fix: [no-ticket] harbor - ops integrity (exhaust-budget instruction lever, fail-closed ledger restore)#36
Merged
Conversation
…fails closed on corrupt restore) Two operational fixes: 1. instruct_exhaust_budget (default True): the instruction's "unspent budget is wasted" persistence bullet becomes a build-config lever, like instruct_multifidelity. On preserves current behavior; off makes stopping-early the agent's own choice, which is the ablation arm for measuring what the exhortation itself contributes to optimizer persistence. The "scores are noisy" fact stays unconditional. 2. _load_or_build_ledger fails CLOSED on a persisted ledger that exists but cannot be parsed: metered budgets restore with zero remaining, and the unreadable file is preserved as ledger.corrupt for the operator. The old fallback restored the CONFIGURED budgets, which refunded the agent everything already spent, so any crash that corrupted the flush minted budget. A missing file is still a fresh boot; admin and finalize are unaffected either way. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Two Greptile P2s on #30, fixed at the stack tip: - An empty transcript file no longer surfaces as "" feedback: empty candidates are skipped (an empty pane falls through to the trajectory), and if everything is empty the search moves to the next failed attempt rather than short-circuiting on "". - The no-verifier-rewards error branch (agent died before scoring) now attaches the failure transcript like any failed sample: a candidate edit that crashes the agent lands exactly here, and the transcript is the only way the optimizer can see the crash it caused. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
… floor default, ordinal resume, SE naming) - Free-baseline flag is claimed BEFORE the eval await and refunded on failure: setting it only after success reopened a window where two concurrent baseline evals both resolved free (asyncio interleaves at await points). Claim-then-refund keeps both properties: concurrent callers see the claim, and a failed eval does not burn the freebie. - build_status defaults k_anonymity_floor to 5, matching the sidecar's enforcement default: a caller that forgets to pass the floor must not advertise a laxer one than gets enforced. - _route_results resumes the eval ordinal past surviving __eN dirs on a reused volume: a restarted sidecar started back at e1 and silently wiped the prior session's evidence, the exact erasure the versioned dirs exist to prevent. - score_se renamed to mean_score_se and documented: it is the SE of the zero-filled mean_score over n_samples, not of the n_scored subset. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…e crash vs infra outage)
Measured live in the transfer matrix: three champions scored 0/72 on an
off-model executor because their optimizers hardcoded temperature=0,
which that provider rejects. The harness handled it safely (rewards
floored, finalize shipped) but the durable record could not say WHY, and
"why" decides opposite actions: a deterministic candidate crash is a
real, reportable portability failure; an infra outage means invalidate
and re-run. Two changes:
- Collation's no-verifier-rewards error string now names the dead
attempts' exception types ("attempts died: UnsupportedParamsError
x6"). The error string is the one field that flows to the DB, the
per-sample files, and the verifier.
- _admin_eval_score returns (score, failure_cause); a floored target's
target_errors entry carries the dominant per-sample causes (frequency
summary, top 3). Diagnostics are fail-safe: any surprise result shape
degrades to a fixed string, never fails finalize.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…ask crashes cluster Greptile follow-up on #37: _dominant_sample_errors keyed on the raw error string, which embeds the task name, so identical exceptions across a multi-task slice landed in 1x singletons instead of one dominant cause. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
… finalize Home-model evals cannot see model-specific couplings the optimizer bakes in. Measured live in the wave-1 transfer matrix: three of five champions independently hardcoded temperature=0 (a variance trick on their home model, gpt-4.1-mini) and scored 0/72 on claude-opus-4-8, which rejects it, while looking healthy on every eval the optimization loop ever ran. The portability failure was invisible until a separate, manual, after-the-fact probe. VerificationTarget gains `model`: a target with an executor override scores the selected commit under a model it was NOT optimized on, in the same finalize battery as its home-model reward. The baseline is scored under the same override so the comparison stays like-for-like. Plumbing: build.yaml TargetSpec -> compiler -> serve.json _TargetCfg -> VerificationTarget -> engine.evaluate_admin(model=...) -> task_params ["harbor_model_override"] -> HarborRunner -m flag. The override rides task_params, so Mode A ignores it and the runner needs no new state; the shared run_constraints are copied, never mutated. Test fakes of evaluate_admin widened to accept the new kwarg (a strict signature turned the new call into a retried TypeError, flooring rewards, which is itself a nice demonstration of the floor's fail-safe). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…utage retry, key-budget alarm Three infra failure modes measured live in the E5 matrix runs, fixed at the measurement layer: - Dead attempts are classified infra vs candidate (conservative exception-type allowlist + the litellm key-budget message signature). Labels flow into dead_exception_types, error strings, and a new n_dead_infra metric. Classification never moves a score: every dead attempt still zero-fills, or faking infra would excuse failures. Exception type names are candidate-authored, so brackets are neutralized before labeling (a class named 'XError[infra]' cannot walk in pre-suffixed). - An OPT-IN, bounded, backoff-spaced within-eval retry re-measures samples whose every attempt died of a transient infra cause (the 65-second DNS blip that killed 44/72 attempts of one eval). Off by default: against an adversarial optimizer the qualifying predicate is a re-roll lever, since a stochastic candidate that raises allowlisted exceptions on failing attempts converts all-bad rounds into fresh draws. Retry rounds run in fresh sibling jobs dirs (nesting would pool dead attempts into later resumed means) and recovered samples carry an infra_retry audit marker naming the discarded attempts. - An ERROR-level alarm names key-budget exhaustion (a spent key fails every later call identically; two matrix cells of budget-exceeded zeros were nearly booked as a portability finding), with hedged wording since the signature reads candidate-process exceptions. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…d audit history - HarborConfig rejects infra_retry_delay_s <= 0 when retries are enabled (a zero delay silently nullified the backoff) and negative infra_retry_rounds. - The infra_retry audit marker now lists EVERY discarded round in order (discarded_rounds), not just the one immediately before recovery; recovered_round names when the sample finally measured. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
feat: [no-ticket] harbor - infra resilience (dead-attempt classification, opt-in outage retry, key-budget alarm)
feat: [no-ticket] harbor - transfer targets (per-target executor-model override at finalize)
fix: [no-ticket] harbor - floored rewards name their cause (candidate crash vs infra outage)
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 free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
Stacked on #35 (harbor-8-honest-signals). Closes out the pre-paper hardening series (#34 access tiers, #35 honest signals, #36 ops).
Changes
1.
instruct_exhaust_budgetbuild-config lever (build/config.py,compiler.py,instruction.md.j2, mirrored into serve.json likeinstruct_multifidelity). The compiled instruction's "unspent budget is wasted: re-measure your champion or try one more variant" bullet was unconditional; it is now a flag, default on (current behavior byte-for-byte in spirit; the bullet splits from the "scores are noisy" fact, which stays unconditional). Off makes stopping-early the agent's own choice. This is the ablation arm for the persistence experiments: with the exhortation always baked in, there is no way to measure what it contributes to optimizer persistence vs what the optimizer does on its own.2. Ledger restore fails closed (
serve.py)._load_or_build_ledgerrestored the configured (full) budgets whenever the persistedledger.jsonexisted but could not be parsed. The persisted ledger exists precisely so a sidecar restart cannot refill spent budget; falling back to full on a parse failure refunds the agent everything already spent, so any crash that corrupts the flush mints budget. Now: a missing file is still a fresh boot (configured budgets), but an unreadable file restores every metered split with zero remaining, preserves the unreadable original asledger.corruptfor the operator, and logs at error level with the recovery instruction (deleteledger.jsondeliberately to boot fresh). Admin and finalize paths are unaffected; only agent metering closes.Tests
instruct_exhaust_budget=Falseomits it while keeping "Scores are noisy".ledger.corruptbyte-for-byte; structurally-malformed JSON entries (KeyError path) also fail closed.🤖 Generated with Claude Code
Greptile Summary
This PR closes out the pre-paper hardening series with two focused integrity fixes and several supporting signal improvements.
_load_or_build_ledgernow treats an unreadable-but-presentledger.jsonas fully spent (zeroes all metered budgets) rather than refunding configured budgets, closing the crash-to-budget-refund path. Missing files still boot fresh. The corrupt file is copied toledger.corruptas forensic evidence.instruct_exhaust_budgetflag: Splits the "Unspent budget is wasted" instruction bullet from the unconditional "Scores are noisy" fact; defaults on (current behavior) and enables an ablation arm that removes the persistence exhortation to measure the optimizer's own stopping behaviour.BaseException), volume-reuse ordinal resumption,score_se→mean_score_serename, andk_anonymity_floordefault corrected from 1 → 5.Confidence Score: 5/5
Safe to merge; all changes harden existing invariants rather than relaxing them.
The ledger fail-closed logic is correct and well-tested (missing, unparseable, and malformed-entry paths all covered). The infra-retry feature is disabled by default with documented adversarial-candidate risks. The freebie race-condition fix correctly uses the claim-before-await / refund-on-exception pattern. The only notable nit is that the error log in _load_or_build_ledger claims the corrupt file is preserved before the copy is attempted, so a disk-full failure would leave an incorrect log trace — but this does not affect the core zeroing logic.
No files require special attention; serve.py has the minor log-ordering nit but the fail-closed behaviour itself is correct.
Important Files Changed
Flowchart
%%{init: {'theme': 'neutral'}}%% flowchart TD A[sidecar startup: _load_or_build_ledger] --> B{ledger.json exists?} B -- No --> C[Fresh boot: configured budgets\nremaining = total] B -- Yes --> D[Parse JSON] D -- Success --> E[Restore persisted remaining_* values\nReturn BudgetLedger] D -- Fail\njson.JSONDecodeError / KeyError / OSError --> F[logger.error: failing CLOSED] F --> G[shutil.copyfile to ledger.corrupt] G --> H[Zero out all metered remaining budgets] H --> I[Return BudgetLedger - all splits exhausted]%%{init: {'theme': 'base', 'themeVariables': {"darkMode": true, "background": "#0d1117", "primaryColor": "#21262d", "primaryTextColor": "#e6edf3", "primaryBorderColor": "#8b949e", "lineColor": "#8b949e", "textColor": "#e6edf3", "edgeLabelBackground": "#161b22", "actorBkg": "#21262d", "actorBorder": "#8b949e", "actorTextColor": "#e6edf3", "actorLineColor": "#8b949e", "signalColor": "#8b949e", "signalTextColor": "#e6edf3", "noteBkgColor": "#373320", "noteBorderColor": "#d4a72c", "noteTextColor": "#f0e6c0", "labelBoxBkgColor": "#21262d", "labelBoxBorderColor": "#8b949e", "labelTextColor": "#e6edf3", "loopTextColor": "#e6edf3", "activationBkgColor": "#30363d", "activationBorderColor": "#8b949e"}}}%% flowchart TD A[sidecar startup: _load_or_build_ledger] --> B{ledger.json exists?} B -- No --> C[Fresh boot: configured budgets\nremaining = total] B -- Yes --> D[Parse JSON] D -- Success --> E[Restore persisted remaining_* values\nReturn BudgetLedger] D -- Fail\njson.JSONDecodeError / KeyError / OSError --> F[logger.error: failing CLOSED] F --> G[shutil.copyfile to ledger.corrupt] G --> H[Zero out all metered remaining budgets] H --> I[Return BudgetLedger - all splits exhausted]Reviews (3): Last reviewed commit: "Merge pull request #37 from scaleapi/har..." | Re-trigger Greptile