	0	0	0	0	已完成
P-000001 (D)	告发同学 L	主导	25	7	4	0	0	1	60	已完成
P-000001 (D)	整治如实上报的 G	主导	486	89	119	0	0	28	5060	已完成
P-000001 (D)	虚报产量与第一次饥荒	主导	470	92	117	0	0	28	5060	已完成

• 两份拓谱的全部主事件都已传导到最后一层。
• 群体只是迫于现在的算力；将来每个人都是一个节点，两个人之间往往有很多件事、很多层关联。

开放的结构

• 伤害类型分层、开放，可以无限细分。
• 组织是人身上的标签，不是行为者。
• 持续与反复的行为带起止、次数与频率。
• 身份对齐与更正：旧记录保留，计算用对齐、更正之后的视图。
• 只重算受新数据影响的那一部分。

jesus_v0.1

• 校准：复现了书中第34章的数例——四十八条因素中唯一存心者占32%，对应刑期25天1小时5秒。
• 测试：42 项不变量测试全部通过。
• 参数：除书中给定与校准的参数外，其余数值由 Claude 设定（DESIGN）；现阶段只设计结构。
• 反事实推演：只给人看，不进入任何计算，有测试保证。
• 输出形态：书中的账本形态，是候选之一，不预先定死。
• 用途与边界：只用于研究，不得用于评判任何真实的人。