Update (2026-09-28)
experiment_list returns 632 usable primary sessions for this card, against 102 on 2026-09-15. Almost all of the difference is the library acquisition matrix that began on 15 September (mostly the matrix-* and explore-library-* arms; 59 capture requests name this card), claimed here through the plans. The researcher decided on 2026-09-28 to keep the matrix under this card rather than split it into its own. The matrix is an acquisition study, a fixed reference beside rotating comparisons of channel, width, rate and payload, so its sessions answer "what does each configuration deliver" and not "how many people". The first result drawn from it is the 2026-09-28 width pilot, recorded on EXP-013.
Executive summary (2026-09-15)
The acquisition plane has produced data; its complete acceptance case is still open. On 2026-09-15, experiment_read("EXP-P1") returned 102 usable primary sessions. The card declares 135 sessions. These are provenance counts, not independent count trials or a statistical result. The previous statements that nothing had run, no data existed and this was the programme's only real-hardware experiment were stale.
The current thesis asks what intermittent calibration makes identifiable, how calibration information decays, and when a count must become coarser or be withheld. Its immediate bottleneck is independently timed headcounts crossed with activity, followed by chronological evaluation scored before refitting. EXP-P1 should deliver the acquisition and label contracts for that test. It should not block it on bringing every owned chip family into one deployment. See thesis/Thesis Narrative and Occupancy inference with identifiable calibration.
What the instrument must distinguish
CSI is a radio channel measurement. Its variation can reflect bodies and movement, but also acquisition rate, transmitter mixture and the instrument. BLE observations describe detectable devices. A phone's broadcast, QR dwell or declared stay provides evidence about that participant; none establishes the total number of people in the room.
An independent reference names the spatial scope, count convention, observation time and activity condition without using the detector to choose its boundaries. The label and the sensor stream are joined only after both have their own provenance. A scheduled action is an instruction; an observed action is evidence.
This separates the two acceptance questions: did the instrument preserve what happened, and did the calibrated estimator recover the count? EXP-P1 owns the first. EXP-F1–F3 own the second and must retain unsuccessful or withheld predictions.
Results and evidence boundary
The archive read succeeded after an initial timeout. Its primary-session table covers hires-listen-44-80, hires-listen-44-160 and chan-6ghz-psc5 captures on monad01…monad10. A profile's requested band or width is acquisition intent; it does not certify the decoded frame class, bandwidth or source.
The audit retains quarantined sessions with no-records and clock-seam findings, including September 10–11 6 GHz attempts. They remain excluded from usable totals. No integrity flag is waived by this refresh. The 102-session total does not establish sustained useful CSI, a verified 6 GHz sensing link, clock alignment, BLE/person agreement or a field-ready counting model.
The live PostGIS arrangement fiit-ground-fleet contains ten surveyed nodes on fiit-ground-0: eight in library-open, two in library-silent. That is a placement inventory, not a live health verdict. The fourteen Pi hosts and ESP32 quantities in the device list are the historical hardware inventory, not today's required acquisition roster.
Whole-register caveat: the 2026-09-15 audit reported six phantom-session errors on EXP-F2. Those references need separate repair. This review does not certify the whole register as consistent.
The final audit also reports unattributed-illumination for EXP-P1: the 102 receiver sessions have no paired injector claim. Their illuminating source remains unverified on this card. P1.4 therefore keeps source attribution open; this review does not infer an ambient source or assign an injector from a profile name.
Revised plan: P1.1 to P1.9
All nine original work items are retained below so older references remain interpretable. “Deferred” means outside the current critical path; “open” means acceptance evidence is still required. No new capture was executed in this review.
| Work item | Current disposition | Acceptance evidence and thesis purpose |
|---|---|---|
| P1.1 — ESP32-C5 CSI | Deferred extension | Pin firmware and decoded record layout before using this chip as another acquisition domain. No ESP32 result was verified here. It is not a prerequisite for the deployed AX210 calibration test. |
| P1.2 — BLE listener and ESP32-C6 | Split scope | Validate the BLE reference actually used by the fleet and handset. Treat a C6-specific implementation as an optional extension. Pair device observations with independent people counts; resemblance to a published RSSI distribution is not an acceptance test. |
| P1.3 — C5/S3 logging sidecar | Deferred extension | Reopen only if a selected experiment requires this logger. Measure bytes written and recoverable records under its actual load. The old 10–20 MB/s target was a proposed engineering budget, not a result. |
| P1.4 — Pi 5/AX210 capture | Acquisition evidenced; quality gate open | Pin kernel, driver, csid version and radio configuration. Decode CSIQ and sidecars; verify nonzero matrices, record class, actual bandwidth, source identity and useful delivery. Produce a reproducible source-scoped reduction from exact session IDs. |
| P1.5 — Distributed timing | Open | Preserve monotonic stamps and clock-transfer samples. Report offset, drift, residual uncertainty and discontinuities under acquisition load. Choose alignment tolerance from the estimator's window and test sensitivity to that uncertainty. NTP/PTP alignment does not establish carrier-phase coherence. |
| P1.6 — Collection and provenance | Archive evidenced; full chain open | Trace recorder → sealed archive → decoder → derived data → published snapshot. Reconcile counts and gaps at each boundary; preserve versions and input IDs. Distinguish packet loss, all-zero records and upload/decode backlog. |
| P1.7 — Participant and reference loop | Quest definitions live; scientific gate open | Verify start, broadcast, scan/dwell or observer event, finish and stored timestamps on each supported platform/app state. Count visible people only within a declared scope; do not substitute enrolments or BLE detections for headcount. Coordinate with EXP-P3. |
| P1.8 — App interruption and network state | Replace AP-switch prerequisite | The deployed fleet has no AP for a handset to join. Test foreground/background/lock, early finish, loss of connectivity and delayed upload instead. Record the missing interval explicitly; do not author connect_to_ap. This replaces the old commanded AP-1→AP-2 test. |
| P1.9 — End-to-end rehearsal | Open | Rehearse a specified zone, independently timed count/activity script and exact acquisition regime. Verify label/sensor joins and preserve predictions before refitting. Repeat on independent days before a reliability or drift claim; a single unattended take is a rehearsal. |
The current critical path is P1.4 → P1.5/P1.6 → P1.7/P1.8 → P1.9. Existing field captures can be assessed against the capabilities they actually require. There is no blanket requirement to rerun all historical bring-up work before reading those captures.
What closes a gate
For each gate, record the acquisition regime, software versions, exact input IDs, independent replication unit, planned threshold, measured summary and remaining limitation. The original sub-100 ms timing, ±2 s task alignment, <1% loss and four-hour duration were design targets. They are not universal sufficient conditions and are not reported as achieved.
For repeated reliability measurements, resample independent capture episodes or days while preserving within-episode temporal structure. Receivers sharing an illuminator are correlated observations. Report effect and uncertainty at that level; a large packet count does not increase the number of independent deployments. Select the number of repetitions from a declared operational tolerance and pilot variability before the confirmatory run. At insufficient replication, report a descriptive rehearsal and keep the gate open.
Source-scoped CSI evaluation follows the measurement contract: select one decoded record class and transmitter, exclude quarantined material, preserve gaps, distinguish useful yield from delivery, and exclude the first 30 minutes when evaluating steady-state CSI features. Short phone dwells must therefore coincide with an already stabilised capture.
Quest review — 2026-09-15
Publication completed: six corrected September revisions are live, five public and one operator-only. Their complete stored payloads match the reviewed files. The six originals were hidden through window updates; their step rows, enrolments and completions were preserved. Together with the two previously hidden scripts, the service now lists fourteen quests, six live and eight hidden. This is a content publication check, not a handset acceptance run.
The following table records the pre-refresh findings. Initially the live service returned eight quests: six live, two hidden, none scheduled for a future opening. All eight full step sequences were read. The six live quests had session-level broadcast enabled; none of the eight contained connect_to_ap. The tracked survey additionally enabled handset tracking.
| Quest | Current value | Correction required |
|---|---|---|
| Survey loop | Repeated fixed-route observations across days | Retain a fixed route for repeatability. The stored sequence starts at FP-15 and finishes at OPEN-D3-IN, so it does not measure closure at the starting point. Remove the closure claim or add a verified return and scan. Its tracking instructions also exceed the untracked configuration. |
| Survey loop (tracked) | Operator trajectory reference | Restrict to the operator audience. Apply the same closure correction. Handset tracking is an uncertain positioning reference, not an exact ruler or a total headcount. |
| Study session | Self-reported participant-presence interval | Preserve early finish and foreground requirements. Remove the claim that one participant labels the whole room. Its target-free, selectable-duration probe is a presence workflow, not a surveyed fingerprint. |
| Fingerprint | QR-anchored short dwell | Refresh targets from surveyed placements. The stored list includes unplaced markers and wrong room labels. Remove the unsupported universal “thirty seconds before drift” explanation; thirty seconds is the protocol duration. |
| Treasure hunt | Natural movement and node dwells | Its ten node targets match the surveyed room assignments. Retain the route; describe BLE positions as estimates and dwells as participant observations. A successful quest is not proof of count accuracy. |
| Counting | Human observations independent of device detection | Keep the visible-count question, with a clear field of view and self-count convention. Do not sum overlapping views into room occupancy. Remove the claim that every radio/count gap is app penetration: missed people, scope, timing and model error are alternatives. |
| MONAD · Zone run — ZONE-A | Historical hidden script | Keep hidden. Its prescribed occupancy fields are instructions, not verified labels; screen check-in/out receipts and observer completeness would need validation before reuse. |
| MONAD · 30-minute chain showcase | Historical hidden chain test | Keep hidden. Fixed August timing and old hardware/profile references require a new version before reuse. Its scripted counts cannot validate the sensing result by themselves. |
The placement cross-check found that FP-13…FP-16 are in the open room, whereas Fingerprint calls them silent-room targets. FP-11 and SILENT-D1-IN have no resolved room in the queried arrangement. OPEN-D1-IN, OPEN-D2-IN/OUT and SHOWCASE-IN/OUT occur in the live target list but not in the queried anchor placements. Treat them as unresolved targets, not as invented locations. Regenerate from surveyed placements and check physical accessibility before an attended take.
The replacements are Fingerprint (2026-09), Study session (2026-09), Treasure hunt (2026-09), Counting (2026-09), Survey route (2026-09) and Survey route (tracked) (2026-09). Fingerprint now has 35 generated, surveyed targets. The surveys retain their twenty-stop route and no longer claim a return scan or loop-closure measurement. Counting explicitly asks for other visible people, excludes the observer and keeps overlapping views separate. The tracked version is operator-only. Points and estimated durations are unchanged.
Corrected quest payloads and publication receipts are under infra/lab/quests/; the detailed review is the refresh review. Remaining handset gates include the short-run clock fit, target-free study timer and broadcast capability handling on Counting. They are not declared passed by publishing new instructions.
Next scientific protocol
The existing quests mainly supply device trajectories and participant observations. They do not yet implement the thesis's independent count-by-activity experiment. Before an attended session, specify the zone and who is counted, independent start/end observations, seated/still versus walking conditions at each feasible headcount, and a record of arrivals/departures. Choose feasible counts and dates with the operator; none is assumed here.
Assign fitting, uncertainty-calibration and evaluation episodes in advance. Save each evaluation prediction before any later refit. Compare calibration policies at the same independent-label budget. A phone observation can contribute a reference only after its missingness and device/person relationship have been checked.
Review panel
Re-executed locally on 2026-09-15. These are review perspectives applied by one reviewer to the live archive, quest definitions, surveyed placements and current thesis narrative. They are not independent expert endorsements or a paid pipeline run.
| Perspective | Verdict | Evidence needed to resolve it |
|---|---|---|
| Research contribution | High relevance as the measurement instrument; no counting result established | Link each accepted capability to the independent calibration test in the September 10 synthesis. |
| Software and provenance | Acquisition exists; “nothing run” withdrawn | Reproduce the archive-to-reduction-to-publication chain with exact IDs and failure visibility. |
| RF and deployment | Validate actual delivered CSI separately for each regime | Class/source-scoped decoding and useful yield; profile names alone do not certify band or sensing quality. |
| Timing and identifiability | Align observations with measured uncertainty; separate devices from people | Independent time/count references and an estimator-specific tolerance. No carrier-coherence claim from PTP. |
| Statistics | Session totals are provenance, not sample size | Independent days/episodes, activity crossing, planned tolerances and frozen chronological scoring. |
| Sim-to-real | Real hardware is present; joint model validation remains open | Test the calibrated estimator against independent labels. Do not use successful transport as evidence for simulation transfer. |
| Mobile and participant experience | Live scripts contain semantic and spatial drift | Correct audience, target list, closure claims, app-state limits and observer scope while preserving run history. |
| Communication | The platform's purpose is clear once evidence and acceptance are separated | Keep dated evidence, open gates and the distinction between participant observations and headcounts visible. |
Limits
This pass does not establish capture-rate stability, a timing distribution, labelled count accuracy, calibrated coverage or a successful recalibration policy. It does not verify ESP32 bring-up or certify the physical accessibility of every marker. The archive audit is not a lake decode-coverage audit. No fleet arm or new participant session was run.
The July Wireless Laboratory Framework remains useful for role vocabulary, but its AP, deployment and scheduler-as-truth statements are historical. The current service contract and independently observed labels govern this plan.
Literature and traceability
- guarino2026_e72c ↗, retrieved section “Lack of standardization and benchmarking”: acquisition, preprocessing and evaluation need explicit comparable protocols.
- brunello2025_d781 ↗, retrieved RQ2 and introduction: temporal and environmental variation make recording conditions and chronological evaluation material.
- bocus2022_ce7f ↗, retrieved acquisition section: its Intel 5300 instrument is a multimodal provenance precedent, not an AX210/80 MHz layout reference.
- thesis/Thesis Narrative and synthesis/2026-09-10-calibration-identifiability-research-model: the current identification, calibration and independent-label requirements.
experiment_read,experiment_audit,lab_quest_list, all eightlab_quest_readresponses andgis_placements, queried 2026-09-15: the operational evidence for this refresh.
The targeted corpus retrieval supports these design cautions. It does not supply an acceptance threshold for this particular fleet or validate the current handset/person reference. Those are measurement obligations.
Cross-experiment links
- EXP-P2: historical illuminator qualification, with its retired-instrument limits.
- EXP-P3: handset and participant instrument acceptance.
- EXP-F1: independently validated reference quality.
- EXP-F2: prospective drift and recalibration.
- EXP-F3: held-out transfer and honest uncertainty.