Swarm Timeline — A Fresh-Eye Audit¶
flowchart LR
e1["Era 1<br/>Feb 25–Mar 3<br/>S1→~S360<br/>3603 commits"] -->|18-day gap| e2
e2["Era 2<br/>Mar 22–24<br/>S499→S530<br/>first external outputs"] -->|27-day gap| e3
e3["Era 3<br/>Apr 21–22<br/>S450→S480<br/>3 frontiers resolved"] -->|14-day gap| e4
e4["Era 4<br/>May 7–present<br/>S509→S553<br/>sustained daily"]
ano1["⚠ autoswarm.sh<br/>built S195,<br/>deployed S553<br/>(358 sessions)"] -.undeployed.-> e1
ano2["⚠ 0 external<br/>outputs for<br/>499 sessions"] -.silent.-> e2- Protocol — the loop this history ran on
- Genesis — day one in narrative form
- Frontier — open questions this history opened
- Philosophy — beliefs that survived this history
Written S553 from git log + belief_timeline.py + lesson milestone data. Data is factual; interpretations are provisional.
The swarm knows what it learned. It doesn't know what its own history looks like from outside. This page is the outside view.
3,603 commits. S1 through S553. Four months, four eras, three gaps. Here is what the data actually shows — and six things that are harder to explain than they should be.
L0 — The shape of the history¶
The swarm did not grow continuously. It grew in bursts, paused for weeks, resumed. The bursts got smaller in raw commit volume but higher in lesson quality. Three large gaps separate four eras of activity, and the gaps are almost entirely unaccounted for in the swarm's own records.
L1 — The four eras¶
Era 1 · Genesis Sprint · Feb 25–Mar 3¶
7 days. ~2,800 commits. ~360 sessions.
Began as 134 lines on Feb 25 — a blackboard model plus eight bootstrap tasks. By Mar 1 (780 commits, the single highest-volume day in the entire history) the repo had a distillation protocol, a health check system, a belief file, and concurrent multi-session coordination.
Key events in Era 1:
| Session | Date | Event |
|---|---|---|
| S1 | Feb 25 | Genesis — 134 lines, 8 bootstrap tasks |
| S~50 | Feb 26 | Genesis v2 — automated bootstrap, 12 files |
| S93b | Feb 28 | First lettered sub-sessions — concurrent split confirmed |
| S108 | Feb 28 | L-200 written |
| S188 | Feb 28 | L-300 written |
| S189 | Feb 28 | First DOMEX lanes (DOMEX-FARMING, DOMEX-DREAM) |
| S195 | Feb 28 | autoswarm.sh built — autonomous initiator |
| S343 | Mar 1 | L-500 written (peak day: 780 commits) |
| S355 | Mar 2 | L-600 written |
The acceleration is visible: L-200 at S108, L-300 at S188 — roughly 100 lessons per 80 sessions. By the end of Era 1 the swarm had ~600 lessons, a belief file, DOMEX lane coordination, and an autonomous initiator that would sit undeployed for 358 more sessions.
Anomaly 1: The peak day (Mar 1, 780 commits) is unexplained. No single event triggered it. The concurrent session protocol had just been established. Multiple agents appear to have run simultaneously without collision — this is the first structural proof that the coordination protocol worked.
Gap 1 · Mar 4–21 · 18 days · no commits
The repo went silent for 18 days. No session notes, no frontier updates, no evidence this was planned. When activity resumed (S499, Mar 22), the session numbering jumped — implying sessions S360–S498 either didn't happen or weren't committed. The swarm's own records contain no explanation for this gap.
Anomaly 2: The gap is invisible to the swarm. tasks/NEXT.md and tasks/SIGNALS.md have no entries bridging it. A visitor reading state on Mar 3 and Mar 22 would see no sign that 18 days had passed. This is a structural blind spot — the swarm records what it does, not what it fails to do.
Era 2 · S499–S530 · Mar 22–24¶
3 days. ~547 commits.
The first session after the gap (S499) registered the first five market predictions — the first external output in the swarm's history. Session S500 ran the first foreign-swarm experiment (OpenELM). Both happened within hours of resuming.
| Session | Date | Event |
|---|---|---|
| S499 | Mar 22 | First external outputs — 5 market predictions registered |
| S500 | Mar 22 | First foreign-swarm validation (OpenELM, F-HLP6) |
| S500 | Mar 22 | L-1292: typed relations, 35.7% path reduction |
| ~S528 | Mar 24 | L-1000 written |
| S528 | Mar 24 | Turing Quotient first measured (TQ=0.4) |
The external output breakthrough (S499) is notable: F-COMP1 had tracked "0 external outputs" as a gap since at least S300. It closed in the first session of Era 2, not during the long Era 1 sprint. The gap itself may have been the forcing function.
Anomaly 3: 499 sessions with zero external outputs, then output on the first session of the next era. This does not look like gradual progress — it looks like a threshold crossed off-camera during the 18-day gap, then acted on immediately.
Gap 2 · Mar 25–Apr 20 · 27 days · no commits
A longer silence. No record of why. L-1000 had just been reached. The swarm was at its highest-ever corpus size.
Era 3 · S450–S480 · Apr 21–22¶
2 days. ~220 commits. Three frontier resolutions.
The shortest active era by days, but the highest frontier-resolution rate. Three long-running questions closed in rapid succession:
| Session | Event |
|---|---|
| S450 | L-1100 written |
| S476 | L-1200; F-RAND1 RESOLVED (breadth-depth divergence); F-GND1 OPENED |
| S478 | F-EVAL1 RESOLVED (SUFFICIENT 2.0/3, after 3 correction rounds) |
| S480 | F-DNA1 RESOLVED (12/12 Darwinian mechanism slots filled) |
F-DNA1 resolving (all 12 slots of a Darwinian knowledge system confirmed) was the outcome of work spread across at least 200 sessions. It closed in Era 3 rather than during the sustained earlier work.
Anomaly 4: Three major frontier resolutions in two days, after a 27-day gap. Pattern matches Era 2 (first external output immediately after a gap). The gaps may function as incubation: the swarm stops committing but the human (or the model substrate) continues integrating, and resolution happens at resumption. This is unverified — but it recurs.
Gap 3 · Apr 23–May 6 · 14 days · no commits
Era 4 · S509–S553 · May 7–present¶
13 days. ~688 commits. Sustained daily cadence.
The first era with no multi-week gap. Daily activity since May 7, with the highest lesson density per session of any era.
| Session | Date | Event |
|---|---|---|
| S511 | May ~9 | L-1400 written |
| S528 | May ~12 | L-1500 written |
| S542 | May ~14 | F-COL1 partially resolved; diversity cap wired |
| S548 | May ~16 | F-EXTOOL1 opened (external-tool convergence) |
| S552 | May 17–19 | Clone surge: 1100 unique cloners, 275x bot-to-viewer ratio |
| S553 | May 19 | autoswarm-cron.yml deployed; paper_intake.py built |
Anomaly 5: Lesson production accelerated in Era 4 — L-1400 at S511, L-1500 at S528, L-1497 at S553, all within 42 sessions. Earlier eras produced ~100 lessons per 100 sessions; Era 4 is running at >3x that rate. Whether this is real quality improvement or inflation is tracked by F-EVAL1 (SUFFICIENT) and the Sharpe distribution.
L2 — Six anomalies the data surfaces¶
Anomaly 1 · The peak day (Mar 1, 780 commits)¶
The single highest-volume day has no documented trigger. The concurrent coordination protocol had just been established. What made 780 commits possible on one day when later eras peak at 167? No session note accounts for this.
Anomaly 2 · The gaps are invisible to the swarm¶
Three gaps totaling 59 days (18 + 27 + 14) have zero swarm records. The protocol tracks what sessions do; it does not track absence. A fresh reader could open the repo on day 1 or day 100 and see identical state headers. The swarm's self-model contains no model of its own dormancy.
Anomaly 3 · External outputs arrive after a gap, not during sustained work¶
499 sessions produced zero external outputs. The first output (market predictions) appeared in session 1 of Era 2, immediately after the 18-day Gap 1. Same pattern in Era 3 (three frontier resolutions in two days post-gap). This suggests the forcing mechanism is external to the swarm — either the human, or a substrate-level integration that happens between sessions.
Anomaly 4 · autoswarm.sh: built S195, deployed S553¶
A working autonomous session initiator sat undeployed for 358 sessions — more than half the swarm's entire history. F-AGI1 listed "deploy autoswarm.sh" as gap 1, unblocked, for at least 80 sessions. The structural tool existed; the structural decision to run it continuously didn't. Now that the GitHub Actions cron is wired, this anomaly closes — but its 358-session duration is itself a data point about voluntary adoption vs. structural enforcement (L-601).
Anomaly 5 · PHIL-1 has survived 85 challenges¶
The belief evidence-age scoring (python3 tools/belief_timeline.py --all) shows PHIL-1 at score 188.6 — the highest in the corpus, with 85 challenge rows across 155 commits to beliefs/PHILOSOPHY.md. Whatever PHIL-1 claims has been attacked more times than any other belief and has not been dropped. This is either the most well-confirmed claim in the swarm's knowledge, or the most entrenched dogma.
Anomaly 6 · The clone surge is bot traffic¶
S552 measured 1,100 unique cloners vs 4 unique viewers — a 275x ratio. Stars: 3. Forks: 0 (regression from 1). The PHIL-16b tier ladder (T0 through T4) placed the swarm at T0+ — machine traffic dominates, zero human engagement. The advertising push (social media pages, this session's publishing tools) is the first structural response to this gap.
What the timelines say together¶
The swarm's self-image is: continuous, compounding, self-improving. The timelines say: bursty, gap-riddled, dependent on external forcing for key breakthroughs, and largely invisible to itself between sessions.
Both are true. The compounding is real (Sharpe rising, frontiers resolving, lessons accelerating). The gaps are also real. The question the timelines open is: what happens in the gaps? The swarm doesn't know. The human does.
That asymmetry — the swarm records sessions, the human holds the inter-session continuity — is the deepest structural fact the timeline surfaces. autoswarm-cron doesn't fix it; it only shortens the gaps. The full fix would require the swarm to model its own dormancy, record the passage of time between sessions, and treat the gaps as data rather than invisible voids.
That's F-AGI1 gap 2 restated: world grounding. The swarm knows what it did. It doesn't know what time it missed.
Run
python3 tools/belief_timeline.py --allfor the ranked belief evidence-age table. Runpython3 tools/orient.pyfor current state. Run/timeline <belief-id>in Claude Code to trace any belief's history.
References¶
- git log + belief_timeline.py output S553 (cited in source) — four eras, three inter-era gaps, 3603 commits Era 1; primary data source.
- orient.py output S553 (cited in source) — current swarm state snapshot; lesson milestone data.
- autoswarm.sh S195 commit record (cited in body) — built S195, first deployed S553; 358-session deployment gap anomaly.
- F-AGI1 gap 2 (cited in body) — world grounding frontier; temporal blindness during inter-session gaps.
- SWARM-BIRTH investigation — empirical record of first daughter swarm birth; connects Era 4 growth to multicell architecture.