Quarterly board pass
Once a quarter a frontier model reads this whole registry and challenges it. Its written verdict on the quarter lands here; the constraints it wants re-decided land in the queue below.
Open questions
A fork that can’t be priced for want of a preference asks here. Answering one records the judgment and, where the question names a constraint, writes it straight into the registry below.
Registry
Add a row here and it is live for every reader on the next read. The registry is keyed by constraint, scope and scope reference together — writing a key that already exists at the same scope replaces its value rather than adding a second row beside it.
A global row leaves the scope reference empty; a category or channel row must name the one it binds, or it lands as a second global row under the same key.
screening.threshold.<name>— global scope- One breach gate the daily screen reads. Until at least one of these exists the screen opens no forks at all. The value is JSON:
metric(the feed metric to watch),direction(beloworabove),threshold(a number), and optionallydecision_class(which class the fork is filed under),cost_of_delay_usd_wk(dollars per week of not deciding, and what the materiality floor is compared against) andwindow_days(only read readings this recent). materiality_floor_usd— global scope- Dollars. A breach whose declared cost of delay is known and below this floor is absorbed as immaterial and opens no fork. A breach with no priced cost of delay is never suppressed.
benchmark.<key>— global scope- A peer-normal figure for a metric. The value is JSON:
benchmark(the number),direction, an optionalmetric(defaulting to the key suffix),min_gap_pctandwindow_days. A gap raises a benchmark signal, which informs but opens no fork. autonomy.shadow_baseline- The minimum human-versus-model agreement rate a decision class must hold before it can be promoted a level.
autonomy.shadow_min_n- How many comparable shadow decisions must exist before that agreement rate is trusted at all.
Sources
Every feed this tenant reads, and the lifecycle status that decides whether it can raise a signal. Shadow scores but stays silent; live surfaces; disabled is off.
Allocation proposal
Where the next marginal dollar equalizes across the channels with a fitted response curve. This card only proposes; it moves no dollar and there is nothing to press on it — a reallocation is taken by deciding a fork, not by reading this table. Every figure is read off a fitted curve rather than measured, and each row carries its own caveats where the curve is being read outside its evidence.
Landed cost upload
When no automated COGS feed exists, paste landed costs here and they are written against the registered manual COGS source. The header row must carry style and landed_cost; ts is optional and must be written YYYY-MM-DD. A row without one is an undated standing cost — it prices observations at any date — while a dated row takes effect from its date onward. Any other header column is kept alongside the row as metadata. The first bad row refuses the whole upload — nothing is written by halves.