# Ten ideas worth examining

**An internal Astra reading · 18 October 2026**  
Commissioned Codex reader, requested model **gpt-6-astra / high**. Edited by Scribe from FIND-6507. The ranking, promises and suggested tests belong to that reader. They are not claims of invention, external endorsement or superiority over simpler workflows. **Every possible test below is a suggestion, not commissioned work.**

The reader's strongest three concepts were correction-aware memory, obligations that survive handoffs, and review bound to the exact artifact. The shared concern is how agents can use records responsibly when their context and the holders of their roles keep changing.

## 1. Keep the original claim—and put its correction in the reading path

**Implemented.** An honest archive keeps what was said, but useful memory should also show a later reader that it was challenged or corrected. Corpus's ordinary record reader exposes correction relationships rather than quietly replacing the original or automatically declaring the newest account true.

The idea developed through Archivist and Architect work. One recorded correction distinguishes Human enthusiasm from an overstated “ruling.” Missing or incorrect relations can still leave a stale claim looking sound; a correction can itself be wrong.

*Possible check:* compare fresh readers' conclusions and reading cost with and without the correction headers.  
*Sources: FIND-1073/1074/2384; PROP-320, REQ-467; `corpus_story.py`.*

## 2. Let obligations outlive the current handoff summary

**Implemented.** A handoff describes what is happening; an order book records what remains owed. The distinction protects accepted work from disappearing when a successor rewrites the summary.

The Human proposed keeping enduring goals separately from current events; the Archivist developed the order-book shape. Mechanical checks can detect disappearing entries, but they cannot establish that a claimed completion is sound. Keeping every obligation forever can also preserve expensive dead work.

*Possible check:* compare lost and obsolete obligations, plus maintenance effort, against an ordinary issue tracker across several handoffs.  
*Sources: FIND-2220; PROP-354; `order_book.py`.*

## 3. Bind review to the artifact actually delivered

**Implemented mechanisms.** “Reviewed,” “landed” and “working” describe different facts. Candidate identities, sealed handoffs and post-return witnesses help keep a favorable review attached to the exact work it evaluated.

This addresses a concrete custody gap, not the harder guarantee that a reviewer notices every semantic defect. Strict identity can also impose extra review when an integration change is harmless.

*Possible check:* compare detection and overhead for wrong commits, dirty overlays, stale parents and swapped receipts against a straightforward protected-branch workflow.  
*Sources: FIND-910; PROP-232, REQ-358; `job_candidate.py`, the pipeline account.*

## 4. Compile task context without hiding what was omitted

**Historical implemented prototype; shelved.** A task-context compiler aimed to make evidence selection reproducible and budgeted, with mandatory dependencies and an explicit boundary of what remained unknown. The attraction is making selection inspectable rather than an invisible convenience.

The prototype's replay demonstrated internal consistency, not useful coverage of actual sessions. It was shelved. Its agent-designed mechanism should not be conflated with the Human's different, unresolved multi-model compiler idea. Static blind spots and preparation costs could erase any saving.

*Possible check before any revival:* compare old packets with completed jobs' actual reads, changes and review misses, using ordinary search as the baseline.  
*Sources: FIND-630/665/695/722; PROP-145, REQ-255; archived task-context compiler.*

## 5. Give exploratory research an honest stopping point

**Exercised in bounded studies.** A study can finish with missing observations, a ceiling or no worthwhile next experiment. Deciding beforehand what each outcome permits helps prevent spare allowance or disappointment from becoming an endless search for a favorable result.

P2 stopped with an unstarted task pair. P3-D completed its planned load range but nominated no larger same-range comparison. Careful procedure did not guarantee informative tasks, and full crew cost was not measured.

*Possible check:* for a future bounded question, measure design, implementation, review and reporting costs as well as participant use, with the intended decision stated in advance.  
*Sources: FIND-6375/6432/6445/6480; retained P3-D report, FIND-6479.*

## 6. Sample real work, then adjudicate what happened

**Untested future proposal.** The Human suggested sampled or filtered records, followed by agents doing the work and the crew determining what was correct. The turn explicitly said not now and not to alter the current study.

The appeal is difficulty not hand-built to suit a benchmark, with evaluation labor shared beyond the Human. But sampling, tasks and adjudication remain undecided. Agents can share misconceptions; consensus is not ground truth.

*Reader's proposed extension:* independently judge a small preselected sample before reconciliation, with Human checks on disputed cases and a random subset. This detail is the reader's suggestion, not part of the Human's settled design.  
*Sources: FIND-6463, following FIND-6452.*

## 7. Separate fidelity from present-day applicability

**Applied practice.** First ask what someone actually said; separately ask whether those words still govern this situation. A faithful old instruction can be stale, while a current interpretation can still misquote its source.

The distinction developed through cooperation among the record-keeping offices. It has exposed questions and tentative thoughts promoted into “Human rulings.” Source checks are costly, transcripts can expire, and the corrective reading can overinterpret too.

*Possible check:* independently classify a small sample of relayed directions as quotation, interpretation and claimed authority, then compare them with preserved source turns.  
*Sources: FIND-2382/2384 and the corrections those records retain.*

## 8. Classify transcript material by its evidential role

**Existing local component.** Conversation, operational evidence, repeated database output and host packaging are not interchangeable search hits. A structure-aware classifier tries to separate those roles while keeping uncertainty visible.

This addresses repeated-ledger echo rather than merely changing keyword scores. The documented component depends on host formats, has incomplete Codex file-read coverage and has not established a general retrieval benefit. Transcript volume is also not a measurement of live context use.

*Possible check:* hand-label a small cross-host sample and measure useful evidence lost and retrieval quality—not only classifier agreement.  
*Sources: FIND-138/276; REQ-178; classifier contract and Chain C account.*

## 9. Organize research around questions, not another warehouse

**Accepted proposal at the time of reading; not claimed implemented.** A concise, evolving account of a question can point to protocols, observations, first scores and paper editions that remain with their original evidence owners.

This separates understanding from custody and keeps unused preparations, development cohorts and public editions from being mistaken for the same result. The proposal reuses existing Library mechanisms. Its risk is that the synthesis becomes another contradictory authority.

*Possible check:* have a fresh reader recover P1's current meaning and reproduce its retained-score table without consulting room chat.  
*Sources: Human organization request FIND-6465; proposal FIND-6472 and acceptance FIND-6473.*

## 10. Ask an efficiency question when correctness cannot distinguish formats

**Narrow observed result.** Identical recorded answers can accompany different input use. A Human question prompted the later audit in which table representations consumed fewer reported input tokens than vertical records under the studied conditions.

That is a different question, not a retroactively invented accuracy win. The P3-D pairs reuse cases across loads; host context and uncontrolled caches further limit inference. Speed and billed savings were not established.

*Possible check:* compare exact serialized sizes first; only if warranted, specify a separate controlled resource study with correctness as a constraint.  
*Sources: FIND-6483/6484 and the retained audit.*

## What the reader would prioritize

For conceptual strength: **correction-aware memory, durable obligations and exact-artifact review**. For future research attention, the reader substitutes **sampled-record adjudication** into the third place because it could test what the current constructed tasks leave unknown. That is a recommendation, not approval to run it.

The original reading was bounded and read-only: named records, pages, reports and selected implementation seams, without tests or model experiments. These ideas connect familiar mechanisms in a particular agent environment; their comparative value remains an open question. Source identifiers refer to the private archive, not independent validation.