Record proposal Modeling first-result proof

This commit is contained in:
Meta Vens 2026-08-11 09:23:44 +02:00
commit 9aeb59e9ab
4 changed files with 115 additions and 16 deletions

View file

@ -3886,3 +3886,40 @@ The three failed attempts are not accepted reliability evidence. The next cut
belongs to role-result schema construction/model ingress: preserve strict
duplicate-field rejection while making the first model-produced result conform
to its one-owner schema.
## Runtime-owned proposal Modeling publication
Proposal Modeling owns semantic frontier judgment, not RepoModel causality.
- Owner: authenticated runtime launch context owns request/proposal identity,
model base, Body observation basis, patch identity, deterministic frontier
identity, evidence binding, and publication time. Modeling owns only the
closed `proposalFrontierDraft` plus ordinary role verdict/summary sight.
- Inputs: one immutable proposal Modeling request and launch context, the
current RepoModel projection, and one strict semantic draft.
- Outputs: one ordinary role result containing a runtime-composed
`RepoModelPatch` with exactly one active frontier upsert.
- Derived state: provider response-format strictness is derived recursively
from the closed schema. Every constant declares its JSON type; this is a
provider contract, not a relaxation of runtime admission.
- Forbidden writers: the model cannot author request IDs, proposal IDs, Body
basis, patch IDs, base hashes/revisions, timestamps, evidence provenance, or
arbitrary patch operations. A retry parser cannot merge duplicate fields or
repair missing anatomy.
- Shared paths: direct provider completion and replay both pass through
`role_worker_result_from_ingress` and existing runtime role-result admission.
- Cut line: the generic proposal full-`RepoModelPatch` schema is absent.
Legacy non-proposal Modeling schemas remain separate and retain their
downstream gates.
- Verification: exact `e723058a`, release
`sha256-6428f9194c261806a8859ccff9efeb9bb041a0103c5bbd8fac617aae265a651e`,
witness
`sha256-7ee2e14548bad526e946fbd79b555b7891d2eb6e2917856fefb1ebb7aeebd995`
produced one first-attempt outer Modeling result on fresh copied state, zero
open jobs, and no Hands/Eyes authority. The copied owners stopped at exit
zero under the restored brake.
The proposal grant and outer worker are single-attempt proven. Downstream
coordinator review is not yet single-consumption proven: the live run emitted
two receipt-bound `reviewModelingResult` continuation grants for the same
proposal-bound result before settlement. That route is the next ownership cut.

View file

@ -1,5 +1,51 @@
# Fresh workspace handoff
## Runtime-owned proposal Modeling first-result accepted - 2026-08-11
Actual c006 remains exact `c9a2a979`, alive and cognitively braked. All live
work below used fresh named-volume copies. The actual resident, Heartbeat,
runtime, and local-Verse stores were not mutated or stopped.
The repeated duplicate-field failures were a symptom of the wrong output
owner. Generic proposal Modeling asked the model to author a complete
`RepoModelPatch`, including runtime identity, model base, Body observation
basis, request/proposal provenance, patch timestamps, deterministic frontier
identity, evidence binding, and nested patch anatomy. Pushed `9f82ee38`
removes that mouth. Proposal Modeling now returns one closed semantic
`proposalFrontierDraft`; runtime derives and authenticates every causal patch
field and publishes the exact patch through the existing admission path.
Duplicate JSON keys still fail closed.
The exact `9f82ee38` package passed 24-binary inspection, but its fresh copied
gate exposed a provider-subset error before inference: `roleId` used `const`
without the `type` OpenAI requires. Pushed `e723058a` adds the type and a
recursive test that every constant in the dedicated schema declares one.
Focused core coverage passes. Persistent-cache packaging completed with Cargo
at 28.40s and produced exact 24-binary release
`sha256-6428f9194c261806a8859ccff9efeb9bb041a0103c5bbd8fac617aae265a651e`
and witness
`sha256-7ee2e14548bad526e946fbd79b555b7891d2eb6e2917856fefb1ebb7aeebd995`
with zero warnings. Receipts are under `.epiphany-run/linux-package-e723058a`.
On a new copy, request
`repo-frontier-proposal-modeling-c6e584188580f373360a1b9d5113e2df67a22dd16c7cefcd97361067862b6d4b`
launched exactly one outer Modeling job,
`75a69738-b694-411c-bb6d-f823b39bd9e9`. It used five internal read-only model
tool follow-ups, produced one admitted runtime-owned frontier result on the
first attempt, and terminalized the exact proposal grant. Runtime ended with
zero open jobs and no actionable Hands or Eyes frontier. The copied brake was
restored; resident and Heartbeat both exited zero. Final state hashes and logs
are under `.epiphany-run/proposal-strict-live3-e723058a`.
Do not flatten the remaining coordination wound. Heartbeat subsequently issued
two receipt-bound `reviewModelingResult` continuation grants for the same
completed proposal-bound result before the coordinator settled sleeping with
`modelingResultProposalBound=true` and
`modelingResultFailureReviewed=true`. This was not a second Modeling launch,
but it is not clean single-consumption review either. Next map the result-review
owner and require one proposal grant, one outer worker, at most one review
continuation, and exact settlement on a fresh copy. Yggdrasil never builds.
## Governed directory inventory accepted live - 2026-08-11
Actual c006 remains exact `c9a2a979`, alive and cognitively braked. All work