5.7 KiB
Sanitized detailed source transcription S15, captured 2026-10-24. Original planning steps, acceptance criteria and historical evidence are retained rather than condensed. Read current context and precedence first: dated holds, approvals, test counts and “latest” labels below belong to their original checkpoint and do not grant current execution/publication authority. Privacy replacements preserve roles and test semantics. See section coverage.
Mandatory AI execution timing rule
Operator instruction recorded 2026-10-22. Consult this file EVERY time estimating remaining work, duration, cost-related execution time, or a delivery date. Applies across this operator's development work, not only this release.
Rule
Estimate from measured AI-agent execution on comparable work with the same acceptance exit. NEVER use human developer hours, working days, story points, generic engineering schedules, or a human-to-AI speed multiplier. Never convert execution hours into calendar days using an eight-hour working day. A model allowance (including B1) is not measured throughput or a delivery promise.
Keep the original requirement IDs and exit criteria fixed. Defects and retries belong underneath their existing requirement. Do not claim a complete requirement from a source/test sub-exit. Scope changes require an explicit ledger change, not an expanding denominator hidden in a new forecast.
Formula
For each comparable work class k:
unit_time[k] = observed_execution_elapsed[k] / qualified_exits[k]
remaining_execution = SUM(remaining_units[k] * unit_time[k]) + unestimated_work
- observed_execution_elapsed includes agent investigation, coding, debugging, failed attempts, repeated builds, tests, integration and qualification attributable to those units, not just successful commands. Use timestamped session/tool evidence, not file mtimes or invented start times.
- Include actual attributable execution spent on unfinished/failed units in the observation cohort and disclose that unfinished work; do not select only cheap successes. If there are zero qualified comparable exits, unit_time is UNKNOWN, not zero.
- Exclude separately recorded operator pauses, resource queues and external authority/device/review waits from execution elapsed. Unclassified gaps remain unknown; do not silently classify them as execution or waiting.
- Use elapsed interval unions within a lane: overlapping agent/build/test time counts once. Do not add compiler time already included in an end-to-end measured cycle. Do not add a second arbitrary retry percentage when observed retries are already included.
- remaining_units are actual outstanding exits at the SAME granularity and complexity as measured exits. Do not extrapolate from test counts, lines of code, commits, or a small fixture to an entire installed app. Do not apply measured compile speed to implementation/integration work. Do not assume partially complete work has zero remaining cost or an invented completion percentage.
- unestimated_work is an explicitly listed unknown term, NOT zero. When any material term is unknown, give the measured subtotal and state that a full total/date is not established. A sample of zero whole-app completions cannot support a whole-app delivery estimate.
For multiple real execution lanes:
execution_makespan = resource_constrained_schedule(remaining tasks, dependencies, measured durations)
conditional_finish = current_time + schedule(remaining tasks AND explicit waits/availability)
For one continuously available serial lane, execution_makespan equals remaining_execution. Count concurrency only where workers/resources actually exist and concurrent work is authorized. Critical path alone is merely a lower bound when resources contend. External waits may overlap work: place them at their dependent gates rather than blindly summing every wait. No assumed unattended or 24/7 operation. Unknown availability/waits mean no unconditional calendar date.
Required timing response
- Fixed scope and qualified exits completed since the previous checkpoint.
- Remaining measured units by class and any unestimated classes.
- Observed timestamp window, sample size, comparable acceptance boundary, failed/retry work and evidence location.
- Calculated remaining execution subtotal/range in AI elapsed minutes/hours; distinguish a measured estimate from a model allowance.
- External waits, resource/availability assumptions and conditional calendar implications, separately.
- Reconcile against the previous forecast: exits completed, newly discovered work under existing IDs, changed measurements, explicit scope changes and forecast error. Never restart a rolling countdown.
With sparse observations, report observed ranges and uncertainty; do not label them confidence intervals or p50/p95 without sufficient data. Never manufacture a finite total merely because the operator requests one. A rough scenario may be supplied only with explicit assumptions and unknown terms. Cost needs actual usage and pricing evidence; elapsed time alone is not a billing estimate.
Correction of earlier reporting
The prior 'further weeks' statement was not derived from a measured end-to-end completion rate. It is withdrawn as a quantitative forecast. Existing compile/browser timings remain valid for those bounded activities only. They do not establish the remaining duration of all 35 platform requirements and eight apps. This rule supersedes human-hour estimates; it does not erase recorded failures, remaining work or authority gates.