Date: 2026-08-29
Source run: collaboration-observables.yml run 33229934255, head 16f4ba59, conclusion success, cutoff 2026-08-29T02:50:35.294099Z
Committed evidence: results/collaboration/observables-2026-08-29T02-50-35Z.json, results/collaboration/snapshot-2026-08-29T02-50-35Z.json
This is a snapshot of one run, not a current status. It is a review candidate for interpretation, not a proven claim about the project’s trajectory.
.github/workflows/collaboration-observables.yml runs weekly (cron: "37 5 * * 1") and uploads its output with retention-days: 30. At the time of writing there was exactly one successful production run in the repository’s history, and its output existed nowhere in the tree. Thirty days after 2026-08-29 the only production measurement the observables pipeline has ever produced would have ceased to exist.
Committing the two JSON files makes the measurement durable and citable. results/ is listed in .gitignore, so these were added explicitly, matching the existing convention for results/experiments/r1 and results/benchmarks/.
Two of the eight observables carry data:
| Observable | Value | Model |
|---|---|---|
ci_evidence |
920 / 1005, posterior mean 0.914598, 95% approx interval [0.897344, 0.931851] | beta-binomial-v1 |
cycle_latency |
median 0.051389 h over 50 samples, bootstrap 95% [0.041111, 0.0675], 0 right-censored | deterministic-bootstrap-median-v1 |
Six report zero observations. The reason differs between them, and the distinction is the point of this note.
| Observable | Reading | What it means |
|---|---|---|
first_independent_review_latency |
samples: 0, closed_without_review: 50 |
All 50 pull requests in the window closed without an independent review. There is no latency to estimate because the event never occurs. |
review_concentration |
observations: 0, distinct_actors: 0, hhi: 0.0 |
Follows directly from the above: with no independent reviewer/pull-request pairs, review concentration has no population. |
contributor_recurrence |
trials: 0, posterior falls back to the uniform prior 0.5 |
scripts/collaboration_snapshot.py:203-209 counts a contributor only for a merged pull request whose author is neither the repository owner nor a bot. Zero such authors exist in the window, which the raw snapshot records as contributors: 0. |
An HHI of 0.0 here must not be read as “perfectly distributed review”. It is the value an empty population takes. The observables document already carries this caution; the run is the first concrete instance of it.
review_queue.open_review_ready: 0 is a fourth real zero, but a point-in-time one — it describes the queue at the cutoff instant and carries no information about the window.
| Observable | Reading | Cause |
|---|---|---|
ownership_concentration |
observations: 0 |
scripts/collaboration_snapshot.py writes "changed_file_owners": [] unconditionally and declares ownership_attribution_not_collected in its own limitations list. |
structural_debt |
observed_findings: 0, inventory_complete: false |
Same: "structural_debt_finding_ids": [] is hard-coded, with structural_debt_inventory_not_collected declared. |
Reading either of these as “no ownership concentration” or “no structural debt” would be a category error. Nothing was measured. The collector says so in machine-readable form, and inventory_complete: false is carried all the way into the output.
Superseded for this snapshot only. Collector v0.2 closes both gaps. The follow-up run measured ownership concentration at HHI 1.0 over 66 attributions and loaded four deterministic structural-debt findings, none of which fell inside the observed window — see Ownership Concentration: First Real Measurement. The two zeros recorded above remain the correct reading of this snapshot, which was collected by v0.1.
evidence_derived_strategy_priors: []. The run derived no priors from evidence. The evolution policy therefore still runs on the seven hand-authored uniform 0.142857 priors in config/evolution-policy-v1.json. The relevant checklist item in issue 86 remains genuinely open.maximum_pull_requests: 50, window: most_recent_pull_requests_by_creation). Widening the window would raise the pull-request count, but every additional pull request in this repository’s history has the same single author, so contributors and review_concentration would very likely stay at zero. That is a prediction, not a measurement; it is falsifiable by re-running with a larger --max-pull-requests.ci_evidence counts check runs, not distinct verified outcomes; a flaky check that reruns contributes more than once."causal_claim": false, "github_write": false, "policy_activation": false, and this note does not widen that authority.