An agent starting with confirmed personal criteria needed no clarification replies across 12 synthetic tasks. The same agent with generic instructions and permission to learn and save preferences needed nine replies. Both arms produced matching first submissions after clarification and matching final submissions in all 12 tasks. A separate condition-masked model review passed all submitted plans under the frozen content rubric.
The initialized package received information that the baseline initially lacked. It also received the align skill and its retrieval instructions. This result describes that package in a narrow synthetic workload. It does not isolate ALIGNS.md's file name, establish superiority over equally informed AGENTS.md or another memory system, or measure human effort. The earlier information-equivalent preference-transfer study remains a separate tie.
Four authored profiles cross two preferences: one workflow in depth or a survey of several workflows, and advanced instruction or first-use teaching. Every combination fits the stated task budget. Each profile receives a workshop, guide, and demo task in that order, with a fresh conversation for each task. There is one repetition, giving 24 episodes across the two arms. These are synthetic profiles, not human participants.
A starts with generic AGENTS.md and the current task. B starts with the same instructions and task, plus confirmed preferences from an authored earlier conversation in ALIGNS.md, the align skill, and its instruction block. Both retain workspace files across tasks and may ask questions or save confirmed preferences. A can therefore acquire and reuse the criteria during the run. Neither receives the earlier conversation itself.
Every task includes the full response schema and allowed values. Questions use ordinary language plus explicit scope/audience topic IDs. The scripted user supplies only the requested criteria; corrections would supply only mismatching criteria. This bounded interface does not test open-ended preference discovery. Each episode permits three replies and four actor calls. The frozen protocol and cases contain the complete design and authored prior conversations.
Zero-reply fit requires the first response to be a valid submission matching both preferences. A question does not count as a matching submission, even when asking is appropriate. First-submission fit allows clarification before the first deliverable. These measures answer different questions.
| Measure | A: learns during tasks | B: initialized preferences and method |
|---|---|---|
| Zero-reply preference fit | 3/12 | 12/12 |
| First-submission exact fit, after any clarification | 12/12 | 12/12 |
| Final exact fit | 12/12 | 12/12 |
| Later user replies | 9 | 0 |
| Clarification replies | 9 | 0 |
| Preference corrections | 0 | 0 |
| Formatting corrections | 0 | 0 |
| Repeated authored criterion units | 18 | 0 |
| Runtime failures / unresolved episodes | 0 / 0 | 0 / 0 |
All nine baseline replies supplied both criteria. Four were initial clarifications, one per profile; five were later requests for criteria already supplied within the same trajectory. All 18 units repeat criteria from the authored prior conversation, but only ten units repeat information delivered to A earlier in this run. The unit definition must not be read as nine instances of forgetting information already in the baseline's conversation history.
The paired reply counts were:
| Profile | Preferences | A: workshop / guide / demo | B: workshop / guide / demo |
|---|---|---|---|
| P1 | Focused, experienced | 1 / 1 / 1 | 0 / 0 / 0 |
| P2 | Focused, newcomer | 1 / 0 / 0 | 0 / 0 / 0 |
| P3 | Broad, experienced | 1 / 1 / 0 | 0 / 0 / 0 |
| P4 | Broad, newcomer | 1 / 1 / 1 | 0 / 0 / 0 |
A saved preferences in every trajectory. P1 and P4 used LEARNING_PREFERENCES.md but asked again on both later tasks. P2 updated AGENTS.md after the workshop and reused it immediately. P3 updated AGENTS.md after the guide and reused it for the demo. This variation includes retrieval behavior and the timing of saving, not just whether storage was permitted. The evidence audit checks those writes and reads against native traces.
The authored prior conversation contains one user reply per profile in both workflows. It is a common sunk cost. B additionally assumes one reply approving storage, or four pooled replies. That confirmation and the deterministic file construction were modeled; no setup interview or installation session was run.
Including this assumed incremental setup cost, pooled cumulative replies are A: 4, 7, 9 and B: 4, 4, 4 after workshop, guide, and demo. The arms break even after the first task. After all three tasks, the total is nine observed replies for A versus four assumed setup replies for B. Including the common authored prior conversation would make the totals 13 and eight.
Payback depends on the assumed cost. The sensitivity below reports the first later task where saved replies meet or exceed setup cost. It is descriptive reply accounting, not a measured return on human time.
| Assumed setup replies per profile | P1 | P2 | P3 | P4 | Pooled |
|---|---|---|---|---|---|
| 0 | Workshop | Workshop | Workshop | Workshop | Workshop |
| 1 | Workshop | Workshop | Workshop | Workshop | Workshop |
| 2 | Guide | Not reached | Guide | Guide | Demo |
| 3 | Demo | Not reached | Not reached | Demo | Not reached |
The zero-cost row uses the protocol's first-later-task convention. No setup cost remains to recover before that task. Installation, reading, maintenance, changes in preferences, and abandoned work were not measured. No reply count, text length, token count, or model duration is converted to human minutes.
A fresh reviewer configured as gpt-6-astra at medium effort read 24 shuffled packets without condition labels or the result key. It assessed scope, audience, and rationale for first and final submissions separately. All 24 first and 24 final submissions passed. Within each packet, first and final text was identical because no submission required correction.
This was a maintainer-run model review, not external human validation. Actual response text was preserved, so storage references to AGENTS.md or LEARNING_PREFERENCES.md could reveal clues to condition. The reviewer treated storage claims as unverified from text alone; the separate trace audit checks persistence. Its model setting is recorded from the subagent configuration, without an independently archived CLI model receipt for the reviewer.
The reviewer noted one plan that promised a single meeting example but later introduced a second scenario. Both teach the same capture workflow, so it passed the frozen workflow-scope criterion. This minor inconsistency remains visible in the review limitations. The rubric did not separately score actual teaching feasibility or fictional interface accuracy. The review replay reconstructs all packets and verifies 152 quoted excerpts against the submitted text. It verifies evidence provenance, not the correctness of model judgments.
Actors used Codex CLI 0.162.0-alpha.17.2 with gpt-6-astra and medium reasoning. The freeze commit and its CI receipt, created at 03:31:58 UTC on 2026-10-15, preceded the original run manifest at 03:32:04 UTC. Frozen source hashes remain unchanged.
The user paused the study for unrelated site work. At interruption, eight actor calls in seven episodes had native completion receipts. Their outputs and workspaces were retained. The orchestrator had not saved subprocess stdout, stderr, exit codes, timeout receipts, prompt-audit stdout, or intermediate workspace snapshots for those calls. Those fields remain null. Native completion establishes model-turn completion; it cannot reconstruct missing process receipts.
After the user resumed, the recovery procedure was committed and pushed before further actor calls. It reconstructed the completed responses from native traces, continued P2/A from its exact saved workspace at the guide, and ran the five untouched chains. No completed response was rerun or replaced. The execution lineage records the original trace hashes. The completed matrix has 33 actor calls in 24 fresh task conversations: eight recovered calls and 25 resumed calls.
| Resource | A | B |
|---|---|---|
| Actor calls | 21 | 12 |
| Native input tokens, all calls | 420,132 | 303,435 |
| Native output tokens, all calls | 5,634 | 3,621 |
| Native task duration, summed seconds | 310.795 | 197.566 |
| Calls missing subprocess duration | 2 | 6 |
Native token counters are differenced between completed turns in each session, then summed. Cached input is already included in input tokens. Resumed CLI receipts can repeat cumulative counters, so their raw sums are retained in resource evidence but are not reported as consumed-token totals. Native task duration and subprocess duration are different measurements. Full uninterrupted wall time is unavailable. These figures exclude semantic review and do not establish a general compute or speed advantage.
The separate unblinded evidence auditor checked all 33 responses, delivered replies, 21 command executions, four preference writes, and 15 commentary messages. It found no observed off-interface user request, hidden-answer read, or unauthorized preference write. This is an observable-action audit, not a complete filesystem isolation guarantee. Private reasoning, credentials, identity fields, host paths, and account quota metadata are excluded from publication; private originals remain retained.
The recovery-aware verifier binds all responses and prompts to native traces and replays the frozen scoring and reply policy. Its only adaptation to the frozen exact verifier is explicit handling of unavailable subprocess durations. The original verifier alone cannot process those null resource fields. An early replay made before the last chain finished reported the incomplete matrix; that diagnostic remains archived separately from the passing final verification.
PYTHONDONTWRITEBYTECODE=1 python3 benchmarks/retention/recovery/verify.py \
benchmarks/retention/results/2026-10-15-astra-v1 \
--output /tmp/align-retention-verification.json
PYTHONDONTWRITEBYTECODE=1 python3 benchmarks/retention/reporting/verify_review.py \
benchmarks/retention/results/2026-10-15-astra-v1 \
--output /tmp/align-retention-review-verification.jsonThe chart utility generates the SVG, PNG, and plotted data from verified counts. It is reporting code added after execution and changes no scoring rules.
The tasks are short learning-material plans, the criteria are explicit and stable, the answer interface is bounded, and there is one run per profile. The result supports fewer scripted clarification replies for this initialized package. A broader study would need varied tasks, preference changes, equally informed memory controls, and measured setup and maintenance with consenting human users.