# Product development evidence summary — September 15, 2026

This owner-approved publication summarizes internal product development since the September 9 limited Release 1 announcement. It is not a public reproduction bundle, independent review, new release, or statistical activation. Private source, operational records and implementation tuning are not included.

## Candidate identities and recorded observations

| Candidate | Identity | Bounded change |
| --- | --- | --- |
| Extended range | Request 3.0.0; method 1.1.0-candidate.1 | Extended variance representation and authenticated local MCP integration |
| Core | Request 4.0.0; method 1.2.0-candidate.1 | Seven Welch core values and submitted-result comparison |
| Full | Request 5.0.0; method 1.3.0-candidate.1 | Complete eleven-field result or refusal; clearer actual verification scope |

The full-result development observations used frozen application source identity `0456ef2ecfdded0271028595ab985c9585038375`. The 96-case panel yielded 96 direct full calculations meeting the recorded numerical checks and 90 authenticated local MCP full results meeting those checks. Six original CSVs exceeded the unchanged 1 MiB ingestion limit and remained refusals in the original denominator. The original five-field score and additional six-field diagnostic were recorded separately; this is not a population accuracy estimate or an AI-answer benchmark.

Three alternating repetitions compared request 3.0.0 and 5.0.0 under an unchanged 30-second subprocess limit on the recorded development host. One retained case timed out on all predecessor calls (30.08–30.38 seconds) and completed on all successor calls (0.42–0.46 seconds). The predecessor's true completion time is unknown. No general speed ratio, portable latency guarantee or production measurement is asserted.

Sixteen development roots were checked with a separate 80-digit conditional-distribution calculation, alongside boundary, transformation and deliberate-error controls. These are finite numerical checks, not a rigorous enclosure proof. The later main integration did not announce cloud rollout or default activation.

## Completed milestones and next work

Safe-import and historical meaning-preservation milestones have recorded internal completion. Subsequent repairs strengthen file-path intake, historical-result identity, version-record persistence, replay-unavailable classification and registered compatibility decisions. Internal completion does not itself establish public API availability.

Record assembly and emission has entered development. Invite-free access covering Welch, independent multi-group and paired two-condition capabilities is an adopted initial release plan, not current support or a promised date. Hosted results do not yet provide public Records for local replay.

## Provenance and interpretation

Accountable role: Licklider product/publication owner. Publication was explicitly authorized on September 15, 2026. Candidate implementation and adversarial self-review used the same OpenAI Codex assistant; no independent engineering or statistical review is inferred. This article's preparation likewise used ChatGPT/Codex. Internal source and raw measurements remain private; readers can assess the disclosed scope but cannot independently reproduce this panel from this summary alone.

Current public artifacts and their independently inspectable scope are linked from [nomue documentation](https://www.licklider.ai/docs/). This development summary does not expand those artifacts' claims.