Uh oh!
There was an error while loading. Please reload this page.
Synchronize latest Public updates - #17
Merged
Merged
Conversation
Each role records a single-line decision event so the runtime never has to recover a verdict from prose. The runtime then rendered those payloads back into `KEY: VALUE` lines and re-read them, which gave every field the authority to answer for the fields around it. The two readers of that footer disagree about duplicates — the front door takes the first match, the role-reply reader the last — so a forged line won or lost depending on which reader a call site happened to use and in which order the renderer emitted the fields. A Manager `config` value containing a newline could publish its own `CONTROL: ABORT`. - read the front-door fields straight off the Manager decision - read the round handoff and milestone off the Engineer decision - build the Reviewer verdict from its payload instead of `json.dumps` and back - keep the prose readers for a turn that recorded no decision at all, where the lines really are all there is Also in this change: - redact a credential whose label carries a prefix. `\b` does not open after an underscore, so `api_key=` was redacted while `OPENAI_API_KEY=` was not. This widens where a known label may start, never which labels count; guessing a secret from the shape of its value stays out of scope, and a test now pins that an unlabelled token shape is left alone. - verify the shipped source digest only when a release is being built. The manifest is refreshed by the release build, so comparing it on every commit asserted that every commit is a release. - budget the Reviewer's fixed contract prose rather than the whole prompt. The research stage checklist was 3.8k of the 8.6k measured, so 45% of a budget meant for role prose was being spent by a vertical's checklist and either side could exhaust the other's headroom. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
The report argued for verification as a quality mechanism and left its strongest evidence—that campaigns run without a person in the loop—as an aside. This inverts that: independent review is presented as the condition that makes the operator's absence acceptable, and the absence itself is measured. - retitle to "Who Drives the Harness?" and rewrite the abstract, intro, and conclusion around the Driver role and its three questions - add duty cycle and interruption rate as paired metrics, with definitions, a per-campaign table, and all 38 interruptions classified - rename verification-guided to evidence-driven, and reframe goal drift: departing from an objective the evidence has refuted is correct behavior - add the AI infrastructure vertical (FlashAttention-4, SGLang, four upstream fla-org kernels, local deployment) and the AI-for-AI vertical separating parameter-invariant from parameter-changing self-improvement - add a delivery ledger grouping campaigns by who decides the result counts - frame auto research and RSI in related work, and expand future work into the loop between them - rewrite the endogenous-harnessing discussion: the reviewer plan-challenge channel existed but was unreachable, and was repaired during preparation - correct method-section drift against the implementation: four reviewer verdicts, rolling sessions, skill and stage ownership, eight backends - state the MLE-Bench denominator so the medal rate is interpretable - add 25 references covering self-evolving agents, judge reliability, multi-agent failure modes, materials generation, and formal mathematics Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Reachability analysis from the public entry points found 34 of setup.py's 59 functions unreachable — 806 of 1463 lines. They are not merely uncalled: each is a second implementation of a file another module now owns. setup.py wrote model_api.json capability_vault.py owns it setup.py wrote gpu_resources.json capability_vault.py owns it setup.py wrote gpu_keepalive.json gpu_lease.py owns it setup.py wrote codex config.toml capability_vault.py owns it setup.py wrote the author config nothing reads it at all `run_setup` stopped calling them when it was rewritten, and `gpu_lease` falls back to defaults when its config is absent, so nothing surfaced the loss. Thirteen of them still had tests, which is worse than none: the suite was paying for code the product cannot run, and would have reported a refactor of it as a break. One of those tests earned its intent back rather than being deleted with the rest. It monkeypatched `_apply_git_identity` and `_seed_codex_config` to raise, asserting setup never writes the machine's Git identity or a backend's auth file. That guarantee now holds because the capability is gone, so the test asserts the absence directly. Also removes the tail the same analysis found elsewhere: a figure-policy group in paper_layout_review, the reviewer prompt's `_direct_memory_edit_block` (the main Reviewer is deliberately barred from editing Skills, so its result was already hard-coded empty), and three unused CLI helpers. No module in the package has an unreachable function left. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Authored and independently reviewed by this Argus daemon. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Authored and independently reviewed by this Argus daemon. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
The mechanisms that let a campaign make progress with no score to climb were implemented and scattered across the report as implementation detail. They are the answer to why the Driver's seat could not be automated earlier, and they are stated as such. Given a dense reward, all three Driver questions collapse into reading a number, which is why self-improving agents advanced fastest exactly there and why that progress did not transfer. The seat is not a separate problem from the missing score; it is what the missing score leaves behind. - add a method subsection covering the four mechanisms: frontier transitions in which informative failure counts as progress, separation of execution failure from hypothesis refutation, named non-terminal research outcomes, and an evidence bar that scales with claim maturity - note in the introduction that manufacturing a score would have the same system produce both the work and its grade - add a discussion subsection reading three existing results as progress a scalar objective would have mishandled: the retained falsified route, the abandoned data path that ran slower than native, and the simpler admitted method that raised no metric - carry the framing into the abstract and conclusion Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Uh oh!
There was an error while loading. Please reload this page.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for freeto join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Direct synchronization of the latest Public commits without additional modifications.