# ๐ŸŽจ Agent Execution Logs โ€” Visual Atlas ![Animated hero: 13 diagrams, 2 sealed runs, 236 tool calls, 32 findings, 0 invented numbers](diagrams/atlas-hero.gif) > **What this is** *(the fun version)*: forensic evidence, but you can actually look at it. > This page is the graphical companion to > [`AGENT-EXECUTION-LOGS-REPORT.md`](AGENT-EXECUTION-LOGS-REPORT.md): thirteen color diagrams, one > per major topic of the report, each **generated from and verified against the raw data files in > this folder** (`base-dc-report.json`, `notch-report.json`, the run logs, and the Thymus audit > trails) โ€” no invented numbers. Diagrams are committed as PNG (rendered with mermaid-cli) so they > always display on GitHub; every figure's Mermaid source is in a collapsed block beneath it. > > ๐ŸŽฌ **Moving pictures:** the same topics, animated โ€” the > [**108-second walkthrough video**](https://galvangabriel-web.github.io/agentropix-mcp/docs/12-CASES-REPORTS/srl-2018-report/submission/watch.html) (plays in-browser; > [details](README.md#-the-animated-walkthrough)). > > **Shared palette:** ๐ŸŸฆ blue = agents/tools/files ยท ๐ŸŸช purple = ATT&CK detectors & governance/seals ยท > ๐ŸŸฉ green = success/ALLOW/stable ยท ๐ŸŸฅ red = errors/REJECT/timeouts/gaps ยท ๐ŸŸง amber = correlation & anchors ยท > โฌœ grey = skipped/silent. --- ## 1 ยท ๐Ÿ—ฃ๏ธ Who talked to whom *(spoiler: everyone reports to hunt)* ![Animated: producer agents light up, correlation tokens pop in with their confidence scores, and the hunt correlator collects all eight verified edges](diagrams/atlas-a2a.gif) Thirteen agents were hired; six actually said something useful; one (the hunt correlator) listened to everyone and wrote it all down. Below: the full gossip graph from the report's ยง3 handoff log โ€” producers โ†’ correlation tokens โ†’ the hunt hub. Even the memory agent contributed *by skipping* (its skip notice became the `coverage` token โ€” still counts!). > **Why this representation:** The hunt correlator's 8 base-dc edges are pairwise "cross-source agreement" links keyed on a token, so the most faithful and catchable representation is a tripartite flowchart LR: producer agents (left, colored by category) converge on token nodes (middle, amber, with the real confidence value), which feed the single hunt hub (right). Modeling tokens as nodes instead of edge labels makes the 2-agent vs 3-agent agreements (scan, yara) visually obvious as fan-ins, and keeps the graph under 900px wide. A second, deliberately tiny notch diagram shows the degenerate case โ€” one lone 0.3 'found' edge correlated entirely from failure/negative-result messages โ€” which is itself the story. ### base-dc: agent-to-agent communication graph via the hunt correlator ![base-dc run: six producer agents converge on eight correlation tokens, which the hunt correlator seals as cross-source agreement edges (findings 14-21).](diagrams/base-dc-agent-correlation-graph.png) *base-dc run: six producer agents converge on eight correlation tokens, which the hunt correlator seals as cross-source agreement edges (findings 14-21).* **What to see:** Six of the 13 planned agents fed the correlator: blue disk agents (artifact, yara_hunt, null_session_baseline), purple ATT&CK detectors (t1059.001, t1546.008), and even the gray memory agent, whose skip notice still contributed the 'coverage' token. The two 3-agent agreements stand out as triple fan-ins: 'scan' (null_session_baseline + t1059.001 + yara_hunt) and 'yara' (artifact + t1546.008 + yara_hunt), both at confidence 0.4; the strongest edges are 'case' and 'evidence' at 0.6, the weakest 'host' at 0.3. The red box is honest about who said nothing: timeline timed out after 5452s and filesystem's fls after 60s, while discovery and mail produced no findings at all.
Mermaid source ```mermaid flowchart LR subgraph SRC["base-dc-cdrive.E01 โ€” 13-agent plan"] ART["artifact
chain: case 20180905-001"] YH["yara_hunt
Forge tag 20260505"] NSB["null_session_baseline
0 bursts / 81 windows"] T59["t1059.001 iex-loopback
0 hits / 999 EID 4104"] T46["t1546.008 IFEO hijack
LEG-A Thymus REJECT"] MEM["memory
skipped: not a RAM dump"] ERR["timed out, no token:
timeline (5452s)
filesystem fls (60s)"] QUIET["no token contributed:
injection_detector, t1071.001 (skipped)
discovery, mail (no findings)"] end subgraph TOK["correlation tokens (confidence)"] Kcase(["case (0.6)"]) Kevid(["evidence (0.6)"]) Kyara(["yara (0.4)"]) Kscan(["scan (0.4)"]) Kacross(["across (0.4)"]) Kcomp(["complete (0.4)"]) Kcov(["coverage (0.4)"]) Khost(["host (0.3)"]) end HUNT["hunt correlator
8 cross-source edges"] ART --> Kcase YH --> Kcase ART --> Kevid T46 --> Kevid ART --> Kyara T46 --> Kyara YH --> Kyara NSB --> Kscan T59 --> Kscan YH --> Kscan NSB --> Kacross T59 --> Kacross NSB --> Kcomp T59 --> Kcomp MEM --> Kcov YH --> Kcov T59 --> Khost T46 --> Khost Kcase --> HUNT Kevid --> HUNT Kyara --> HUNT Kscan --> HUNT Kacross --> HUNT Kcomp --> HUNT Kcov --> HUNT Khost --> HUNT classDef disk fill:#bbdefb,stroke:#1565c0,color:#1a1a1a classDef det fill:#e1bee7,stroke:#6a1b9a,color:#1a1a1a classDef skip fill:#e0e0e0,stroke:#616161,color:#1a1a1a classDef err fill:#ffcdd2,stroke:#b71c1c,color:#1a1a1a classDef tok fill:#ffe0b2,stroke:#e65100,color:#1a1a1a classDef hub fill:#ffe0b2,stroke:#e65100,stroke-width:3px,color:#1a1a1a class ART,YH,NSB disk class T59,T46 det class MEM,QUIET skip class ERR err class Kcase,Kevid,Kyara,Kscan,Kacross,Kcomp,Kcov,Khost tok class HUNT hub ```
### notch: the single degenerate correlation edge ![notch run: the same 13-agent plan yields exactly one correlation edge โ€” the token 'found' at confidence 0.3, agreed by three agents whose messages were all failures or negative results.](diagrams/notch-agent-correlation-graph.png) *notch run: the same 13-agent plan yields exactly one correlation edge โ€” the token 'found' at confidence 0.3, agreed by three agents whose messages were all failures or negative results.* **What to see:** On the raw notch image the communication graph collapses to one edge (finding 9): 'found' flagged by null_session_baseline, t1059.001 and yara_hunt at confidence 0.3 โ€” the lowest tier seen in either run. Note the source texts: two evtx_dump rc=1 failures and yara_hunt's 'found no scannable files', so the lone agreement is a token match across negative-result messages, not a detection. The gray box accounts for the other nine planned agents, all skipped, failed, or silent.
Mermaid source ```mermaid flowchart LR subgraph SRC["Challenge.raw (notch) โ€” 13-agent plan"] NSB["null_session_baseline
evtx_dump failed (rc=1)"] T59["t1059.001 iex-loopback
evtx_dump failed (rc=1)"] YH["yara_hunt
found no scannable files"] SKIP["no token contributed:
memory, injection_detector,
t1071.001 (not a RAM image)
t1546.008 (not an E01 image)
filesystem: fls failed (rc=1)
timeline, artifact, discovery,
mail (no findings)"] end K(["found (0.3)"]) HUNT["hunt correlator
1 cross-source edge"] NSB --> K T59 --> K YH --> K K --> HUNT classDef disk fill:#bbdefb,stroke:#1565c0,color:#1a1a1a classDef det fill:#e1bee7,stroke:#6a1b9a,color:#1a1a1a classDef skip fill:#e0e0e0,stroke:#616161,color:#1a1a1a classDef tok fill:#ffe0b2,stroke:#e65100,color:#1a1a1a classDef hub fill:#ffe0b2,stroke:#e65100,stroke-width:3px,color:#1a1a1a class NSB,YH disk class T59 det class SKIP skip class K tok class HUNT hub ```
--- ## 2 ยท โฑ๏ธ One run, two speeds *(geology vs. lightning)* ![Animated: the 91-minute log2timeline wait grows to scale, then the eight-token correlation burst machine-guns in over 217 microseconds](diagrams/atlas-clock.gif) This run spent **96.8% of its life waiting for one stubborn `log2timeline` call** โ€” and then did its smartest work in **217 microseconds**. That's a 25-billion-fold speed difference inside a single clock. The three figures below map ยง4 end to end: the macro chain, where the wall-clock went, and the microsecond zoom. > **Why this representation:** The topic is a causal chain of timestamped events spanning seven orders of magnitude (a 5,452-second timeout down to a 217-microsecond burst), so a vertical flowchart with elapsed-time edge labels is the only form that keeps the sequence readable while making the wild duration asymmetry visible through edge annotations and outcome colors. A companion pie quantifies the headline fact (96.8% of the wall clock was one blocked tool call), and a micro-zoom flowchart resolves the 217-microsecond correlation burst that is invisible at run scale โ€” together they tell the macro and micro story of the same clock. ### base-dc run โ€” end-to-end timestamp chain (14:22:29 โ†’ 15:56:32 UTC) ![The base-dc run's 1 h 34 m 02 s timestamp chain: every key tool-call event from the 14:22:29 memory baseline to the sealed report, colored by outcome with elapsed-time edges.](diagrams/base-dc-timestamp-chain.png) *The base-dc run's 1 h 34 m 02 s timestamp chain: every key tool-call event from the 14:22:29 memory baseline to the sealed report, colored by outcome with elapsed-time edges.* **What to see:** Notice the brutal asymmetry: the very first edge consumes 91 min 02 s (96.8% of the run) waiting for log2timeline to hit its 5452 s timeout (tc[2], exit 1), while everything after 15:54:31 โ€” including the 131-call extract_files storm (61 Thymus REJECTs + 70 rate-limit blocks) and the null_session_baseline running in parallel โ€” fits in the last 2 minutes. The amber agent.hunt node then emits 8 cross-source correlation findings in just 217 microseconds (15:56:32.095393 โ†’ .095610) before the run seals with 22 findings, 176 tool calls, and report_seal 151c9e88โ€ฆ
Mermaid source ```mermaid flowchart TD A["14:22:29.658 UTC
agent.memory baseline
0.11 ms - 1 finding"]:::ok B["15:53:31.501
get_timeline FAIL exit 1
log2timeline timed out after 5452 s"]:::bad C["15:54:31.620
fls FAIL
timed out after 60.0 s"]:::bad D["15:54:31.700
get_image_info OK
79.43 ms - E01 metadata"]:::ok E["15:54:31.702836
extract_files call 1
Thymus REJECT_OUTSIDE_ALLOWLIST"]:::rej F["15:54:35.582 - 15:55:33.829
extract_files storm - 131 calls
61 Thymus REJECT + 70 rate-limited"]:::rej G["15:55:32.320
null_session_baseline done
56.7 s - 1 finding"]:::ok H["15:55:33.811 - .829
yara_hunt + injection + T1546.008
5 findings - YARA mount skipped"]:::ok I["15:56:32.093
T1059.001 IEX + T1071.001 svchost
58.3 s - 3 findings"]:::ok J["15:56:32.095393 - .095610
agent.hunt correlation burst
8 findings in 217 microseconds"]:::hunt K["15:56:32.119754
final list_files sweep
tool call 176 of 176"]:::tool L["SEALED REPORT
22 findings - budget_exhausted - critic 1.0
report_seal 151c9e88 - audit seal 1085b493
146 audit entries - evidence SHA e2b9cf0c"]:::seal A -->|"91 min 02 s blocked - 96.8% of run"| B B -->|"+60.1 s"| C C -->|"+80 ms"| D D -->|"+2 ms"| E E -->|"storm spans 62.1 s"| F E -->|"parallel - 60.6 s later"| G F --> H G --> H H -->|"+58.3 s scan"| I I -->|"+2.3 ms"| J J -->|"+24 ms"| K K --> L classDef ok fill:#c8e6c9,stroke:#2e7d32,color:#1a1a1a classDef bad fill:#ffcdd2,stroke:#c62828,color:#1a1a1a classDef rej fill:#ffcdd2,stroke:#b71c1c,stroke-dasharray:4 3,color:#1a1a1a classDef tool fill:#bbdefb,stroke:#1565c0,color:#1a1a1a classDef hunt fill:#ffe0b2,stroke:#ef6c00,color:#1a1a1a classDef seal fill:#e1bee7,stroke:#6a1b9a,color:#1a1a1a ```
### Where the 5,642 seconds went *(redesigned โ€” the pie was hiding the story)* ![Proportional bar of the base-dc wall clock: log2timeline's 5,461.8 s wait fills 96.8% of the full-run bar, and a magnified second bar reveals the three ~60 s phases the pie hid โ€” the extract_files storm, the fls timeout, and the detector scans.](diagrams/base-dc-wallclock-pie.png) *Wall-clock budget of the 94-minute base-dc run, as a proportional bar with the 3.2% remainder magnified โ€” the earlier pie made everything except the timeout invisible, which is exactly why it got redesigned.* **What to see:** The top bar is the whole run: one blocked log2timeline call (5,461.8 s measured duration_ms on tc[2], against its 5452 s configured cap) is 96.8% of it. The magnified bar below is the surviving 3.2% (180.5 s), where the real texture lives: the 62.1 s extract_files storm (with its 61 Thymus REJECTs), the 60.1 s fls timeout, and the 58.3 s detector scans. Still invisible at any scale: the correlation burst, final sweep, and sealing โ€” under 30 ms combined.
Sources โ€” the redesigned bar (HTML) and the superseded pie (Mermaid) The current figure is rendered from [`diagrams/animated-decks/wallclock-bar.html`](diagrams/animated-decks/wallclock-bar.html). The original pie source remains at [`diagrams/base-dc-wallclock-pie.mmd`](diagrams/base-dc-wallclock-pie.mmd) โ€” kept for provenance; superseded 2026-06-12 because a 96.8% slice rendered every other slice unreadable.
### Micro-zoom โ€” the 217-microsecond agent.hunt correlation burst ![Inside the 217-microsecond window at 15:56:32.095 UTC: agent.hunt emits all 8 cross-source correlation findings, 22-55 microseconds apart, closing the CROSS_AGENT_CORRELATION_DONE proof.](diagrams/hunt-burst-217us-zoom.png) *Inside the 217-microsecond window at 15:56:32.095 UTC: agent.hunt emits all 8 cross-source correlation findings, 22-55 microseconds apart, closing the CROSS_AGENT_CORRELATION_DONE proof.* **What to see:** Each amber node is one real hunt finding from base-dc-report.json with its exact sub-millisecond timestamp: the burst opens with token 'case' at .095393 (artifact + yara_hunt agree) and closes with token 'host' at .095610 โ€” 217 microseconds total inside a 2.33 ms agent.hunt call (tc[151]). Inter-finding gaps shrink from 55 us down to 22 us, and two tokens ('scan', 'yara') reach 3-agent agreement, the strongest cross-source signal of the run.
Mermaid source ```mermaid flowchart TD P["tc 151 - agent.hunt
15:56:32.095698 UTC - 2.33 ms
8 cross-source findings"]:::tool T0["+0 us - .095393
token case
artifact + yara_hunt"]:::hunt T1["+55 us - .095448
token evidence
artifact + T1546.008"]:::hunt T2["+89 us - .095482
token across
null_session + T1059.001"]:::hunt T3["+118 us - .095511
token complete
null_session + T1059.001"]:::hunt T4["+142 us - .095535
token coverage
memory + yara_hunt"]:::hunt T5["+167 us - .095560
token scan
3 agents agree"]:::hunt T6["+195 us - .095588
token yara
3 agents agree"]:::hunt T7["+217 us - .095610
token host
T1059.001 + T1546.008"]:::hunt S["completion proof
CROSS_AGENT_CORRELATION_DONE"]:::seal P --> T0 T0 -->|"55 us"| T1 T1 -->|"34 us"| T2 T2 -->|"29 us"| T3 T3 -->|"24 us"| T4 T4 -->|"25 us"| T5 T5 -->|"28 us"| T6 T6 -->|"22 us"| T7 T7 --> S classDef tool fill:#bbdefb,stroke:#1565c0,color:#1a1a1a classDef hunt fill:#ffe0b2,stroke:#ef6c00,color:#1a1a1a classDef seal fill:#e1bee7,stroke:#6a1b9a,color:#1a1a1a ```
--- ## 3 ยท ๐Ÿ” How 13 agents became 2 *(self-correction, not layoffs)* ![Animated: the 13-agent roster appears, 11 fade to banked-green, the two stubborn gaps pulse red through four futile retries, and the run honestly declares budget_exhausted](diagrams/atlas-iter.gif) After one iteration the engine had already figured out who was useful: 11 agents delivered, got their findings **banked**, and were never re-run. The two that produced nothing (`discovery`, `mail`) got four more chances each โ€” and produced nothing, every time. The run then ends `budget_exhausted`: it doesn't pretend it finished, it just says it ran out. We respect that. > **Why this representation:** The self-correction story has two distinct shapes that need two distinct visuals: (1) a per-run vertical flowchart funnel that shows WHY the plan shrank โ€” the critic banks the stable agents' findings, drops them from the plan, and re-plans only the zero-finding gap agents, then loops on those gaps until the iteration budget runs out; and (2) a single xychart-beta comparing both runs' plan-size trajectory (13โ†’2 vs 13โ†’4), which makes the funnel shape and its iter-2 cliff instantly catchable. The flowcharts carry the real agent names and critic verdicts so the reader sees self-correction (not random shrinkage); the bar/line chart carries the cross-run comparison the flowcharts cannot. Vertical TD layouts keep both funnels at ~586px intrinsic width; all three sources were rendered with mmdc and verified. ### BASE-DC run โ€” plan shrinks 13 โ†’ 2 (gap-only re-planning) ![BASE-DC (base-dc-cdrive.E01): after iteration 1, the 11 finding-producing agents are dropped from the plan and only the two zero-finding gap agents (discovery, mail) are re-planned, four times, until budget_exhausted.](diagrams/base-dc-self-correction-funnel.png) *BASE-DC (base-dc-cdrive.E01): after iteration 1, the 11 finding-producing agents are dropped from the plan and only the two zero-finding gap agents (discovery, mail) are re-planned, four times, until budget_exhausted.* **What to see:** Iteration 1 plans the full 13-agent roster; 11 agents produce findings (green) and are immediately dropped from all later plans (grey) โ€” their results are banked, not re-computed. The critic (score 1.00, should_halt=false) re-plans only the 2 gap agents, discovery and mail, which return 0 findings on every one of the four retries (red chain). The run ends as budget_exhausted at 5/5 iterations with 22 sealed findings โ€” an honest negative: the gaps never close, the loop never pretends they did.
Mermaid source ```mermaid flowchart TD P1["ITER 1 - plan: 13 agents (full roster)"]:::plan P1 --> S1["11 STABLE - findings banked
memory, timeline, filesystem, artifact
null_session_baseline, yara_hunt
injection_detector, hunt, t1546_008
t1059_001, t1071_001"]:::stable P1 --> G1["2 GAPS - 0 findings
discovery, mail"]:::gap S1 --> DR["DROPPED from plan, iters 2 to 5
results kept, never re-run"]:::dropped G1 --> C1["Critic 1.00, should_halt false
continue: plan gap, re-plan gaps only"]:::critic C1 --> P2["ITER 2 - plan: 2
discovery, mail"]:::plan P2 --> G2["0 findings again"]:::gap G2 --> P3["ITER 3 - plan: 2"]:::plan P3 --> G3["0 findings"]:::gap G3 --> P4["ITER 4 - plan: 2"]:::plan P4 --> G4["0 findings"]:::gap G4 --> P5["ITER 5 - plan: 2"]:::plan P5 --> G5["0 findings"]:::gap G5 --> FIN["STATUS: budget_exhausted
5 of 5 iterations, 22 findings sealed"]:::final classDef plan fill:#bbdefb,stroke:#1565c0,color:#1a1a1a classDef stable fill:#c8e6c9,stroke:#2e7d32,color:#1a1a1a classDef gap fill:#ffcdd2,stroke:#c62828,color:#1a1a1a classDef dropped fill:#e0e0e0,stroke:#757575,color:#1a1a1a classDef critic fill:#e1bee7,stroke:#6a1b9a,color:#1a1a1a classDef final fill:#ffe0b2,stroke:#e65100,color:#1a1a1a ```
### NOTCH run โ€” plan shrinks 13 โ†’ 4 (gap-only re-planning) ![NOTCH (Challenge.raw): the same 13-agent roster yields only 9 stable agents โ€” timeline and artifact join discovery and mail as gaps โ€” so iterations 2-5 re-plan 4 agents instead of 2, again ending budget_exhausted.](diagrams/notch-self-correction-funnel.png) *NOTCH (Challenge.raw): the same 13-agent roster yields only 9 stable agents โ€” timeline and artifact join discovery and mail as gaps โ€” so iterations 2-5 re-plan 4 agents instead of 2, again ending budget_exhausted.* **What to see:** Identical self-correction mechanics as BASE-DC but a different gap set: on this raw memory image, timeline and artifact also produce 0 findings, so the re-planned set is 4 agents (timeline, artifact, discovery, mail) rather than 2. The 9 stable agents (green) are dropped after iteration 1 and the run seals 10 findings at budget_exhausted. Comparing the two funnels shows the planner adapts per-evidence: the shrink target is data-driven (13 to 2 vs 13 to 4), not hard-coded.
Mermaid source ```mermaid flowchart TD P1["ITER 1 - plan: 13 agents (same roster)"]:::plan P1 --> S1["9 STABLE - findings banked
memory, filesystem, null_session_baseline
yara_hunt, injection_detector, hunt
t1546_008, t1059_001, t1071_001"]:::stable P1 --> G1["4 GAPS - 0 findings
timeline, artifact, discovery, mail"]:::gap S1 --> DR["DROPPED from plan, iters 2 to 5
results kept, never re-run"]:::dropped G1 --> C1["Critic 1.00, should_halt false
continue: plan gap, re-plan gaps only"]:::critic C1 --> P2["ITER 2 - plan: 4
timeline, artifact, discovery, mail"]:::plan P2 --> G2["0 findings again"]:::gap G2 --> P3["ITER 3 - plan: 4"]:::plan P3 --> G3["0 findings"]:::gap G3 --> P4["ITER 4 - plan: 4"]:::plan P4 --> G4["0 findings"]:::gap G4 --> P5["ITER 5 - plan: 4"]:::plan P5 --> G5["0 findings"]:::gap G5 --> FIN["STATUS: budget_exhausted
5 of 5 iterations, 10 findings sealed"]:::final classDef plan fill:#bbdefb,stroke:#1565c0,color:#1a1a1a classDef stable fill:#c8e6c9,stroke:#2e7d32,color:#1a1a1a classDef gap fill:#ffcdd2,stroke:#c62828,color:#1a1a1a classDef dropped fill:#e0e0e0,stroke:#757575,color:#1a1a1a classDef critic fill:#e1bee7,stroke:#6a1b9a,color:#1a1a1a classDef final fill:#ffe0b2,stroke:#e65100,color:#1a1a1a ```
### Plan size per iteration โ€” BASE-DC vs NOTCH ![Agents in the plan at each of the 5 iterations: BASE-DC (blue bars) collapses 13 to 2 after iteration 1; NOTCH (orange line) collapses 13 to 4 โ€” both then hold flat until budget_exhausted.](diagrams/plan-size-per-iteration-both-runs.png) *Agents in the plan at each of the 5 iterations: BASE-DC (blue bars) collapses 13 to 2 after iteration 1; NOTCH (orange line) collapses 13 to 4 โ€” both then hold flat until budget_exhausted.* **What to see:** Both runs start with the full 13-agent plan and collapse after a single iteration โ€” BASE-DC to 2 agents (blue bars: 13, 2, 2, 2, 2), NOTCH to 4 (orange line: 13, 4, 4, 4, 4). The one-iteration cliff is the self-correction signal: ~85% (11/13) and ~69% (9/13) of the roster proved stable on the first pass and was never re-run. The flat tail at 2 and 4 shows the critic (score 1.00, should_halt=false every iteration) kept retrying only the persistent zero-finding gaps until the 5-iteration budget was exhausted.
Mermaid source ```mermaid --- config: themeVariables: xyChart: backgroundColor: "#ffffff" titleColor: "#1a1a1a" xAxisLabelColor: "#1a1a1a" xAxisTitleColor: "#1a1a1a" xAxisLineColor: "#1a1a1a" xAxisTickColor: "#1a1a1a" yAxisLabelColor: "#1a1a1a" yAxisTitleColor: "#1a1a1a" yAxisLineColor: "#1a1a1a" yAxisTickColor: "#1a1a1a" plotColorPalette: "#1565c0, #e65100" --- xychart-beta title "Plan size per iteration - bar: BASE-DC (13 to 2), line: NOTCH (13 to 4)" x-axis ["iter 1", "iter 2", "iter 3", "iter 4", "iter 5"] y-axis "agents in plan" 0 --> 14 bar [13, 2, 2, 2, 2] line [13, 4, 4, 4, 4] ```
--- ## 4 ยท ๐Ÿšช The bouncer & the wax seal *(governance, made visible)* ![Animated: the ALLOW counter rolls to 85 and the REJECT counter to 61 โ€” each refusal issued before the process spawned โ€” while the one 32-byte key cross-binds both sealed files](diagrams/atlas-gov.gif) The Thymus is the bouncer: it checked **every** file access, let 85 in, and turned **61** away โ€” each one refused *before the process even started*, and every refusal is on the sealed record. The wax seal is the cross-bind: one 32-byte key seals both the report and its audit log, so editing either file after the fact is instantly self-incriminating. Figures below map ยง6. > **Why this representation:** Governance has two distinct stories that need two distinct visual grammars: (1) the boundary's behavior is a simple categorical split, best told by ALLOW/REJECT pies โ€” one per case, because the 42% REJECT rate on base-dc versus 0% on notch is itself a finding (the gate fires when workloads stray, stays silent when they don't); (2) the sealing is a dependency chain, so a vertical flowchart shows how one 32-byte key cross-binds report.json and audit-log.json while the evidence sha256 anchors everything, with the REJECT path in red to show denials are inside (not outside) the sealed record. A consistent palette (green=ALLOW, red=REJECT, purple=seals/governance, blue=files, amber=hash anchor) ties the three figures together. All three diagrams were rendered with mermaid-cli and visually verified; every number was extracted from the raw jsonl/json with jq. ### Thymus policy gate โ€” base-dc case (146 decisions) ![Thymus runtime boundary on the base-dc disk case: 85 of 146 file-access decisions allowed, 61 rejected.](diagrams/thymus-gate-base-dc-pie.png) *Thymus runtime boundary on the base-dc disk case: 85 of 146 file-access decisions allowed, 61 rejected.* **What to see:** The gate is not decorative โ€” 61 of 146 decisions (42%) were REJECTed, every one with reason REJECT_OUTSIDE_ALLOWLIST against /tmp/claude-1001 scratch paths (59x a tasks dir, 1x an extract dir, 1x a w204 dir). All 85 ALLOWs share a single reason, 'within read-only zone', covering only /cases/SRL-2018 (20) and the base-dc-cdrive.E01 image itself (65). Agents never touched anything outside the evidence allowlist, and every denial is on the sealed record.
Mermaid source ```mermaid %%{init: {"theme":"base","themeVariables":{"pie1":"#c8e6c9","pie2":"#ffcdd2","pieTitleTextColor":"#1a1a1a","pieLegendTextColor":"#1a1a1a","pieSectionTextColor":"#1a1a1a","pieStrokeColor":"#1a1a1a"}}}%% pie showData title base-dc Thymus gate - 146 decisions "ALLOW - within read-only zone" : 85 "REJECT - outside allowlist" : 61 ```
### Thymus policy gate โ€” notch case (26 decisions) ![The same Thymus boundary on the notch raw-image case: all 26 decisions ALLOW, zero rejections.](diagrams/thymus-gate-notch-pie.png) *The same Thymus boundary on the notch raw-image case: all 26 decisions ALLOW, zero rejections.* **What to see:** All 26 decisions are ALLOW 'within read-only zone' (20x the /cases/Challenge_NotchItUp directory, 6x Challenge.raw) โ€” the REJECT slice is deliberately kept in the legend at 0 to show the gate was armed but never tripped. Contrast with base-dc's 61 REJECTs: the difference is workload shape (the disk case spawned /tmp scratch traffic the policy refused), not a relaxed policy.
Mermaid source ```mermaid %%{init: {"theme":"base","themeVariables":{"pie1":"#c8e6c9","pie2":"#ffcdd2","pieTitleTextColor":"#1a1a1a","pieLegendTextColor":"#1a1a1a","pieSectionTextColor":"#1a1a1a","pieStrokeColor":"#1a1a1a"}}}%% pie showData title notch Thymus gate - 26 decisions "ALLOW - within read-only zone" : 26 "REJECT" : 0 ```
### Tamper-evidence cross-binding โ€” evidence anchor, HMAC seals, and the Thymus REJECT path (base-dc) ![How one 32-byte session key cross-binds the findings report and the Thymus audit log into a single tamper-evident unit, anchored to the evidence image's sha256.](diagrams/seal-cross-binding-chain.png) *How one 32-byte session key cross-binds the findings report and the Thymus audit log into a single tamper-evident unit, anchored to the evidence image's sha256.* **What to see:** Follow the purple governance lane: the same 32-byte HMAC session key produces both report_seal (151c9e88...e453b8c1) over base-dc-report.json and audit_log_seal (1085b493...6b5927) over the 146-entry audit log โ€” and that audit_log_seal value is embedded byte-identically inside report.json, so editing either file breaks the cross-bind. The red edge is the boundary in action: 61 REJECTs of /tmp scratch paths are themselves part of the sealed record, while the green ALLOW path leads only to the evidence image, whose sha256 (e2b9cf0c...965b31b1) anchors the whole chain to the physical evidence. The audit log's entry_count 146 matching its 146 actual entries is the third independent consistency check.
Mermaid source ```mermaid flowchart TD REQ["Swarm agents
file-access requests"] --> GATE{"Thymus policy gate
base-dc: 146 decisions"} GATE -->|"85 ALLOW
within read-only zone"| EV["Evidence image
base-dc-cdrive.E01"] GATE -->|"61 REJECT
REJECT_OUTSIDE_ALLOWLIST"| REJ["Blocked scratch paths
59x tasks dir, 1x extract, 1x w204
all under tmp claude-1001"] GATE -. "every decision appended" .-> AUD["base-dc-report.audit-log.json
146 entries = entry_count 146"] EV --> SHA["sha256 anchor
e2b9cf0c...965b31b1"] SHA --> RPT["base-dc-report.json
evidence_image_sha256 field"] KEY["per-case session key
32 bytes - HMAC
bytes never printed"] RPT --> RSEAL["report_seal
151c9e88...e453b8c1"] AUD --> ASEAL["audit_log_seal
1085b493...6b5927"] KEY --> RSEAL KEY --> ASEAL ASEAL --> BIND["SEALS-CROSS-BIND
audit_log_seal identical in
report.json and audit-log.json"] RPT --> BIND classDef gov fill:#e1bee7,stroke:#6a1b9a,color:#1a1a1a classDef ok fill:#c8e6c9,stroke:#2e7d32,color:#1a1a1a classDef bad fill:#ffcdd2,stroke:#b71c1c,color:#1a1a1a classDef file fill:#bbdefb,stroke:#1565c0,color:#1a1a1a classDef anchor fill:#ffe0b2,stroke:#e65100,color:#1a1a1a class GATE,KEY,RSEAL,ASEAL,BIND gov class EV ok class REJ bad class REQ,RPT,AUD file class SHA anchor linkStyle 2 stroke:#b71c1c,stroke-width:2px ```
--- ## 5 ยท ๐Ÿ“Š The numbers at a glance *(hunt carried the team ๐Ÿ†)* ![Animated: findings-per-agent bars race out โ€” hunt 8, the mid-tier detectors 2 each โ€” then the base-dc vs notch comparison lands the punchline: 94 minutes vs 72 seconds](diagrams/atlas-metrics.gif) Eight of base-dc's 22 findings came from one agent โ€” the hunt correlator, which makes its living combining everyone else's work (a lesson in there somewhere). And the run comparison: 22 findings in 94 minutes vs 10 findings in 72 seconds. One of them had a timeout problem. > **Why this representation:** The topic is "run metrics at a glance," which is purely quantitative โ€” bar charts (xychart-beta) are the only representation that makes magnitudes instantly comparable. Chart 1 ranks the 11 base-dc agents by findings so the hunt agent's dominance (8 of 22) pops immediately; Chart 2 puts the two runs side-by-side on all three headline metrics (calls, findings, runtime) using alternating blue/amber bars via a zero-placeholder two-series technique, so the 176-vs-60 and 94-min-vs-1.2-min contrasts are caught in one look. Every number was extracted from base-dc-report.json and notch-report.json with python3 (findings[] grouped by agent field; trace.tool_calls length; trace.total_duration_ms), and both Mermaid sources were render-verified with mermaid-cli to PNG. ### base-dc: Findings per Agent ![Findings per agent in the base-dc run (22 total, sorted descending) โ€” the correlation hunt agent alone contributed 8.](diagrams/base-dc-findings-per-agent.png) *Findings per agent in the base-dc run (22 total, sorted descending) โ€” the correlation hunt agent alone contributed 8.* **What to see:** The hunt (correlation) agent dominates with 8 of the 22 findings โ€” 4x any single detector โ€” showing most value came from cross-agent correlation rather than any one detector. The mid tier (artifact, yara, t1546 IFEO hijack, t1059 IEX loopback C2) contributed 2 each; six agents contributed 1 each. Note the memory bar is a skip marker, not a detection: the MemoryAgent recorded an honest 'disk-only image, Volatility requires RAM' skip with confidence 0.0.
Mermaid source ```mermaid %%{init: {"xyChart": {"width": 880, "height": 430}, "themeVariables": {"xyChart": {"plotColorPalette": "#1565c0", "titleColor": "#1a1a1a", "xAxisLabelColor": "#1a1a1a", "xAxisTitleColor": "#1a1a1a", "xAxisLineColor": "#1a1a1a", "yAxisLabelColor": "#1a1a1a", "yAxisTitleColor": "#1a1a1a", "yAxisLineColor": "#1a1a1a"}}}}%% xychart-beta title "base-dc run: 22 findings across 11 agents (94 min)" x-axis ["hunt", "artifact", "yara", "t1546", "t1059", "timeline", "filesys", "null-sess", "inject", "t1071", "memory"] y-axis "findings" 0 --> 9 bar [8, 2, 2, 2, 2, 1, 1, 1, 1, 1, 1] ```
### Run Comparison: base-dc vs notch ![Headline metrics of the two SRL-2018 submission runs side by side: tool calls, findings, and wall-clock runtime.](diagrams/base-dc-vs-notch-run-comparison.png) *Headline metrics of the two SRL-2018 submission runs side by side: tool calls, findings, and wall-clock runtime.* **What to see:** base-dc (blue) used 176 tool calls to seal 22 findings in 94.0 min; notch (amber) used 60 calls for 10 findings in just 72.4 seconds (1.2 min). The near-invisible 'min notch' bar IS the story: the 78x runtime gap is almost entirely one tool โ€” base-dc's get_timeline call ran 5,461.8 s of the 5,642 s total (log2timeline timed out at 5452 s on the E01), while everything else on both runs completed in seconds. Both runs ended with status budget_exhausted after 5 iterations and a critic_score of 1.0.
Mermaid source ```mermaid %%{init: {"xyChart": {"width": 860, "height": 430}, "themeVariables": {"xyChart": {"plotColorPalette": "#1565c0, #ef6c00", "titleColor": "#1a1a1a", "xAxisLabelColor": "#1a1a1a", "xAxisTitleColor": "#1a1a1a", "xAxisLineColor": "#1a1a1a", "yAxisLabelColor": "#1a1a1a", "yAxisTitleColor": "#1a1a1a", "yAxisLineColor": "#1a1a1a"}}}}%% xychart-beta title "base-dc (blue) vs notch (amber): 176 vs 60 calls, 22 vs 10 findings, 94 vs 1.2 min" x-axis ["calls base-dc", "calls notch", "findings base-dc", "findings notch", "min base-dc", "min notch"] y-axis "count or minutes" 0 --> 180 bar [176, 0, 22, 0, 94, 0] bar [0, 60, 0, 10, 0, 1.2] ```
--- *The section animations above are Animotion-built, deterministically rendered (virtual-time), and loop forever โ€” their deck sources live in [`diagrams/animated-decks/`](diagrams/animated-decks/). Built from the sealed records in this folder; cross-checked claims live in the [gold report](AGENT-EXECUTION-LOGS-REPORT.md) (ยง3โ€“ยง6) and are independently re-verifiable via its "verify in 60 seconds" block. Folder guide: [README.md](README.md).*