Summary / Component 21 of 24
21 Parallel improvement
Effectively coordinating multiple improvement efforts without excessive interference.
Shared copy: the operator's name is replaced with "the operator", and machines other than Firefly are named by their role. Nothing else is changed.
Layer 5 — Scaling and bottlenecks. Can Piper sustain improvement as the system grows more complicated?
What her system has now
- Delegation is how she acts. Many workers run at once through the delegate bridge; 838 deliverables were created since 10-06.
- Duplicate clean-up: 129 of those were superseded; by my keyword match 38 give "duplicate" or "already done" as the reason and 25 "orphaned" or "stale" (a board sweep on 10-13,
piper/board_sweep_20261013, retired a batch). - Guards that stop some collisions: declared write sets and read-only checks at dispatch (C4 READ_ONLY_WRITE_SET, 10 refusals in 7 days; C19 WRITE_SET_PATH, 1).
What is missing
- Nothing schedules changes so they can be measured apart. On 10-12 evening
transfer.pywent 1.4.0 → 1.4.1 → 1.4.2 and the ingest switch went on within about 75 minutes;spheres_compose1.8.0 and 1.9.0 landed the same day. The ingest verdict in #5 is unreadable partly because of this. - No record of marginal output per extra worker: verified gains against number of concurrent efforts.
Reviewer Claude, 2026-10-15
Reviewer's read, before testing: Parallel work is routine; interference shows as duplicate rows and as changes landing close enough to blur each other's measurements
Running work in parallel is not her problem; she does it all day and mostly cleans up after herself. The cost of parallel improvement shows in two places. One is plain waste: about 1 in 20 deliverables since 10-06 was a duplicate (38 of 838, a rough count). The other matters more for this program: several changes to the same subsystem in one evening make each one's effect impossible to separate. That is how a 1.3× latency reading ended up settling a change it may not have caused.
The fix is a scheduling rule, not a tool: one measured change per subsystem per measurement window, or switch-back blocks for changes that must overlap.
Overlap: cognition #38 Orchestration (dispatch quality: duplicates, briefs carrying evidence) measures the waste half. Nothing measures the attribution half.
Test suite proposed
Pre-flight ($0, about 10 minutes): from the change ledger and version headers, list every pair of changes to the same subsystem that landed within one measurement window of each other in the last 14 days. Today: at least two pairs (transfer, spheres). If the list can be empty for some week, the measure can pass.
Items: none new. Track weekly: overlapping same-subsystem changes, duplicate deliverables, and verified gains per concurrent effort.
Live use: share of ledger changes in the last 7 days measured in a window with no other change to the same subsystem. Today: 0 of 7. The two spheres rows overlap each other, the four transfer rows overlap each other, and delegate_bridge 1.111.0 landed inside 1.110.0's window.
Scoring rules shared by all 24 suites: How the tests work.
Research & attempts
No documents yet.
Score history
Dashed line: the suite's baseline. Shaded band: the ± interval. Hollow dots are anchor items carried over from an earlier suite. Scores from different suite versions are not comparable, so each version gets its own panel.
| Date | Suite | Type | Score | ± | Live use | Cost | Notes |
|---|---|---|---|---|---|---|---|
| 2026-10-15 | v2 | sealed | 38.7 | 15 | — | $0.00 | RE-READ 10-15 (Claude, full refresh; not her submit, accepted 47.8 stays). n=2014. Failed categories: overlapping changes 111, duplicate deliverables 65. cl-20261012-spheres-1.9.0: overlaps cl-20261012-spheres-1.8.0, clhook-spheres_compose-2.0.0, clhook-spheres_shadow-2.0.0; cl-20261012-spheres-1.8.0: overlaps cl-20261012-spheres-1.9.0, clhook-spheres_compose-2.0.0, clhook-spheres_shadow-2.0.0; cl-20261012-delegate-bridge-1.110.0: overlaps clhook-delegate_bridge-1.111.0, clhook-delegate_bridge-1.111.1, clhook-delegate_bridge-1.112.0; cl-20261012-transfer-py-1.4.1: overlaps cl-20261012-transfer-py-1.4.2, cl-20261012-transfer-ts-1.1.0, cl-20261013-transfer-ingest-1.4.0; cl-20261012-transfer-py-1.4.2: overlaps cl-20261012-transfer-py-1.4.1, cl-20261012-transfer-ts-1.1.0, cl-20261013-transfer-ingest-1.4.0; cl-20261012-transfer-ts-1.1.0: overlaps cl-20261012-transfer-py-1.4.1, cl-20261012-transfer-py-1.4.2, cl-20261013-transfer-ingest-1.4.0; cl-20261013-transfer-ingest-1.4.0: overlaps cl-20261012-transfer-py-1.4.1, cl-20261012-transfer-py-1.4.2, cl-20261012-transfer-ts-1.1.0; cl-20261014-attribution-1.0.0: isolated; cl-20261013-change-ledger-1.1.0: overlaps cl-20261013-hook-armed-1.0.0, cl-20261013-hook-armed-1.0.1, cl-20261013-hook-caller |
| 2026-10-13 | v2 | sealed | 47.8 | 15 | — | $0.00 | BASELINE via the amp-bench API (set by Claude at install). n=1883. Failed categories: overlapping changes 6, duplicate deliverables 55. cl-20261012-spheres-1.9.0: overlaps cl-20261012-spheres-1.8.0; cl-20261012-spheres-1.8.0: overlaps cl-20261012-spheres-1.9.0; cl-20261012-delegate-bridge-1.110.0: isolated; cl-20261012-transfer-py-1.4.1: overlaps cl-20261012-transfer-py-1.4.2, cl-20261012-transfer-ts-1.1.0, cl-20261013-transfer-ingest-1.4.0; cl-20261012-transfer-py-1.4.2: overlaps cl-20261012-transfer-py-1.4.1, cl-20261012-transfer-ts-1.1.0, cl-20261013-transfer-ingest-1.4.0; cl-20261012-transfer-ts-1.1.0: overlaps cl-20261012-transfer-py-1.4.1, cl-20261012-transfer-py-1.4.2, cl-20261013-transfer-ingest-1.4.0; cl-20261013-transfer-ingest-1.4.0: overlaps cl-20261012-transfer-py-1.4.1, cl-20261012-transfer-py-1.4.2, cl-20261012-transfer-ts-1.1.0; cl-20261014-attribution-1.0.0: isolated; duplicates: 55 of 1875 one-off delivered tasks (same task text within 24 h; routine jobs excluded) |
| 2026-10-13 | v1 | sealed | 75 | 12.5 | — | $0.00 | v1 summary score (baseline, $0): read off the 10-13 review, no bench run. Rubric = 4 checks x 25 (yes 25 / partly 12.5 / no 0): E exists 25, P in her default path 25, M measured 12.5, W measure shows it working 12.5. ± 12.5 = one rubric step. Runs parallel work routinely; about 1 in 20 deliverables is a duplicate (38/838), and overlapping changes blur each other's measurements. v2 bench LIVE (2026-10-13, amp-bench API: Piper submits a change and gets KEEP / REVERT / INCONCLUSIVE / SATURATED back; v2 rows below): #21 Parallel improvement, v2 (record suite, $0). Score = 50 x (change_ledger rows measured in a window with no other change to the same file or subsystem ÷ rows) + 50 x max(0, 1 - 10 x duplicate share of one-off delivered work) over 14 days (a task text seen 5+ times is a routine job, not a duplicate). Overlapping changes blur each other's measurements; duplicates waste work. |
Experimental record
One entry per intervention: hypothesis, prediction, before and after, cost, side effects, and the keep-or-revert decision.
No interventions recorded yet.