Merge pull request #636 from AssetOverflow/feat/proposal-review-d
feat(proposal-review): RPT-d — boundary ledger [merge LAST]
This commit is contained in:
Shay
GitHub
commit
3ec5e91433
1 changed files with 79 additions and 0 deletions
79
docs/analysis/proposal-review-reporter-2026-06-07.md
Normal file
79
docs/analysis/proposal-review-reporter-2026-06-07.md
Normal file
|
|
@ -0,0 +1,79 @@
|
||||||
|
# Proposal review reporter — boundary ledger (RPT)
|
||||||
|
|
||||||
|
**Module:** `core/proposal_review/` · **CLI:** `python -m core.proposal_review`
|
||||||
|
**As of:** 2026-06-07, on the contemplation-batch main (N1–N6 merged).
|
||||||
|
|
||||||
|
A **read-only** reporter that surfaces the comprehension-failure proposals emitted by the
|
||||||
|
contemplation pass (N5/N6) so they become a reviewable queue instead of inert files. It closes
|
||||||
|
the growth loop **without** crossing into self-modification:
|
||||||
|
|
||||||
|
```text
|
||||||
|
failure → family → proposal_only artifact → review visibility → human/actionable queue
|
||||||
|
```
|
||||||
|
|
||||||
|
## What it is — and is not
|
||||||
|
|
||||||
|
> **This is L10-adjacent, not L10. It is a proposal review reporter, not an idle loop.**
|
||||||
|
|
||||||
|
| It DOES | It does NOT |
|
||||||
|
|---|---|
|
||||||
|
| read `teaching/proposals/comprehension_failures/*.json` | advance the teaching loop |
|
||||||
|
| validate + report (deterministic summary) | ratify anything |
|
||||||
|
| **independently verify** every artifact is inert | mount anything |
|
||||||
|
| flag malformed / unsafe artifacts | modify any reader |
|
||||||
|
| exit non-zero on a safety violation | write/move/delete any file (it mutates nothing) |
|
||||||
|
| — | affect serving in any way |
|
||||||
|
|
||||||
|
## Relationship to `ChatRuntime.idle_tick`
|
||||||
|
|
||||||
|
`ChatRuntime.idle_tick()` (L11, chat/runtime.py) **remains the only `idle_tick`.** It *writes*
|
||||||
|
proposals over a **different** stream — the reviewed-learning flywheel's discovery backlog →
|
||||||
|
the persistent proposal log in the engine-state dir. This reporter *reads* the **comprehension_failures**
|
||||||
|
stream and reports. Different stream, different verb (write vs read). To avoid the ambiguity of a
|
||||||
|
second "idle tick" / parallel path, this is a standalone reporter with its own name.
|
||||||
|
|
||||||
|
A clean **future** PR may call this reporter from `idle_tick` as a **read-only sub-pass** (surface
|
||||||
|
pending comprehension-proposals alongside the learning passes) — writing only a summary to the
|
||||||
|
existing engine-state report surface, never mutating proposal artifacts, never ratifying.
|
||||||
|
|
||||||
|
## Components
|
||||||
|
|
||||||
|
| Phase | Module | Role |
|
||||||
|
|---|---|---|
|
||||||
|
| RPT-a | `scan.py`, `model.py` | read `*.json` → typed `PendingProposal`; flag `MalformedArtifact` |
|
||||||
|
| RPT-b | `report.py` | deterministic summary (total / by family / by status / malformed / review-needed) |
|
||||||
|
| RPT-c | `safety.py`, `__main__.py` | independent safety dry-check + read-only CLI |
|
||||||
|
| RPT-d | this doc | the boundary ledger |
|
||||||
|
|
||||||
|
## The safety dry-check (the load-bearing part)
|
||||||
|
|
||||||
|
The reporter's value is not just visibility — it is **verification**. The dry-check confirms,
|
||||||
|
without trusting the emitter, that every artifact is inert:
|
||||||
|
|
||||||
|
```text
|
||||||
|
status == "proposal_only"
|
||||||
|
mounted == false
|
||||||
|
requires_review == true
|
||||||
|
content-address consistent: filename == sha256(failure_family : problem_text_sha256)
|
||||||
|
path under the sink
|
||||||
|
no malformed (unverifiable) file
|
||||||
|
no serving-path module imports/reads the sink
|
||||||
|
```
|
||||||
|
|
||||||
|
Each assertion is proven meaningful-fail in the tests; the CLI exits non-zero on any violation.
|
||||||
|
|
||||||
|
## Determinism note
|
||||||
|
|
||||||
|
The proposal artifacts are **content-addressed and carry no timestamp** (the emitter is clock-free
|
||||||
|
so the same failure is idempotent). The report is therefore fully deterministic, and **time-based
|
||||||
|
"oldest/newest" is intentionally omitted** — an honest temporal order is not in the data, only in
|
||||||
|
non-deterministic filesystem mtime. If a temporal queue is ever wanted, it must come from a
|
||||||
|
separate, clearly-non-deterministic index, not from the report.
|
||||||
|
|
||||||
|
## After this lands
|
||||||
|
|
||||||
|
The natural next step is **Option A**: wire this reporter into `ChatRuntime.idle_tick()` as a
|
||||||
|
read-only sub-pass (summary only, no mutation, no ratify) — connecting the growth organ to the
|
||||||
|
already-existing idle mechanism without changing learning semantics. Option B (R3 capability
|
||||||
|
families: rates / time / state) is the other branch. The always-on heartbeat over real uptime —
|
||||||
|
the actual lived-spine gap — remains separate and unbuilt; this reporter does not claim it.
|
||||||
Loading…
Reference in a new issue