Skip to content

feat(autonomy): LIVE model proposals at flight time - the ledger's last gap (P6) - #307

Merged
IanFrelinger merged 5 commits into
masterfrom
feat/autonomy-live-proposal
Aug 15, 2026
Merged

feat(autonomy): LIVE model proposals at flight time - the ledger's last gap (P6)#307
IanFrelinger merged 5 commits into
masterfrom
feat/autonomy-live-proposal

Conversation

@IanFrelinger

Copy link
Copy Markdown
Owner

Summary

The ledger's last open item, executed: a model called LIVE at flight time, its proposal judged by the real loop. run-first-flight.ps1 -Live calls the local provider — ollama codellama:7b, the seam the general-generation arc named as production (model:ollama:isolation-enforced) — with a committed witness-blind prompt, records the raw exchange as a committed artifact (record/replay discipline), and mounts the recording into the flight as a read-only input. The spike extracts source mechanically and validates nothing beyond locating a class name: the gate is the judge, and every -Live run is a fresh acceptance sample. Base: master (the P-stack landed first).

The four-sample campaign

SampleVerdict
1REJECT mutation 0.08 — the IndexOf + 5 offset mutant, trim-shadowed by every existing witness message (witness gap #3, expressible only by this implementation shape)
2REJECT mutation 0.08 — same survivor; the colon-adjacent hardening was necessary but not sufficient
3CERTIFIED (signed, escape 0) → HELD by the swap host — the model omitted the constructor; the compiled brick self-declared an empty Id and verify-at-load refused the mismatch (brick-id-mismatch). Defense-in-depth on a live candidate: the gate certifies, the host independently re-verifies
4PASS — AdmittedAndSwapped in 13.3s: escape 0 against the six-case witness, full session containment, autonomous Tier-0 swap of live model output, 3/3 serving invocations, zero leaks

Measured live acceptance: 1/4 — reported, not targeted (the P3-S3 discipline). The campaign forced two further contract-derived witness hardenings (colon-adjacent marker; marker-terminal line) and one prompt reshape (skeleton-completion form — the constructor arrives verbatim, the model contributes the behavior: the A2.3 manifest-scaffold shape).

The observation worth keeping

Every distinct implementation shape that has traversed this loop — hand-authored, recorded model, live model — surfaced a witness gap the previous shapes could not express. Four of the six witness cases were demanded by the gate rejecting real candidates. Proposer diversity is adversarial witness-hardening.

Limitation 5 closes to its end state: the mechanism is proven on every leg the spec demands (model identity hash-bound in lineage, witness-blind prompting, full containment, honest rejections, a swap-host identity hold, an autonomous swap of live output). What remains is breadth — a standing proposer loop over many objectives — which is host operations on seams that all exist.

Testing

🤖 Generated with Claude Code

PlzTouchGrassand others added 5 commits August 14, 2026 12:11
run-first-flight.ps1 -Live calls the LOCAL provider (ollama codellama:7b,
the seam the general-generation arc named as production:
model:ollama:isolation-enforced) at flight time with a committed
witness-blind prompt (objective + interface contract only; the witness's
values never appear), records the raw exchange to
spikes/autonomy-first-flight/recordings/ (record/replay discipline), and
mounts the recording read-only into the flight container as an INPUT -
the proposal is data the flight consumes, exactly as a proposer cluster
would hand it over.
The spike's --live leg extracts the source mechanically (fenced block,
else first using-to-last-brace) and does NO validation beyond locating a
class name: the certification gate is the judge, and pre-filtering would
blur whose verdict counts. Live proposals require full session
containment like --proposed - model code never executes in the harness
process. The provider identity rides the lineage proposer signature,
hash-bound into generation-depth.
Acceptance is measured, not targeted (the P3-S3 discipline): whatever the
gate decides about a 7B model's proposal - ADMIT or any REJECT - is the
evidence, and every -Live run is a fresh sample.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
… last gap closes
Four live samples from ollama codellama:7b, each called AT FLIGHT TIME,
each judged by the real gate under full session containment:
1: REJECT mutation 0.08 - the IndexOf+5 offset mutant, trim-shadowed by
every witness message (witness gap #3, found only by this shape)
2: REJECT mutation 0.08 - same survivor; the colon-adjacent hardening
was necessary but not sufficient
3: CERTIFIED (signed, escape 0) then HELD by the swap host: the model
omitted the constructor, the compiled brick self-declared an empty
Id, and verify-at-load refused the identity mismatch - defense in
depth demonstrated on a live candidate
4: PASS - AdmittedAndSwapped in 13.3s, escape 0 against the six-case
witness, autonomous Tier-0 swap of live model output
Measured live acceptance 1/4 - reported, not targeted. Limitation 5 is
rewritten to its end state: the mechanism is closed on every leg the spec
demands; what remains is breadth (a standing proposer loop over many
objectives), which is host operations on seams that all exist.
The proposer-diversity observation is recorded: every distinct
implementation shape that traversed the loop surfaced a witness gap the
previous shapes could not express. Four of the six witness cases were
demanded by the gate rejecting real candidates.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
@cursor

cursorBot commented Aug 14, 2026

Copy link
Copy Markdown
Contributor

Bugbot is not enabled for your account, so this pull request was not reviewed.

Enable Bugbot in the Cursor dashboard to get automatic reviews on future PRs.

@IanFrelinger
IanFrelinger merged commit f1e8778 into masterAug 15, 2026
3 checks passed
@IanFrelinger
IanFrelinger deleted the feat/autonomy-live-proposal branch August 15, 2026 02:46
Sign up for freeto join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants

@IanFrelinger@PlzTouchGrass